======================================================================= ======================================================================= *** GASNet-EX Software Release *** GASNet-EX is the next generation of the GASNet-1 communication system. The GASNet interfaces are being redesigned to accommodate the emerging needs of Exascale supercomputing, providing communication services to a variety of PGAS programming models on current and future HPC architectures. This upgrade is being done over the next several years as part of the DOE Exascale Computing Program (ECP). GASNet-EX is a work-in-progress. Many features remain to be specified, implemented, and/or tuned. Users interested in learning what's new in GASNet-EX are recommended to peruse the ChangeLog file in this directory for a summary of recent developments. The docs/GASNet-EX.txt file provides more detailed information about the evolving EX specification. GASNet-EX notably includes a backwards-compatibility layer to assist in migration of current GASNet-1 client software. Existing GASNet clients can get started by relying on this layer (provided in gasnet.h), and incrementally add calls to the new gex_* interfaces (defined in gasnetex.h, which is automatically included by gasnet.h) to access new EX features and capabilities. For details, see docs/gasnet1_differences.md. Feedback or questions on any matters related to the GASNet-EX project are welcomed at: gasnet-devel@lbl.gov EX-Specific Implementation Notes: * gasnet_fwd.h provides a dependency-free useful subset of the API declarations, suitable for inclusion into application-level code. * The following public headers may be included in any order: - gasnetex.h : GASNet-EX API + Tools - gasnet.h : GASNet-1 API (also includes gasnetex.h) - gasnet_fwd.h : Forward declaration subset for app inