README file for ibv-conduit =========================== Paul H. Hargrove @ TOC: @ @ Section: Overview @ @ Section: Where this conduit runs @ @ Section: Terminology @ @ Section: Build-time Configuration @ @ Section: Job Spawning @ @ Section: Runtime Configuration @ @ Section: Multi-rail Support @ @ Section: On-Demand Paging (ODP) Support @ @ Section: HCA Configuration @ @ Section: Advice to Client Authors @ @ Section: Known Problems @ @ Section: Core API @ @ Section: Extended API @ @ Section: GEX_FLAG_IMMEDIATE Support @ @ Section: Graceful exits @ @ Section: References @ @ Section: Overview @ ibv-conduit implements GASNet over InfiniBand networks using the Open Fabrics Verbs API (www.openfabrics.org). The name "ibv" comes from a time when this API was known as "InfiniBand Verbs" (its library is still named libibverbs). @ Section: Where this conduit runs @ ibv-conduit runs over networks using the Open Fabrics Verbs API, such as InfiniBand. Platforms known to support this API include Linux, Solaris, AIX, Windows and FreeBSD. However, at this time only Linux and Solaris have been confirmed to run GASNet's ibv-conduit. While Open Fabrics Verbs also covers iWARP and RoCE, ibv-conduit does not currently support these networks. Users interested in contributing the necessary support (primarily use of "rdmacm" for connection setup) are encouraged to contact gasnet-devel@lbl.gov. While ibv-conduit may work on Intel Omni-Path (OPA) HCAs, we instead recommend use of ofi-conduit on such systems. Nearly all InfiniBand cards (known as Host Channel Adapters, or HCAs) have support for Open Fabrics Verbs, available from www.openfabrics.org or from the HCA or OS vendor. We have tested numerous InfiniBand HCAs from Mellanox, InfiniPath HCAs from Pathscale/QLogic, and TrueScale HCAs from Intel. In addition to the Open Fabrics Verbs, vendor-specific APIs also exist and some have GASNet conduits: + ofi-conduit for the PSM (version 1) API on Intel TrueScale HCAs + ofi-conduit for the PSM2 API on Intel Omni-Path HCAs The performance of ibv-conduit relative to others on the same hardware will depend on your applications' communications patterns, the number of nodes you run on, and other parameters. Therefore, we do not make any specific recommendation as to which to use. We recommend that you benchmark your own workload if you are concerned with the very best possible performance. While ibv-conduit runs on Solaris, the libibverbs on Solaris is not compatible with GASNet's implementation of PSHM over POSIX shared memory. This is not normally of concern, because on Solaris GASNet defaults to an implementation of PSHM over SystemV shared memory. However, if one configures GASNet to use PSHM-over-POSIX on Solaris, then be advised that ibv-conduit will lack PSHM support and thus perform all communication within a compute node via the InfiniBand HCA instead of shared memory. @ Section: Terminology @ This document will use "NIC" (short for Network Interface Card) to refer to the physical object installed in a host and "connector" to refer to the external connections from the NIC to a network. The term "HCA" (short for Host Channel Adapter) will be used to refer to a device as enumerated by `ibv_get_device_list()` or the command-line utility `ibv_devices`. The HCA is the device driver's logical representation of the NIC, but there is not always a one-to-one correspondence, as described next. When a NIC has multiple connectors, the driver may present these either as a single HCA with multiple "ports" or as multiple single-port HCAs. Additionally, some systems will present more than one HCA per connector. This is typically done on systems where the NIC is connected to multiple I/O buses. On a compute node of the Summit system at OLCF, there are two external network cable connectors on a single NIC which is connected internally to two I/O buses. The driver presents four HCAs, one for each combination of external connector and internal I/O bus. So, for a single NIC with two connectors there are at least three ways the system may present the same resources: with 1, 2 or 4 HCAs. See the end of the description of the `GASNET_IBV_PORTS` environment variable for information on multiple ways to get a listing of HCAs and ports which ibv-conduit detects as available. The term "rail" is used in place of "HCA" in some contexts, but has the same meaning. In particular, one must enable multi-rail support in ibv-conduit if one is to use multiple HCAs (as defined above) in a given process. However, one can use multiple ports of a single multi-port HCA without enabling multi-rail. @ Section: Build-time Configuration @ Ibv-conduit can ensure good network attentiveness (timely processing of incoming AMs) by spawning an extra thread that remains blocked until the arrival of an Active Message. One can disable this thread by configuring GASNet with the flag '--disable-ibv-rcv-thread'. It is recommended that one NOT use this option, but instead disable the thread at runtime (see Runtime Configuration section). If the extra thread will never be needed, disabling it at build time will yield a small reduction in latencies by allowing some locking operations to compile away. By default, each ibv-conduit process in a GASNet job will open at most one Host Channel Adapter (HCA). To allow a process to utilize more than one HCA, specify '--with-ibv-max-hcas=N' at configure time (where 'N' is the number of HCAs to support per process). Alternatively, specifying '--enable-ibv-multirail' is equivalent to '--with-ibv-max-hcas=2' unless an explicit '--with-ibv-max-hcas=N' option provides a different value. Passing '--disable-ibv-multirail' overrides any explicit '--with-ibv-max-hcas=N' options. Note that multirail support includes provisions for correctness which can be relevant if using mutiple HCAs per *host*, even if using only a single HCA in each process. So, '--with-ibv-max-hcas=1' is distinct from '--disable-ibv-multirail'. Enabling multirail support (using '--with-ibv-max-hcas=1' if appropriate) is strongly recommended if one might ever use multiple HCAs per host. See 'GASNET_USE_FENCED_PUTS' in the Runtime Configuration section and "Bug 3447" in the Known Problems section for more information regarding the correctness issues. The use of specific HCA ports is controlled at run time by the environment variable GASNET_IBV_PORTS, described below. The default value of this variable can be set at configure time using '--with-ibv-ports=...'. Closely connected to selection of HCA ports is the setting of the environment variable GASNET_USE_FENCED_PUTS, also described below. Its default value can be set using '--with-ibv-fenced-puts=val' where 'val' is either '0' or '1'. Alternatively, '--with-ibv-fenced-puts' (with no argument) and '--without-ibv-fenced-puts' can be used to select defaults of 1 and 0, respectively. When using dynamic connections (see GASNET_CONNECT_DYNAMIC env var, below) there is an extra thread spawned to block for the arrival of connection requests. If needed, this can be disabled at configure time using '--disable-ibv-conn-thread'. By default, ibv-conduit uses 64KB buffers for AM Mediums, to yield gasnetc_AM_LUB{Request,Reply}Medium() values tens of bytes smaller due to message headers. This default can be overridden by passing '--with-ibv-max-medium=N' for 'N' equal to any power-of-two from 1024 to 262144, inclusive. The default spawner to be used by the gasnetrun_ibv utility can be selected by configuring '--with-ibv-spawner=VALUE', where VALUE is one of 'mpi', 'pmi' or 'ssh'. If this option is not used, mpi is the default when available, and ssh otherwise. Here are some things to consider when selecting a default spawner: + The choice of spawner only affects the protocol used for parallel job setup and teardown; in particular it is NOT used to implement any part of the steady-state GASNet communication operations. As such, the selected protocol needs to be stable and co-exist with GASNet communication, but its performance efficiency is usually not a practical consideration. + mpi-spawner is the default when MPI is available precisely because it is so frequently present on systems where GASNet is to be installed. Additionally, very little (if any) configuration is required and the behavior is highly reliable. + pmi-spawner uses the same "Process Management Interface" which forms the basis for many mpirun implementations. When support is available, this spawner can be as easy to use and as reliable as mpi-spawner, but without the overheads of initializing an MPI runtime. + ssh-spawner depends only on the availability of a remote shell command such as ssh. For this reason ssh-spawner support is always compiled. However, it can be difficult (or impossible) to use on a cluster which was not setup to allow ssh to (and among) its compute nodes. For more information on configuration and use of these spawners, see README-{ssh,mpi,pmi}-spawner (installed) or other/{ssh,mpi,pmi}-spawner/README (source). By default, ibv-conduit serializes calls to `ibv_poll_cq()` in a manner which reduces time spent blocked on the mutex internal to its implementation. One can configure using `--disable-ibv-serialize-poll-cq` to disable this behavior. For more information, see the `GASNET_RCV_THREAD_POLL_MODE` environment variable documentation, below. @ Section: Job Spawning @ If using UPC++, Chapel, etc. the language-specific commands should be used to launch applications. Otherwise, applications can be launched using the gasnetrun_ibv utility: + usage summary: gasnetrun_ibv -n [options] [--] prog [program args] options: -n number of processes to run (required) -N number of nodes to run on (not supported by all MPIs) -E list of environment vars to propagate -v be verbose about what is happening -t test only, don't execute anything (implies -v) -k keep any temporary files created (implies -v) -spawner=(ssh|mpi|pmi) force use of a specific spawner (if available) There are as many as three possible methods (ssh, mpi and pmi) by which one can launch an ibv-conduit application. Ssh-based spawning is always available, and mpi- and pmi-based spawning are available if the respective support was located at configure time. The default is established at configure time (see section "Build-time Configuration"). To select a non-default spawner one may either use the "-spawner=" command- line argument or set the environment variable GASNET_IBV_SPAWNER to "ssh", "mpi" or "pmi". If both are used, then the command line argument takes precedence. It has been noted that some InfiniBand driver implementations may not allow for multiple open()s of the adapter. In this case, spawning via MPI is not possible because the MPI and GASNet implementations cannot share the adapter. If your GASNet jobs fail to spawn via MPI, but spawn correctly with ssh or pmi, then this may be the reason. If you need mpi-based spawning, our recommendation is to attempt to set the MPIRUN_CMD such that your MPI will not use InfiniBand (see mpi-spawner's README). If that is not possible, you may need to select a different MPI implementation. @ Section: Runtime Configuration @ There are a number of parameters in ibv-conduit which can be tuned at runtime via environment variables. General settings: ---------------- Ibv-conduit supports all of the standard GASNet environment variables and the optional GASNET_EXITTIMEOUT and GASNET_THREAD_STACK families of environment variables. See GASNet's top-level README for documentation. + GASNET_BARRIER In addition to the barrier algorithms in the top-level README, there is an implementation specific to IBV: IBDISSEM - like RDMADISSEM, but implemented using lower-level operations for lower latency. Currently IBDISSEM is the default on IBV. Spawner settings: ---------------- + GASNET_IBV_SPAWNER To override the default spawner for ibv-conduit jobs, one may set this environment variable as described in the section "Job Spawning", above. There are additional settings which control behaviors of the various spawners, as described in the respective READMEs (listed in section "Build-time Configuration", above). Connection settings: ------------------- Under normal conditions, Host Channel Adapters and Ports will be located and configured automatically. However, in the event you have multiple adapters or multiple active ports on a single adapter, you may wish to set environment variables to identify the correct HCAs and Ports. Or, you may wish to use non-default values for configuring connections. These parameters are permitted to take different values on each process. However, please see bug 4314 if process-specifc values are needed. https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=4314 See "Build-time Configuration", above, for information on enabling use of multiple HCAs in GASNet ibv-conduit. + GASNET_HCA_ID + GASNET_PORT_NUM ** UNSUPPORTED ** These environment variables, used in older releases, are no longer supported. Setting them to anything but the empty string will result in a run-time warning. + GASNET_NUM_QPS This variable gives the number of IB Queue Pairs (QPs) over which to stripe traffic between each pair of peers. This can yield an increase in throughput and bandwidth when multiple physical ports are used on one or more adapters. If the number of QPs exceeds the number of available physical ports then multiple QPs will be mapped round-robin to the ports. Be aware that mapping multiple QPs per port may yield either a performance improvement or a degradation, depending on traffic pattern. The default is 0, which means one QP per HCA/port used. + GASNET_IBV_PORTS By default, GASNet will open and use one active IB port on each HCA used, which will be all HCAs (when GASNET_NUM_QPS is zero), or the first GASNET_NUM_QPS HCAs found (when GASNET_NUM_QPS is non-zero). Setting GASNET_IBV_PORTS will specify a filter for which ports will be used. This can be used for instance to cause multiple physical ports to be used per HCA, or to specify specific ports and/or HCAs to be considered (up to GASNET_NUM_QPS if it is non-zero). This variable is a string of one or more HCA/port specifications, separated by '+' characters. Each such specification gives an HCA identifier and an optional comma-separated list of port numbers. The list of port numbers, if provided, is separated from the HCA id by a ':'. If a list of ports is given, only those ports may be used. Otherwise the first active port on the given HCA may be used. The following example allows the first active port on HCA mlx5_0, and only port 2 on mlx5_1: GASNET_IBV_PORTS="mlx5_0+mlx5_1:2". Note that this list is a *filter*, which means: + Duplicate entries do not cause multiple opens of a port or HCA + Entries describing non-existent HCAs are silently ignored + Entries describing inactive ports are silently ignored + Order is not significant. In particular if GASNET_NUM_QPS is less than the number of entries in GASNET_IBV_PORTS, ports are opened in the order detected, regardless of their order in GASNET_IBV_PORTS See 'GASNET_IBV_LIST_PORTS' and 'GASNET_IBV_LIST_PORTS_NODES' for how to enumerate available HCAs and the status of their ports. In most IBV distributions the 'ibv_devinfo' utility is also available to list the HCAs and the status of their ports. The default can be set at configure time using '--with-ibv-ports=...', and is empty (no filter) in the absence of that configure option. See also 'GASNET_IBV_PORTS_*', immediately below. + GASNET_IBV_PORTS_* The environment variable 'GASNET_IBV_PORTS', described immediately above, provides only a single setting and unless one uses some external means to give per-process settings this cannot provide per-process control. This can make it difficult to get the best performance from multi-rail systems with multiple processes per node and architectural locality properties that affect PCI/adapter access efficiency. However, if 'hwloc' is detected at configure time, then it is possible to give ibv-conduit values for 'GASNET_IBV_PORTS' which will vary per-process based on cpu-binding and machine topology information as follows. 1. The variable 'GASNET_IBV_PORTS_TYPE' names an object type using hwloc's terminology, with the default being "Socket" (aka "Package"). If the value is "None" (case-insensitive) then the logic below is disabled and the value of 'GASNET_IBV_PORTS' is used by all processes. 2. GASNet queries the set of objects of the given type which intersect the process's cpuset, to construct a variable name 'GASNET_IBV_PORTS_[suff]' where '[suff]' is an underscore-delimited ordered list of logical object ids. For example, with the default object type, a process bound only to cores in the first socket would have a variable name of 'GASNET_IBV_PORTS_0'. Meanwhile, if the cpuset spans sockets 0 and 1 (such as for an unbound process on a 2-socket system) then the variable 'GASNET_IBV_PORTS_0_1' is used. 3. If the environment variable determined in step 2 is set, then it is used. Otherwise the un-suffixed 'GASNET_IBV_PORTS' is used. As a concrete example, on OLCF's Summit there are four HCAs in software, which represent connections from two I/O buses (one per socket) to two distinct InfiniBand rails. Use of the further I/O bus introduces a latency penalty, but achieving peak aggregate bandwidth requires the job to split traffic over both I/O buses and both rails. For most applications, we have observed the best latency and aggregate per- node bandwidth is achieved using a single HCA connected to the local socket's I/O bus. For an unbound process (or one bound to cores in both sockets) the performance suffers relative to the bound case, but the best average-case is achieved using two HCAs chosen to span both buses and both network rails. This yields the following recommendation as a good default for most applications running on this system: GASNET_IBV_PORTS='mlx5_0+mlx5_3' # Spanning both sockets (e.g. unbound) GASNET_IBV_PORTS_0='mlx5_0' # Bound to socket0 GASNET_IBV_PORTS_1='mlx5_3' # Bound to socket1 For an application which needs to maximize bandwidth of communication to/from processes in a single socket at a time, one must allow process to make use of both I/O buses and network rails (at the cost of increased latency and potentially reduced aggregate per-node bandwidth). This can be accomplished using two HCAs per processes as follows: GASNET_IBV_PORTS='mlx5_0+mlx5_3' GASNET_IBV_PORTS_1='mlx5_1+mlx5_2' This example illustrates the use of un-suffixed 'GASNET_IBV_PORTS' as a default when lacking a more specific setting. In particular, unbound processes and those bound to socket 0 will both use 'mlx5_0+mlx5_3' while processes bound to socket 1 will use 'mlx5_1+mlx5_2'. These specific recommendations are appropriate to the specific composition of a node of OLCF's Summit, and should not be considered as generic advice for use of all multi-HCA systems. Of course, even on the same system, your mileage may vary. By default 'GASNET_IBV_PORTS_TYPE' is "Socket" and all other variables in this family are unset. + GASNET_IBV_PORTS_VERBOSE This integer setting controls the detail of any warnings printed when there are non-fatal issues related to the selection of HCAs and ports to be used. A value of 0 supresses the warnings entirely. A value of 1 (the default) or higher will warn if one or more HCAs are excluded from consideration other than due to failure to match GASNET_IBV_PORTS (or the numerically suffixed variants). In particular, a warning is issued if more HCAs are detected than supported by '--with-ibv-max-hcas=N' (where N defaults to 2 when configured using `--enable-ibv-multirail`, and 1 otherwise). Values of 2 or higher are reserved to request printing additional details in future releases. This setting defaults to 1. + GASNET_IBV_LIST_PORTS The value is a boolean: "0" to disable or "1" to enable the reporting of all detectable HCAs and and the status of their ports. See 'GASNET_IBV_LIST_PORTS_NODES' for how to limit which nodes report. The default is "0" (no report). + GASNET_IBV_LIST_PORTS_NODES If GASNET_IBV_LIST_PORTS is enabled, then this setting may be used to limit which nodes report HCAs/ports. The value is a list which may contain one or more integers or ranges separated by commas, such as "0,2-4,6". If unset, empty, or equal to "*" then all nodes will report (if enabled by GASNET_IBV_LIST_PORTS). The default is no limit on which nodes report. + GASNET_IBV_PKEY If set, this specifies the 15-bit InfiniBand Partition Key to use. Valid values are in the range 2 to 0x7fff. For compatibility, the membership bit (0x8000) is ignored. The default is to use the Partition Key installed at table index 0. + GASNET_QP_TIMEOUT This sets the timeout value used to configure InfiniBand QueuePairs. The IB specification uses (4.096us * 2^qp_timeout) as the length of time an HCA waits to receive and ACK from its peer before attempting retransmission. The default is currently 18 (roughly 1 second). + GASNET_QP_RETRY_COUNT This sets the maximum number of retransmissions due to ACK timeout before the HCA signals a fatal error. The default is currently 7 (the max supported by early Mellanox HCAs) + GASNET_QP_RD_ATOM This sets the number of per-connection resources allocated by the HCA for responding to RDMA Reads (and atomics, which GASNet does not use currently). Lower values use slightly less memory but may reduce the throughput of Get-intensive communications patterns. The default value is '0', which means to use the maximum supported value reported by the HCA. Other valid setting are typically in the range from 1 to 4. + GASNET_MAX_MTU This sets the maximum MTU to be used, and has the following valid values: -1, 0, 256, 512, 1024, 2048 or 4096. If the value is 0 GASNet will automatically select the MTU size. If the value is -1 GASNet will use the HCA port's active value. Otherwise the lesser of this setting or the port's active value will be used. The default is 0: automatic MTU selection. + GASNET_CONNECT_DYNAMIC This boolean setting determines if connections can be established on demand. The default value is TRUE. When GASNET_CONNECT_DYNAMIC is enabled, a node will connect on demand to any peer not previously connected at startup. However, if a node is fully connected to all peers at startup, then dynamic connections are automatically disabled on that node. Therefore, unless GASNET_CONNECT_STATIC or GASNET_CONNECTFILE_IN is set to a non-default value this variable has no effect. + GASNET_CONNECT_STATIC This setting determines if connections are established at startup. When GASNET_CONNECT_STATIC is enabled, a node will connect at startup to all peers indicated by the GASNET_CONNECTFILE_IN setting (see below), or to ALL peers if that variable is unset. The value is a boolean with a default of TRUE. + GASNET_CONNECTFILE_IN This setting provides a filename used to limit the connections established at startup, and is ignored if GASNET_CONNECT_STATIC is FALSE. Any '%' character in the value is replaced with the node number to allow (but not require) separate per-node files. The format of a connect file is a series of lines of the form: node: peer1 peer2 ... without leading whitespace. For example, to request that node 7 connect to nodes 0, 4 and 6: 7: 0 4 6 Line lengths are not limited, but the same node number may appear to the left of the colon on multiple lines to limit line lengths. So, the following is equivalent to the previous example: 7:0 4 7:6 Ranges are supported. So, the following connects node 6 with nodes 9, 10, 11 and 12: 6:9-12 Order is not significant (except in ranges), so neither lines nor peer numbers need to be sorted. Connections are bidirectional so the following: 1:0 0:1 describes only 1 connection between nodes 0 and 1 and only one of these two lines is required to establish it (though there is no error in specifying both). This is true regardless of whether using a single file or per-node files. An optional line size: N indicates the number of nodes in the job, and is validated against the size of the current job if present. An optional line base: N specifies a numeric base for interpretation of all node numbers on lines that follow. The default is 10 (decimal), and legal values range from 2 (binary) to 36 (uses digits '0'-'9' and 'a'-'z'). If present, the 'base' line only affects node numbers read from later lines, and therefore should appear at the start of the file. Values on the 'size' and 'base' lines are always read as decimal. The default is unset/empty (no limit on which nodes are connected at startup). + GASNET_CONNECTFILE_OUT This setting specifies a filename in which to generate connection information suitable for later use as GASNET_CONNECTFILE_IN. Any '%' character in the value is replaced with the node number to allow separate per-node files. Use of per-node files is strongly recommended, and on some file systems (notably NFS) is REQUIRED for correct operation. If desired, the separate files may be concatenated together after the run completes to produce a single file suitable for use as GASNET_CONNECTFILE_IN. Alternatively, the following perl one-liner will concatenate the files while removing all but the first instance of the 'base' and 'size' lines: perl -ne 'print unless (/(base|size)/ && $X{$_}++);' -- [FILES] where [FILES] denotes the list of per-node connection files and the combined file is generated on stdout. The connection information produced in the output file(s) lists only those connections actually used in the current run. Therefore a common use case is to set GASNET_CONNECTFILE_OUT on a fully-connected run, and then use the generated file(s) to limit static connections in subsequent runs. The default is to use base-36 for node numbers, which results in more compact files but is difficult for a human to read. See GASNET_CONNECTFILE_BASE, below, for how to change this. The default is unset/empty (no output files are generated). + GASNET_CONNECTFILE_BASE This setting controls the numeric base used for node numbers in GASNET_CONNECTFILE_OUT files. Valid values range from 2 (binary) to 36 (uses digits '0'-'9' and 'a'-'z'). The value of the setting is always parsed as base-10. The default value is 36. + GASNET_CONNECT_SNDS + GASNET_CONNECT_RCVS These two settings control the number of small buffers allocated to send and to receive dynamic connection requests, and are ignored if GASNET_CONNECT_DYNAMIC is FALSE, or on any node that is already fully connected at startup. Because the buffers are small and allocation is page granular there is seldom any benefit to reducing the default values. However, there are conditions under which increasing one or both may help reduce the latency of dynamic connections: + Dynamic connection setup is blocking at the initiator, but if using pthreads it is possible that one node may have dynamic connection requests in-progress to multiple nodes. So, if an application is highly-threaded it may be beneficial to increase GASNET_CONNECT_SNDS for greater concurrency of sends. + If a given node receives many simultaneous connection requests, any requests in excess of the allocated buffers will be dropped. The connection will be delayed until the requester retransmits. So, the average connection setup time in the presence of "bursty" requests may be reduced by increasing GASNET_CONNECT_RCVS. The default value of GASNET_CONNECT_SNDS is 4. The default value of GASNET_CONNECT_RCVS is MAX(6, 4 + 2*ceil(log_2(N_remote))) where "ceil()" denotes rounding up to an integer, "log_2()" is the base-2 logarithm and "N_remote" is the number of GASNet nodes minus "self" and any nodes reachable through shared memory (PSHM). + GASNET_CONNECT_RETRANS_MIN + GASNET_CONNECT_RETRANS_MAX These two settings control the minimum and maximum intervals between retransmission of messages used in establishing dynamic connections, and are ignored if GASNET_CONNECT_DYNAMIC is FALSE, or on any node that is already fully connected at startup. Values are in units of microseconds (10^-6 sec). The value of GASNET_CONNECT_RETRANS_MIN is the interval between sending an initial request and the first retransmission. Each retransmission doubles the interval before the next, up to the maximum value given by GASNET_CONNECT_RETRANS_MAX, after which the connection setup fails. Adjustment of these settings may help resolve timeouts on networks with high rates of UD packet loss. However, this is not recommended without consulting with the author and the defaults are therefore not documented here. Software configuration settings: ------------------------------- There are some optional behaviors in ibv-conduit that can be turned ON or OFF. These parameters are permitted to take different values on each process, though doing so may not be useful. However, please see bug 4314 if process-specifc values are needed. https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=4314 + GASNET_RCV_THREAD This gives a boolean: "0" to disable, or "1" to enable, the use of an extra thread per-HCA port to block waiting for an Active Message Request or Reply to arrive. This allows ibv-conduit to remain attentive to incoming AM traffic even while the application is not making any calls to GASNet. The down side is that when this thread wakes it must contend for CPU resources and for locks. Therefore, for an application that is calling GASNet sufficiently often, use of this thread may significantly INCREASE running time. However, on an SMP where an otherwise idle processor is available the use of this thread can REDUCE running time by relieving the application thread of the burden of servicing incoming AM Requests and Replies. Note that if '--disable-ibv-rcv-thread' was specified at build time then the extra thread is unavailable and this environment variable is ignored. Currently the default is disabled (0), but this is subject to change. NOTE: In releases prior to GASNet 1.18.2 the AM receive thread was unavailable for ibv-conduit, but that is no longer the case. + GASNET_RCV_THREAD_RATE If GASNET_RCV_THREAD is enabled, then this setting can be used to impose a limit on how frequently the AM receive thread may wake. This may be used to limit interference between the AM receive thread and the main application thread(s), while providing some network attentiveness when the application is not making GASNet calls. A non-zero value gives the maximum rate in wake-ups per second. The default value is 0, which means no limit is imposed. NOTE: A future release may implement GASNET_RCV_THREAD_LOAD to impose a limit on the *fraction* of time the thread spends awake. + GASNET_RCV_THREAD_POLL_MODE If GASNET_RCV_THREAD is enabled, then this setting determines if/how the AM receive thread participates in the serialization of calls to `ibv_poll_cq()`. This serialization can improve the throughput of clients with multiple threads (the AM receive thread included) by reducing the time spent blocked while polling. The following values (case-insensitive) are recognized: * 'serialized' - The AM receive thread will poll the completion queue for AM arrivals while participating in the same serialization protocol as client threads. (default) * 'unserialized' - The AM receive thread will poll the completion queue for AM arrivals without regard to the serialization protocol observed by client threads. * 'exclusive' - The AM receive thread will utilize the serialization protocol observed by client threads to become the only poller of the completion queue for AM arrivals. The default is "serialized". The value "exclusive" is the only option which directly affects the behavior of application threads. The others have only an indirect affect via the degree of contention they may experience due to the receive thread. If the AM receive thread is disabled (via `'--disable-ibv-rcv-thread` at configure time or via the `GASNET_RCV_THREAD` environment variable), then this setting is ignored. In particular, setting `GASNET_RCV_THREAD=0` and `GASNET_RCV_THREAD_POLL_MODE=exclusive` does not prevent application threads from polling the completion queue for AM arrivals. If serialization of calls to `ibv_poll_cq()` was disabled at configure time via `--disable-ibv-serialize-poll-cq`, then this setting is ignored and the behavior is equivalent to "unserialized". Pinnable memory probe configuration: ----------------------------------- In normal operation of ibv-conduit it is necessary to know how much memory may be registered (aka pinned) with the InfiniBand HCA(s). This is limited by multiple factors and thus cannot be determined by a simple query. Therefore, the default behavior is to attempt to mmap and register as much memory as possible at startup, and then release all the memory. When there are multiple GASNet processes on a shared memory node, one representative process will perform this probe. There are at least two well-known reasons why one may desire to limit or eliminate this probe. The first is the time spent performing the probe. The second is the possibility that the O/S or a batch execution environment may terminate a process that exceeds some limit on the virtual memory size of a process and/or may terminate the process with the largest size when memory is exhausted. Use of the following parameters allows one to bound, or to eliminate, this probe. + GASNET_PHYSMEM_MAX If set, this parameter is used to determine the maximum amount of memory ibv-conduit may pin. This limits how large the GASNet segment can be, and how much memory is available for firehose (see below). The value gives an upper bound on pinnable memory per host, which is divided equally among processes on each host. The value may specify either a relative or absolute size. If the value parses as a floating-point value less than 1.0 (including fractions such as "5/8"), then this is taken as a fraction of the (estimated) physical memory. Otherwise the value is taken as an absolute memory size, with "M", "G" and "T" suffixes accepted to indicate units of Megabytes, Gigabytes, and Terabytes, respectively. This parameter *may* validly differ among processes. However, if the value differs among processes on the same host, the implementation will select just one value per host (the algorithm for the selection is unspecified). The default may be set at configure time using --with-ibv-physmem-max=VALUE, and otherwise is "2/3" (pin up to 2/3 of the estimated physical memory). The following two parameters must be equal across all processes, and the behavior otherwise is undefined. + GASNET_PHYSMEM_PROBE This gives a boolean: "1" to enable or "0" to disable the validation (potentially slow) of the GASNET_PHYSMEM_MAX value. By default the environment variable GASNET_PHYSMEM_MAX is trusted if set at runtime or when using a default given at configure time. If neither source has provided a value, the "2/3" default is taken as a maximum amount of memory that might be possible pinned, but the final limit is determined by probing the limits imposed by the O/S and HCA. This probe can take a significant period of time on large memory nodes. Therefore, enabling this probe may greatly slow startup, but can prevent unexpected runtime failures if the user-provided values exceed those imposed by the O/S and HCA. Therefore, it is recommended to enable this probe if one experiences any runtime failure consistent with an out-of-memory condition. The O/S limits may be hard limits from the kernel (Linux often allows at most 80% of physical memory to be pinned) or from resource limits (see 'ulimit' in a Bourne shell or 'limit' in a C-shell). The default is OFF (probe disabled) if configure or environment sets a value of GASNET_PHYSMEM_MAX. However, this can be changed to a default of ON or OFF by configuring using, respectively, --enable-ibv-physmem-probe or --disable-ibv-physmem-probe. + GASNET_PHYSMEM_WARN This gives a boolean: "1" to enable or "0" to disable the warning printed if/when the GASNET_PHYSMEM_MAX value is probed. Protocol configuration: ---------------------- The following environment variables control the selection of protocols for performing certain transfers. These parameters must be equal across all nodes, and the behavior otherwise is undefined. + GASNET_INLINESEND_LIMIT IBV includes an "inline send" operation that transfers the data to the HCA at the same time it transfers the request. This normally provides a measurable performance improvement, but is only available up to an hardware- and firmware-dependent maximum size. A value of 0 disables use of inline sends. A value of -1 causes use of the maximum value reported by the HCA. The default of 72 is normally correct. + GASNET_PACKEDLONG_LIMIT To perform an AMLong with non-empty payload, ibv-conduit must transfer both the payload and the header. For sufficiently small payloads, it is more efficient (in terms of both CPU overhead and network latency) to pack the header and payload together and copy the payload into place on the target before running the handler. Thus, for payload up to and including this size this packing is used. The default value is the maximum that, together with the maximum sized header, fits into a 4KiB transfer (currently 4012). A value of zero ensures the payload and header always travel separately. + GASNET_NONBULKPUT_BOUNCE_LIMIT This parameter sets the limit on the use of bounce buffers to achieve local completion of "non-bulk" PUT and AMLong payload transfers. When passing GEX_EVENT_NOW to perform a PUT or AMLong, the implementation must block until local completion. For PUTs with nbytes larger than GASNET_INLINESEND_LIMIT, and for AMLongs with nbytes larger than both GASNET_INLINESEND_LIMIT and GASNET_PACKEDLONG_LIMIT, ibv-conduit must either copy the data into bounce buffers, or block until remote completion is signaled by the HCA. Such transfers up to and including size GASNET_NONBULKPUT_BOUNCE_LIMIT are performed using bounce buffers while larger transfers stall return from injection until the RMA is acknowledged. The default value is 64KB. A value of zero disables use of bounce buffers. + GASNET_PUTINMOVE_LIMIT (only for GASNET_SEGMENT_{LARGE,EVERYTHING}) When the firehose algorithm (see below) is in use for managing the pinning of remote memory, a PUT that misses in the firehose cache may be accelerated by piggybacking data on the AMMedium that is used to obtain a remote pinning. The value of GASNET_PUTINMOVE_LIMIT is the maximum number of bytes to send in this way. The value is bounded by the maximum value set at compile time, and it is an error to request a larger value. Note that in a GASNET_SEGMENT_FAST configuration, the remote segment is pinned statically and this optimization is never applicable. The default value is 3KB (the current maximum value). A value of zero disables this optimization. + GASNET_PUT_STRIPE_SZ and GASNET_GET_STRIPE_SZ (experimental) When multiple HCA ports are used, the performance of a sufficiently large isolated RMA operation (one not overlapped with other communication) can be increased by subdividing it into multiple pieces striped over more than one path. These parameters specify the threshold above which striping is applied to RMA Put and Get operations, respectively. Suffixes "K", "M" and "G" can be used to specify units of Kilobytes, Megabytes and Gigabytes, respectively. The default units are Kilobytes. Use of a too-small value may limit the performance of large RMA operations by subdividing them into stripes too small to saturate the network. Use of a too-large value may limit the performance of RMA operations which are large enough to benefit from striping but below the value. Values exceeding the HCA's maximum transfer size will be silently reduced. These parameters are ignored if only a single HCA/port is in use. A value of zero uses the HCA's maximum transfer size as the stripe size, effectively disabling this optimization. The current default is zero (disabled). + GASNET_AM_GATHER_MIN This parameter sets the minimum payload size at which a multi-segment gather may be used to concatenate an AM header and payload. Below this minimum payload data is copied. The default value is 1500. A value of -1 disables multi-segment gather, using payload data copy at all payload sizes. + GASNET_USE_SRQ This controls whether IBV Shared Receive Queue (SRQ) support is used, but is ignored if GASNet was configured with --disable-ibv-srq. This setting defaults to -1, which means that SRQ will be used only if doing so would reduce memory usage (as determined from the value of the GASNET_RBUF_COUNT setting, described below). If set to a non-negative value, this setting give the minimum GASNet node count at which SRQ will be used, regardless of whether or not the memory usage would increase or decrease. A value of zero will disable SRQ. Examples: - GASNET_USE_SRQ unset or explicitly set to -1: SRQ is used ONLY if GASNET_RBUF_COUNT is less than the number of receive buffers required for the non-SRQ case. - GASNET_USE_SRQ <= job size [includes GASNET_USE_SRQ == 1] SRQ is used and GASNET_RBUF_COUNT is enforced as a maximum - GASNET_USE_SRQ > job size SRQ is NOT used and GASNET_RBUF_COUNT is ignored Note that the interpretation of the values 0 and 1 allow one to use this setting as a simple boolean if desired. + GASNET_USE_XRC This controls whether IBV eXtended Reliable Connection (XRC) support is used. However, it is is ignored if GASNet was configured with --disable-ibv-xrc, if XRC support was not found at configure time, or if SRQ support is not used (regardless of why), This setting defaults to 1 if SRQ support was enabled at configure time. As a result XRC will be used anytime SRQ is used. + GASNET_USE_ODP This boolean setting controls whether IBV On-Demand Paging (ODP) support should be used. If true, then ibv-conduit uses ODP on any compute nodes where the IBV library reports support is available. When ODP is NOT available on one or more compute nodes, a warning is issued. See GASNET_ODP_VERBOSE for information on controlling this warning. This setting is ignored if GASNet was configured with --disable-ibv-odp, or if ODP support was not found at configure time. This setting defaults to 1 if ODP support was enabled at configure time. + GASNET_ODP_VERBOSE This integer setting controls the detail of the warning printed when ODP support is enabled but is not present on one or more HCAs in the job. A value of 0 supresses the warning entirely. A value of 1 will report a count of processes lacking support. Values of 2 or higher give increasing levels of detail concerning the missing support. This setting defaults to 1 if ODP support was enabled at configure time. + GASNET_USE_FENCED_PUTS This boolean setting controls the use of atomic operations to provide for correct remote completion detection in the presence of multiple HCAs. See "Bug 3447" in the Known Problems section for information on when one may wish to enable this setting. If enabled when multirail support was not enabled at configure time, a warning will be issued. The default can be set at configure time via '--with-ibv-fenced-puts=...' and is 0 (disabled) in the absence of that configure option. Resource usage parameters: ------------------------- The following environment variables control how much memory is preallocated at startup time to serve various functions. Because these resource pools do not grow dynamically, it is important that these parameters be sufficiently large, or performance degradation may result. The default settings should be sufficient for most conditions. You may need to lower some values if you have insufficient memory. + GASNET_RBUF_SPARES This gives the number of AM receive buffers used to hold header and payload of executing AM Request, and thus bounds the number of threads which may concurrently execute AM handlers "in place". Any threads beyond this limit must copy the header and payload before executing the handler. The default value reflects a heuristic estimate of the number of threads which might concurrently poll for AM arrivals. Reducing this parameter may reduce Active Message throughput. The following parameters must be equal across all nodes, and the behavior otherwise is undefined. + GASNET_NETWORKDEPTH_PP This gives the maximum number of ops (RDMA + AMs) which can be in-flight simultaneously from a node to each of its peers. Here "in-flight" means queued to the send work queue and not yet reaped from the send completion queue. This value is the depth of each send work queue. This limit is on the number of ibv-level ops in-flight, and the number of GASNet-level operations may be less (for example, when the length of a PUT or GET is larger than the HCA's maximum message length, or because an AM Long uses separate ops for the payload and header). The default value is 24. Reducing this parameter may limit small message throughput. If you believe your small message throughput is too low, you may try increasing this value. + GASNET_NETWORKDEPTH_TOTAL This gives the maximum number of ops (RDMA + AMs) which can be in-flight simultaneously from each node (with "in-flight" defined as in GASNET_NETWORKDEPTH_PP). The depth of the send completion queue is min(GASNET_NETWORKDEPTH_TOTAL, GASNET_NETWORKDEPTH_PP*(N-1)). If set to zero, the value is set to the maximum usable value computed from GASNET_NETWORKDEPTH_PP and the HCA's reported capabilities. The default value is 255. Reducing this parameter may limit small message throughput. If you believe your small message throughput is too low, you may try increasing this value (or setting it to zero), at a cost in additional memory consumption. + GASNET_AM_CREDITS_PP This give the maximum number of AM Requests which can be in-flight simultaneously from a node to each of its peers. Here "in-flight" means the Request is queued to the send work queue, but the matching Reply has not yet been processed for AM flow control (described in another section of this README). This is the number of buffers which must be preposted to each receive work queue for AM Requests. The default value is 12 (12*MaxMedium*(N-1) allocated for Request buffers). Reducing this parameter may limit Active Message throughput. If you believe your Active Message throughput is too low, you may try increasing this value. + GASNET_AM_CREDITS_TOTAL This gives the integer number of AM Requests which can be in-flight simultaneously from each node, with "in-flight" defined as in GASNET_AM_CREDITS_PP. If set to zero, the value is set to the maximum usable value computed from GASNET_AM_CREDITS_PP and the HCA's reported capabilities. The default value is MIN(256, (nodes-1)*GASNET_AM_CREDITS_PP). Reducing this parameter may limit Active Message throughput. If you believe your Active Message throughput is too low, you may try increasing this value (or setting it to zero), at a cost in additional pinned memory. + GASNET_AM_CREDITS_SLACK This gives the maximum number of flow-control credits that can be delayed at the responder. If a Request handler does not produce a Reply, a credit may be "banked" to be piggy-backed on the next Request or Reply headed to the requesting node. The value of GASNET_AM_CREDITS_SLACK gives the maximum number of credits that can be banked before a hidden Reply is generated to convey credits back to the requester. The default value is 1. GASNET_AM_CREDITS_SLACK will be silently reduced if needed to ensure deadlock will not occur, and is ignored when SRQ is used. Reducing this parameter to zero or setting it too high may increase the latency of Active Message traffic. + GASNET_RBUF_COUNT If SRQ support is unavailable or disabled, this parameter is ignored. See GASNET_USE_SRQ documentation for details of when SRQ is enabled. When SRQ is enabled this gives the max number of AM receive buffers allocated on each node. These buffers are needed for reception of AM headers and the payload of mediums, but are not used for RDMA. The actual number of buffers allocated is the lesser of the value of GASNET_RBUF_COUNT or a value computed from the GASNET_AM_* and GASNET_NETWORKDEPTH_* parameters described above. If set to zero, the value is limited only by the HCA's capabilities. The default value is 1024 (up to 1024*MaxMedium for buffers). Reducing this parameter may limit Active Message throughput. If you believe your Active Message throughput is too low, you may try increasing this value (or setting it to zero), at a cost in additional pinned memory. + GASNET_BBUF_COUNT This gives the max number of pre-pinned buffers allocated on each node. These buffers are needed for assembly of AM headers and the payload of mediums, and for some PUTs (see GASNET_NONBULKPUT_BOUNCE_LIMIT). The actual number of buffers allocated is the lesser of the values of GASNET_BBUF_COUNT and GASNET_NETWORKDEPTH_TOTAL, since the total network depth bounds the number of in-flight operations that might need these buffers. If set to zero, the value is set to GASNET_NETWORKDEPTH_TOTAL. The default value is 1024 (up to 1024*MaxMedium for buffers). Reducing this parameter limits the number of in-flight operations which consume bounce buffers. This includes AMs too large for an inline send and PUTs subject to the GASNET_NONBULKPUT_BOUNCE_LIMIT. If you believe that throughput of these operations is too small, you may try increasing this value (or setting it to zero), at a cost in additional pinned memory. + GASNET_PINNED_REGIONS_MAX This provides a limit on the number of pinned regions to be created. Similar to GASNET_PHYSMEM_MAX, the value gives an upper bound on pinned regions per host, which is divided equally among processes on each host. This may constrain dynamic registration via firehose (below). The value may specify either a relative or absolute size. If the value parses as a floating-point value less than 1.0 (including fractions such as "5/8"), then this is taken as a fraction of the maximum supported region count reported by the HCA(s). Otherwise the value is taken as an absolute region count. The default is to use a fraction of the HCA pinning resources equal to the fraction of physical memory given by GASNET_PHYSMEM_MAX, subject to a system-dependent maximum value. Firehose configuration: ---------------------- These parameters must be equal across all nodes, and the behavior otherwise is undefined. The following environment variables control the per-process resources used by the "firehose" [ref 1] dynamic registration library. By default, firehose will use as much pinned memory as the HCA and O/S will permit, bounded by GASNET_PHYSMEM_MAX. Resource use is divided into two pools. The main pool is for managing of pinning of the GASNet segment on remote nodes, while the "victim" pool is used to manage pinnings for local use. By default in a GASNET_SEGMENT_LARGE or GASNET_SEGMENT_EVERYTHING configurations, 75% of the pinnable memory will go in the main pool and 25% into the victim pool. In a GASNET_SEGMENT_FAST configuration, firehose is not needed for management of the statically pinned GASNet segment, and by default only a small fraction of the available memory is placed in the main pool for internal uses and the majority is placed in the victim pool. + GASNET_USE_FIREHOSE This environment variable is only available in a DEBUG build of GASNet (one configured with --enable-debug). This gives a boolean: "0" to disable or "1" to enable the use of the firehose dynamic pinning library in a GASNET_SEGMENT_FAST configuration. In a GASNET_SEGMENT_FAST configuration, the GASNet segment is registered (pinned) with the HCA at initialization time, because pinning is required for RDMA. However, GASNet allows for local addresses (source of a PUT or destination of a GET) to lie outside of the GASNet segment. So, to perform RDMA GETs and PUTs, ibv-conduit must either copy out-of-segment transfers though preregistered bounce buffers, or dynamically register memory. By default firehose is used to manage registration of out-of-segment memory. (default is ON). Setting this environment variable to "0" (or "no") will disable use of firehose, forcing the use of bounce buffers for out-of-segment transfers. This will result in a significantly lower peak bandwidth for large PUTs and GETs, with little or no effect on small message latency. It is available only for debugging purposes. In a GASNET_SEGMENT_LARGE or GASNET_SEGMENT_EVERYTHING configuration, the GASNet segment is not preregistered and use of firehose is required. Thus it is an error to disable firehose in such a configuration. + GASNET_FIREHOSE_M and GASNET_FIREHOSE_MAXVICTIM_M GASNET_FIREHOSE_M gives the amount of memory to place in the main pool, while GASNET_FIREHOSE_MAXVICTIM_M gives the amount of memory to place in the victim (local) pool. The suffixes "K", "M" and "G" are interpreted as Kilobytes, Megabytes and Gigabytes respectively, with "M" assumed if no suffix is given. When neither variable is set, the defaults are respectively 75% and 25% of the total pool. In a GASNET_SEGMENT_LARGE or GASNET_SEGMENT_EVERYTHING configuration, this pool's size is the maximum pinnable memory (but see below), while in a GASNET_SEGMENT_FAST configuration it is the same size as the prepinned bounce buffer pool. Note that, as used here, "maximum pinnable memory" may be less than determined from GASNET_PHYSMEM_MAX, and in particular may be constrained by the product of the number of pinnable regions and their maximum size. See, GASNET_FIREHOSE_MAXREGION_SIZE, GASNET_FIREHOSE_R and GASNET_FIREHOSE_MAXVICTIM_R for more information. If only one of these variables is set, then the other defaults such that their sum equals the total pool size. Therefore, to enlarge or reduce the total pool, one must set both. Since enlarging the total pool risks exhausting resources, potentially leading to crashes at runtime, doing so will result in a warning. + GASNET_FIREHOSE_R and GASNET_FIREHOSE_MAXVICTIM_R GASNET_FIREHOSE_R gives the maximum number of pinned regions to allocate for the management of the main pool, while GASNET_FIREHOSE_MAXVICTIM_R gives the maximum number of pinned regions to allocate for the management of the victim (local) pool. When neither variable is set, the default is to split the available pool of pinnable regions (see GASNET_PINNED_REGIONS_MAX) in proportion to the values of GASNET_FIREHOSE_M and GASNET_FIREHOSE_MAXVICTIM_M. If only one of these variables is set, then the other defaults such that their sum equals the total pool size. Therefore, to enlarge or reduce the total pool, one must set both. Since enlarging the total pool risks exhausting resources, potentially leading to crashes at runtime, doing so will result in a warning. The value of GASNET_FIREHOSE_R will be silently truncated if larger than (GASNET_FIREHOSE_M / GASNET_FIREHOSE_MAXREGION_SIZE), since additional regions would not be used. Similarly, GASNET_FIREHOSE_MAXVICTIM_R will be silently reduced if it would address more than GASNET_FIREHOSE_MAXVICTIM_M. + GASNET_FIREHOSE_MAXREGION_SIZE This gives the maximum size of a single dynamically pinned region, should be a multiple of the pagesize, and preferably a power of two. The suffixes "K", "M" and "G" are interpreted as Kilobytes, Megabytes and Gigabytes respectively, with "M" assumed if no suffix is given. The maximum addressable size of the main and victim pools are limited by the product of this region size and the number of firehose regions allocated to each pool. If the value of this parameter is set to 0, then it will be automatically adjusted to allow the main and victim pools to be addressed within the available number of regions, if doing so is possible subject to a system-dependent maximum (pagesize squared, or larger). The default value of this parameter is 128KB. + GASNET_FIREHOSE_TABLE_SCALE This parameter gives a floating point factor, used to scale the size of hash tables used in the firehose library relative to their default sizes. Smaller values producer smaller tables, saving memory at the expense of performance. The default value is 1. + GASNET_FIREHOSE_VERBOSE This gives a boolean: "0" to disable or "1" to enable the output of internal information of use to the developers. You may be asked to run with this environment variable set if you report a bug that appears related to the firehose algorithm. External library settings: -------------------------- These parameters must be equal across all nodes, and the behavior otherwise is undefined. + MLX4_SINGLE_THREADED and MLX5_SINGLE_THREADED In a SEQ or PARSYNC build, ibv-conduit will set these environment variables to '1' under appropriate conditions (which include the default configure and environment settings) to instruct libibverbs to elide locking. These variables influence the libibverbs implementations for, respectively, the "mlx4" and "mlx5" drivers. Here "appropriate conditions" means that ibv-conduit is not starting any asynchronous threads. In particular, these variables are *not* set by the conduit if use of GASNET_RCV_THREAD has requested the asynchronous AM receive thread, or if the combination of GASNET_CONN_* settings may require starting a thread to receive dynamic connection requests. For more information on these variables, see their respective documentation, above. In all cases, a user's setting of either variable is preserved, allowing an explicit setting of '0' to prevent the conduit from setting it to '1'. @ Section: Multi-rail Support @ Multi-rail support is OFF in GASNet ibv-conduit by default. By default, ibv-conduit will use only the first active port on the first active InfiniBand Host Channel Adapter (HCA). However, if more than one HCA port is enabled for use, ibv-conduit will stripe communications over them. See the sections "Build-time Configuration" and "Runtime Configuration" for information on how to enable use of more HCAs/ports, or to control which HCAs/ports are used. To first order, the use of multiple ports or multiple adapters will yield increases in both bandwidth (good) and software overhead (bad). How the resulting trade off works for a given application may be hard to predict. If one is concerned with obtaining the maximum possible performance for a given application, then experiment with the GASNET_NUM_QPS and/or GASNET_IBV_PORTS environment variables (documented in "Runtime Configuration") to determine how a given application runs best. IMPORTANT NOTE: The multi-rail support in ibv-conduit makes the assumption that the number of HCAs used in every process will be identical. One must use GASNET_IBV_PORTS to ensure that this property is true. Otherwise, the behavior is undefined (startup crash being the most likely). @ Section: On-Demand Paging (ODP) Support @ Recent Mellanox HCAs (ConnectX-4 and newer) support a feature known as On-Demand Paging (ODP). Where this support is available, ibv-conduit can use it to reduce resource pressure on physical memory. By default, ibv-conduit will attempt to warn at runtime if this support does not appear to be available despite having either the required software or hardware. If you are using Mellanox Connect4-X HCAs or newer and ibv-conduit warns about missing ODP support, we recommend that you install the latest "Mellenox OFED" (aka MLNX_OFED) distribution, available for download from Mellanox. Mellanox provides documentation on installing the software and upgrading the HCA firmware (if needed). At configure time, ibv-conduit does not know the hardware, firmware, or driver versions on the compute nodes and cannot, in general, assume the host running configure represents every node in the system. Therefore, ODP support is enabled by default whenever the necessary library support is detected by the configure script. If ibv-conduit warns about missing ODP support on a system in which there are no ODP-capable nodes, then we recommend reconfiguring GASNet using --disable-ibv-odp to both eliminate the warnings and avoid the small overhead of ODP support. However, in a heterogeneous system in which some subset of compute nodes do support ODP, one should set GASNET_ODP_VERBOSE=0 in the environment to suppress warnings from nodes lacking ODP support, while allowing ibv-conduit to continue using ODP on the remaining nodes. @ Section: HCA Configuration @ GASNet ibv-conduit should *not* require any specialized configuration of your HCAs to achieve normal, correct operation. However, this section documents any configuration that *may* help improve performance. We recommend you backup your configuration data prior to attempting any modification, and that you confirm that any changes made produce a measurable benefit before deciding to keep them. If trying a suggestion here results in no measurable improvement, then we recommend that you return the modified parameter(s) to their previous value(s). WE DISCLAIM ALL RESPONSIBILITY IF FOLLOWING ANY SUGGESTION HERE RESULTS IN AN UNSTABLE OR UNUSABLE SYSTEM. Please consult the documentation provided with your HCA drivers, and/or your vendor or system integrator for information on how to query or change your HCA's configuration parameters. + The HCA configuration parameter MAX_QP_OUS_RD_ATOM controls the number of simultaneous RDMA Reads for which a QP may act as Responder. Our testing on one system with a default value of 8, showed that increasing the value to 16 yielded approximately a 30% bandwidth improvement in an RDMA-GET benchmark. @ Section: Advice to Client Authors @ + Negotiated-payload Active Messages (NPAM) TL;DR: The results below, from one particular system, can be summarized in the following rules-of-thumb for use of NPAM with ibv-conduit in the *current* release: + Use of client-provided buffer is never advantageous. + Use of gasnet-provided buffer may be advantageous for Medium with sufficiently large payloads, where both latency and bandwidth can exceed those of FPAM. + Use of gasnet-provided buffer may be advantageous for Long payloads of sufficient length, where bandwidth is better than FPAM (but at the expense of worse latency). Note that as development continues, these finding are subject to change. Calls to gex_AM_Prepare{Request,Reply}{Medium,Long}() with client_buf != NULL are known as "client-provided buffer" calls. In this mode of operation, there is a small penalty in CPU overhead relative to the fixed-payload AM (FPAM) calls gex_AM_{Request,Reply}{Medium,Long}(), due primarily to the split-phase calling convention. While the design of NPAM allows for the possibility that NPAM with client-provided buffer could enable larger Medium payloads than FPAM, ibv-conduit currently does not provide that capability. Measurements of both AM Mediums and Long with client-provided buffer NPAM on OLCF's Summit show the latency penalty relative to FPAM in a Request/Reply "ping-pong" test is around 2% for payload sizes below a couple hundred bytes, and 1% or lower for payloads of 512 bytes or larger. For AM Longs the penalty eventually approaches zero for payloads of about 512KiB or larger. Throughput of a "flood" test with client-provided buffer NPAM shows penalties of about 8% at the smallest payload sizes, declining smoothly to 5% at the largest Medium payloads and approaching zero for Long payloads of about 512KiB or larger. Calls to gex_AM_Prepare{Request,Reply}{Medium,Long}() with client_buf == NULL are known as "gasnet-provided buffer" calls. In this mode of operation, in which GASNet allocates a buffer where client code assembles/generates the payload at AM injection time, there is a measurable advantage to NPAM for sufficiently large payloads, but a small penalty for small payloads. Measurements with gasnet-provided buffer NPAM on OLCF's Summit show the latency penalty in a Request/Reply "ping-pong" test is around 2% for payload sizes below a couple hundred bytes for both Medium and Long. For Mediums of 512 bytes or larger, the latency is improved over FPAM (by about 9% for large payloads). For Longs, the large payload latency is worse than for small payloads. Throughput of a "flood" test with Mediums shows similar behavior to ping-pong, with a throughput penalty of up to 5% for payload sizes 512 bytes and below, but improvements in throughput above 512 bytes (by about 17% for large payloads). For Longs there is a throughput penalty of up to 5% for payload sizes below about 2KiB, but for large payloads a latency improvement of 40% or more can be seen. Your mileage may vary. Relative performance may change in future releases. @ Section: Known Problems @ + Slow PHYSMEM probe at start-up As described in more detail above, the environment variables GASNET_PHYSMEM_MAX and GASNET_PHYSMEM_PROBE can be used to control an upper-bound on the amount of memory that ibv-conduit will attempt to register/pin. If configure was passed --with-ibv-physmem-max=VALUE, then the given VALUE is used as the default value of GASNET_PHYSMEM_MAX. However, if this environment variable is not set, and no value was given at configure time for GASNET_PHYSMEM_MAX, then the values "2/3" and "yes" are used for these two environment variables and a message will direct the user to this text. Running with GASNET_PHYSMEM_PROBE=1 will also direct one to this text. If all compute nodes probed are found to allow the same amount of memory to be pinned (either in absolute or relative terms), the message issued will indicate the value determined and how to use it at configure or run time. If there is a single memory configuration in use on your system, then configuring (or setting GASNET_PHYSMEM_MAX) as directed in that message will eliminate the time spent on the probe at startup, as well as the message. If the memory configuration varies among nodes on the system, it may not be safe to use a single recommended setting. There are at least three options available: 1. To avoid the probe at the potential cost of registering less than the maximum possible memory for communication, one can elect to use the minimum probed value reported by the probe output across all of the node configurations to be used. Depending of the actual configurations using an absolute (whole number) or relative (fraction or decimal) value may result in using nearer to the maximum memory on all nodes. The most robust option would be to configure using --with-ibv-physmem-max=VALUE, but running with GASNET_PHYSMEM_MAX=VALUE may be simpler. Unless one has configured with --enable-ibv-physmem-probe, or is running with GASNET_PHYSMEM_PROBE=1 in the environment, the given value will be used without running the probe. 2. If the time spent by the probe is not a concern, one can obtain the behavior (default in older releases) of probing for UP TO a (safe) default of 2/3 of apparent physical memory size on every run (without any message) by running with environment variable GASNET_PHYSMEM_WARN=0. 3. If 2/3 is too conservative, one can force a probe with an alternative value by configuring using --with-ibv-physmem-max=VALUE --enable-physmem-probe of running with GASNET_PHYSMEM_MAX=VALUE GASNET_PHYSMEM_PROBE=1 In either case, one may also want to set GASNET_PHYSMEM_WARN=0. + Crashes have been seen using QLogic's InfiniPath HCAs with ibv-conduit with default parameters. If you see crashes with a message containing FATAL ERROR: aborting on reap of failed send then we recommend setting the following two environment variables GASNET_NETWORKDEPTH_PP=8 GASNET_QP_RD_ATOM=1 In our testing this resulted in about a 2% reduction in peak bandwidth, but eliminated all instances of "aborting on reap of failed send". + Lack of XRC support XRC is a optional feature in the Open Fabrics Verbs API, for which GASNet's configure script will probe support. However, that probe is limited to determining if the XRC-related function calls will compile and link, and cannot distinguish platforms on which the calls are present but always fail. On such systems, the failure will be reported at runtime above a certain number of processes (86 with the defaults for all environment variables) with the following message: *** FATAL ERROR: Unable to create an XRC domain. Please see "Lack of XRC support" under Known Problems in GASNet's README-ibv. If you experience this error it is recommended that you reconfigure your build of GASNet with --disable-ibv-xrc. If that is not possible one can also set GASNET_USE_XRC=0 in your environment. + Bug 495 and 955 The "firehose" implementation of dynamic memory registration in ibv-conduit is susceptible to incorrect behavior if it caches registration information for a page of virtual memory that is returned to the OS by munmap() and the same virtual address is allocated to the process again later. To avoid this in most cases, ibv-conduit defaults to GASNET_DISABLE_MUNMAP=1 (documented in the top-level README) on 64-bit systems. However, this can be undermined by the GASNet client in a number of ways. A (non-exhaustive) list of known potentially problematic behaviors include: + Calls to `mallopt(M_TRIM_THRESHOLD, X)` for X != -1. + Setting environment variable MALLOC_MMAP_THRESHOLD_ other than to -1. + Calls to `mallopt(M_MMAP_MAX, Y)` for Y != 0. + Setting environment variable MALLOC_MMAP_MAX_ other than to 0. + Replacing the glibc implementation of malloc/free with one that will munmap() memory at any time other than process exit. If your GASNet client code exhibits any of the above-listed behaviors, you should explicitly set GASNET_DISABLE_MUNMAP=0. This not only prevents ibv-conduit from applying its default of 1, but additionally (and more importantly) modifies other behaviors within GASNet to avoid memory management behaviors that could trigger the bug. However, this setting cannot protect client code which may use malloc()ed memory as the source or destination of GASNet communication calls if such memory is later returned to the OS via munmap(). The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=495 That bug report describes the problem in more detail, and lists the best known recommended work-around(s). + Bug 3447 In this release the implementation of ibv-conduit on multi-HCA InfiniBand networks may yield incorrect results for certain communication patterns. However, the environment variable GASNET_USE_FENCED_PUTS (described above) can correct for this problem at the expense of higher latency and reduced bandwidth for Put-based communications due to additional communication. The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=3447 That bug report describes the conditions under which the problem may manifest, and the recommended work-arounds. + Bug 3693 In some circumstances, one may see the following message at startup: *** FATAL ERROR: Unexpected error Bad address (errno=14) when registering the segment If your system is configured to allow large SystemV shared memory segments, then this can be resolved by switching to SystemV instead of POSIX for GASNet's shared memory support. However, it is also possible that one just needs to increase the systems limit on POSIX shared memory allocations as described under "System Settings for POSIX Shared Memory" in GASNet's top-level README. The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=3693 That bug report describes more about the problem, and provides more detail regarding the recommended work-around(s). + Bug 3816 When using Mellanox On-Demand Paging (ODP), it is possible that abnormal exits (SIGKILL or _exit(), in particular) may lead to leaking memory at the system level, requiring a reboot to reclaim the lost memory. In extreme cases, the cumulative effect of many leaks could eventually lead to the Linux Out-Of-Memory killer being invoked, or even a kernel panic. If you use ODP and see evidence of available memory declining over time, and have reason to suspect a GASNet application is experiencing an abnormal exit, then it may be advisable to disable ODP (see On-Demand Paging (ODP) Support section). The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=3816 That bug report describes more about the problem, and provides more detail regarding the recommended work-around(s). + Bug 3838 When using Mellanox On-Demand Paging (ODP), it has been observed that certain patterns of many small RMA operations (as seen in GASNet's testvis) may lead to communications "locking up" such that RMA operations issued do not complete and eventually new RMA operations cannot be initiated. If you use ODP and experience hangs from an application with many fine-grained RMA operations, then you should retry with ODP disabled (see On-Demand Paging (ODP) Support section). The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=3838 That bug report describes more about the problem, and provides more detail regarding the recommended work-around(s). + Bug 3997 When using ibv-conduit on Linux/AARCH64 (aka ARM64 or ARMv8) systems, it has been observed that some traffic patterns may lead to crashes of the IB HCA or of the compute node. The most up-to-date information on this bug is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=3997 That bug report describes more about the problem, and provides more detail regarding the recommended work-around(s). + Bugs 4008 and 4009 On some systems, RMA Puts and AM Longs with source buffers on read-only pages (such as those that might be generated for const-qualified static variables) are not handled correctly. The most up-to-date information on this issue is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=4008 (ODP) https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=4009 (Solaris) Those bug reports provide more details, and will provide the most current information regarding any recommended work-around(s). + Bug 4314 While this file documents several environment variables as permitting different values on different processes, it is not always possible to achieve such a scenario. The most up-to-date information on this issue is maintained at: https://gasnet-bugs.lbl.gov/bugzilla/show_bug.cgi?id=4314 + See the GASNet Bugzilla server for details on other known bugs: https://gasnet-bugs.lbl.gov/ @ Section: Core API @ + Flow-control for AMs. The AMs in ibv-conduit are just implemented as send/recv traffic. Therefore a send without a corresponding recv buffer preposted at the peer will be stalled by the RNR (receiver-not-ready) flow control in IB. However there are two reasons why we want to avoid this situation. The first is that if such a send is blocked by flow control, then the ordering semantics of IB tell us that all the gets and puts that we've initiated after the AM was sent are also stalled. Rather than let that happen, we should manually delay those which are dependent on the AM. The second reason is that under some conditions the RNR flow control is very poor. The problem is that once the intended receiver sends a RNR NAK to indicate no available recv buffers, IB has the SENDER's hardware/firmware poll for the receiver to become ready again! That leaves us with a choice between configuring a small polling interval and consuming a lot of bandwidth for this polling, or a large interval which leads to performance which is degraded more than necessary when IB flow control is asserted. For these reasons we implement some flow control at the AM level. The basic idea is that every REQUEST consumes one credit on the sending endpoint and every REPLY grants one credit on the receiving endpoint. Thus if M is the initial number of credits on each endpoint and every REQUEST has exactly one matching REPLY, then M becomes a limit on the number of un-acknowledged REQUESTS in flight on an endpoint. If we want to avoid RNR conditions, then we should start with M credits and M preposted recv buffers on each endpoint. This allows for only the receipt of M REQUESTS. In addition, a recv buffer will be posted on demand for a REPLY just before sending each REQUEST. It is a simple matter to count the credits when a REPLY is received and to poll for credits when needed to send a REQUEST. It is also simple to ensure the exactly-one-reply. We already ensure that at-most-one reply is sent by the request handler. Additionally we must check upon handler return for the case that the request handler sent no reply, and send one implicitly. We just use a special "system category" handler, gasnetc_SYS_ack, which doesn't even run a handler. To avoid using up 1/2 our bandwidth in the event of a REQUEST-REQUEST ping-pong, we perform some coalescing to avoid sending too many SYS_ack REPLIES. We keep up to GASNET_AM_CREDITS_SLACK "banked" on the responding node, sending the SYS_ack REPLY only if the number banked exceeds this limit. Credits which are banked get piggybacked on the next REQUEST or REPLY headed back to the original requester. To avoid a window of time between when we send a REPLY (credit) and when we post the recv buffer, we must post the replacement recv buffer BEFORE running an AM REQUEST handler. To do this we keep a pool of unposted recv buffers (also used for the on-demand posting of buffers needed for REPLIES). So, when we recv an AM REQUEST, we grab a free recv buffer from the pool and post it to the endpoint, and only then run the handler. We send an implicit reply if a REQUEST handler didn't send any REPLY. Finally we take the recv buffer containing the just-processed AM and we return it to the unposted pool. There is a corner case we must deal with when there are no spares left in the unposted pool. In this case we will copy the received REQUEST into a temporary (non-pinned) buffer before processing it. This allows us to repost the recv buffer immediately. Since the temporary buffer is not pinned, it cannot be used for receives. Therefore, we free the temporary buffer when the handler is done, rather than placing it in the unposted pool. If we reap multiple AMs in a single Poll, then we reuse the previous buffer as the "spare" for the next one, in place of grabbing one from the unposted pool each time. Thus, we touch the unposted pool at most twice per Poll, once for the first AM we receive and once at the end to put the recv buffer of the final AM back in the unposted pool. For the dedicated receive thread we can do even better, never touching the unposted pool at all, by always keeping a single thread-local "spare", initially acquired at startup. Note that when SRQ is used, no flow control is used. @ Section: Extended API @ [This section is still *mostly* accurate, but has not been kept up-to-date with respect to EX updates. Most notably: "LongAsync" is gone, "_bulk" has been supplanted by lc_opt options, and all the function names have changed.] Notes for myself for extended API: + The send completion facility consists of two pointers to counters, associated with each sbuf. If these pointers are non-NULL then the counter is decremented atomically when the send is complete. One counter is for awaiting reuse of local memory and is only be used for sbufs which are doing zero copy. This counter provides the mechanism for Longs and non-bulk puts to block before they return, and should be allocated as an automatic variable. The second counter is for request completion and should be non-NULL for every sbuf for which request completion would be checked (all gets & puts, but not the Longs). For nb and nbi the counter is waited on at sync-time. Therefore the explicit event is a struct containing the counter. + Similar to the reference implementation's cut-off between Mediums (which typically do a source-side copy) and Longs (which may not), we have a cut-off size, below which the RDMA-put operation will do source-side copies _iff_ local completion is desired (Long, put_nb, and put_nbi). + The gets are done w/ RDMA-reads, and use the sbuf bounce buffers if the local memory is not in the segment (or otherwise registered). The value gets also pass though the bounce buffers. Clearly there is no bulk/non-bulk distinction in terms of local memory reuse, just the alignment and optimal size distinctions. So, only the outstanding request counter on the sbuf is needed for syncs of all types of gets. + Table of when synchronization is needed Local Remote Operation Sync Sync -------------------------- LongAsync X X Long I X put_nb I S put_nbi I S put_nb_bulk X S put_nbi_bulk X S put_nb_val X S put_nbi_val X S put X I put_bulk X I put_val X I get_nb X S get_nbi X S get_nb_bulk X S get_nbi_bulk X S get X I get_bulk X I get_val X I X = Not needed at all (or not even applicable with _val forms) I = Needed before (I)nitiating function returns S = Needed before (S)ynchronizing function returns + Some minor tweaks are used to avoid allocation of counters in some cases. - For all the functions which require waiting on a counter in the initiating function, the counter can be allocated on the stack (as an automatic variable). - For the implicit-handle forms the request counter is in the thread-specific data, possibly in an access-region. - For the explicit event forms the request counter must be allocated from some pool, requiring some memory management work. This is done with a modification to the code from the reference implementation, and uses thread-local data to avoid locks. + The memsets can be more efficiently implemented as a _local_ memset followed by a PUT, for small enough sizes. This is not currently implemented. @ Section: GEX_FLAG_IMMEDIATE Support The following summarizes the current state of support for GEX_FLAG_IMMEDIATE. Last reviewed against 2022.3.0 | SEGMENT | SEGMENT | SEGMENT | OPERATION | FAST | LARGE | EVERYTHING | ------------+-------------+------------+------------+ FPAM Short | Y | Y | Y | FPAM Medium | Y | Y | Y | FPAM Long | Note 1 | Note 1 | Note 1 | | | | | NPAM Medium | Y | Y | Y | NPAM Long | Note 2 | N | N | | | | | RMA Put | N | N | N | RMA Get | N | N | N | ------------+-------------+------------+------------+ Y = GEX_FLAG_IMMEDIATE is fully implemented N = GEX_FLAG_IMMEDIATE is ignored FPAM = Fixed-payload AM NPAM = Negotiated-payload AM For the notes which follow, define a "packed Long" as an AM Long having a payload size of GASNET_PACKEDLONG_LIMIT or less. Note 1: FPAM Long + For a packed Long the IMMEDIATE flag is fully implemented. For the LARGE and EVERYTHING segment modes, this notably includes all Reply calls because in those cases the payload is always packed. + Otherwise, the IMMEDIATE flag is ignored. Note 2: NPAM Long / FAST + For a packed Long the IMMEDIATE flag is fully implemented. This notably includes all calls with a GASNet-allocated buffer because in this case the payload is always packed. + Otherwise, the IMMEDIATE flag is *partially* implemented. The logic in Prepare is sensitive to AM flow-control credits and buffer allocation for the AM header, but ignores the possibility of a stall in injection of the payload transfer at Commit time. @ Section: Graceful exits @ On June 24, 2003 ibv-conduit now passes all 9 (I added two recently) of the cases in testexit. By "Pass" I mean that the entire gasnet job (tested up to 8-way across my 4 dual-processor machines) terminates with no orphans, and with tracing properly finalized (if tracing is enabled). On August 11, 2003 the graceful exit code was revised to send O(N) network traffic in the worst case, as opposed to the O(N^2) required in all cases in the first implementation. Additionally, the exit code is properly propagated through the bootstrap, to yield a correct exit code for the parallel job as a whole. If using MPI for bootstrapping, the actual exit code will depend on supported in a given MPI implementation (some ignore the exit code of the individual processes). This code is heavily commented, but for the curious, here is a description of the code. There are three paths by which an exit request can begin. The first is through gasnetc_exit(), which may be called by the user, by the conduit in certain error cases, and by the default signal handler for "termination signals". The second is via a remote exit request, passed between nodes to ensure full-job termination from non-collective exits. The third is via an atexit/on_exit handler, registered by gasnetc_init(), used to catch returns from main() and user calls to exit(). There are slight variations among the code in these three cases, but most of the work is common, and is performed by three functions: gasnetc_exit_head(), gasnetc_exit_body() and gasnetc_exit_tail(). The first of these, _head, is used to determine the "first" exit and store its exit code for later use. This is important because even a collective exit will involve receiving remote exit requests. Only if a remote exit request is received before any local calls to gasnetc_exit(), should the request handler initiate the exit. Note that even in the case of a collective exit it is possible for the first remote request to arrive before the local gasnetc_exit() call. However, that is made very unlikely by the timing and is nearly harmless since the only difference is the raising of SIGQUIT in response to a remote exit request, which is not done for locally-initiated ones. The second common function, _body(), is used to perform the "meat" of the shutdown. It begins by ignoring SIGQUIT to avoid re-entrance, and then blocks all but the first caller in a polling loop to avoid multiple threads from executing the shutdown code. While the template uses a mutex to block additional threads, ibv-conduit is using atomics to avoid problems with signal context. Once additional thread are blocked from making progress through _body(), the AM progress thread (if any) is terminated to prevent it from interfering. Because strange things can happen if we are trying to shutdown from a signal context, a signal handler is installed for all the "abort signals". This signal handler just calls _exit() with the exit code stored by _head(). Because we may have problems shutting down if certain locks were held when a signal arrived, we also install the signal handler for SIGALRM, and use the alarm() function to bound the time spent blocked in the shutdown code. While there is the risk that this alarm might go off "too soon" if the shutdown has lots of work to do, we can be certain that the correct exit code is still generated. An additional step to address signal context is the definition of GASNETC_FATALSIGNAL_CALLBACK in ibv-conduit/gasnet_core_fwd.h, which gives ibv-conduit the chance to set gasnetc_exit_in_signal=1 just before the conduit-independent signal-handling code can reach the exit path. If this variable is non-zero then certain operations known to be especially risky in signal-handler context are skipped. After signal handlers are established, _body calls gasnetc_disable_AMs() to zero the table of client-registered AM handlers (though preserving the internal ones). This helps avoid interference. Then _body calls gasnetc_exit_reduce() to try to perform a collective reduce-to-all over the exit codes. If this completes within a given timeout then we know the exit is collective (and "graceful" is set non-zero) and skip over the leader/member logic described in the next two paragraphs. If the reduction does not complete within the timeout, then _body next calls gasnetc_get_exit_role() to "elect" a leader node for the exit. This is done with an alarm() timer in force. The use of an "election" with a timeout ensures that we will exit, even if node 0 is wedged. The election of a leader proceeds by sending a system-category AM request to node 0, and spinning to wait for a corresponding reply, which will indicated if the local node is the "leader" or a "member" in the coordination of the graceful exit. The logic on node 0 ensures that the first "candidate" is always made the leader, not waiting for multiple AMs to arrive. Additionally the member nodes may, under circumstances described below, know before entering gasnetc_get_exit_role() that they are members, and will not bother to send an AMRequest to node 0. In either case gasnetc_get_exit_role() indicates to _body which role the local node is to assume. From _body, the single leader node will enter gasnetc_exit_leader() and will begin sending an remote exit request (system-category AM, so this will all work between _init and _attach) to each peer. Then the leader waits (with timeout, of course) for a reply from each peer. This request conveys the desired exit code to each node. It also will wake them out of a spin-loop, barrier, or other case where they were not yet aware of the need to exit. In the handler for the exit request, a node will send a reply back to the leader, so it knows all the nodes are reachable. It will set its role to "member" and, if no exit is in-progress, it will start the exit procedure, as described later. From _body, the member nodes all call gasnetc_exit_member(), which simply spins until the remote exit request has arrived from the leader. Regardless of whether exit coordination (the reduction, or exit requests and replies) completed within their timeouts, _body proceeds to flush stdout and stderr one last time and closes stdin, stdout and stderr. Finally, _body shuts down its bootstrap support. If either coordination was completed within the timeout, then the gasnetc_bootstrapFini() routine is called indicating that we'll not be making any more calls to the bootstrap code and expect to exit shortly. However, if both coordinations did fail we call gasnetc_bootstrapAbort(exitcode). This call is meant to request that the bootstrap terminate our job "with prejudice" since we failed to coordinate a graceful shutdown on our own. We do this to try to avoid orphans, but risk lots of unsightly error messages and possible loss of our exit code. Assuming we did not call _bootstrapAbort (which does not return) we finish _body by canceling our alarm timer and return to our caller. The final common routine is gasnetc_exit_tail(). This function just does the last bit of work to terminate the job. It is not included in _body because we let the atexit/on_exit() case terminate "normally" after _body returns. However, in the case of exits initiated via gasnet_exit() or remote exit request we call _tail to complete the exit. In _tail we set an atomic variable to wake any threads which were stuck polling in _body due to being other than the first thread to enter. Those threads should eventually wake and also call _tail to terminate. Next, we call gasneti_killmyprocess() to do any platform- specific magic required to get the entire multithreaded application to exit. Finally we call _exit() with the saved exit code. Given the routines gasnetc_exit_{head,body,tail}() the code for the three types of exit are pretty trivial. In particular, gasnetc_exit() just calls _head, _body and _tail with no additional logic. In the request handler for the exit request AM, we look at the return from _head to determine if this exit request is the first we've seen (inclusive of local calls to gasnet_exit() and our atexit/on_exit handler). If it IS the first exit request, then we raise a SIGQUIT, as required by the GASNet spec, to allow the user's handler to perform its cleanup. However, to get the most robust exit code we don't want to run the _body code from a signal handler context if we can avoided it. Therefore we inspect the signal handler and skip the raise() call if the handler is the gasnet default handler, SIG_DFL or SIG_IGN. After the raise() returns, or is skipped all together, we are certain that the user's handler, if any, has executed and has NOT called gasnet_exit(). If a user handler had called gasnet_exit(), then raise() would not have returned. So, if we reach the code after the possible raise(), we proceed to call gasnetc_exit_body() and _tail to complete the (hopefully) graceful exit of the gasnet job. It is important to note that if we get a remote exit request that initiates an exit, then we will never return from the handler. However, the design of the AM code in IBV conduit ensures that this will actually work without deadlock. For one, we never run handlers from signal context or with locks held. Thus we can expect a "clean" set of locks. Furthermore, we don't expect to do anything useful with the network once the request handler calls _body anyway. The atexit handler just calls _head and _body before returning to allow the exit to complete. In this case we have a little problem with the lack of access to the return code. Therefore we just pass 0 to _head, which _body then sends in the remote exit requests. Experience has shown that, at least with LAM/MPI for bootstrap, when all but one task exits with zero, the single non-zero exit code becomes the exit code for the parallel job. Therefore, using zero here gives the specified exit code from the parallel job for both collective and non-collective returns from main. If support is detected at configure time for on_exit(), then it is used rather than atexit(), and the problem of the missing return code vanishes. In the normal case of a collective exit, the reduce-to-all-with-timeout is performed in 3 steps. The first is an intra-supernode reduction. The second is a reduce-to-all over supernodes using the same communication pattern as the dissemination barrier, requiring ceil(log_2(SN)) rounds in which each supernode sends and receives one AM (where "SN" is number of supernodes). The third step is a supernode-scoped broadcast. For non-PSHM builds, only a dissemination based reduce-to-all is performed (steps 1 and 3 are eliminated and "supernode" is replace by "node" in the description of step 2). For the non-collective exits, there is both a "best case" and a "worst case" to consider: Best case: one node is way ahead of the others and can win the leader election and send remote exit requests before the others attempt the election. In this case the coordinated shutdown needs 1 round-trip for the election, followed by (N-1) round-trips for the remote exit request/reply, for a total of 2*N AMs sent (not counting those from the failed reduction). Worst case: all nodes attempt the election at roughly the same time and a full N round-trips take place for the election, followed by (N-1) round trips for the remote exit request/reply, for a total of 4*N-2 AMs sent (plus those from the failed reduction). The average case for non-collective exits is somewhere between those two. @ Section: References @ [1] Bell, Bonachea. "A New DMA Registration Strategy for Pinning-Based High Performance Networks", Workshop on Communication Architecture for Clusters (CAC'03), 2003. Also at https://gasnet.lbl.gov/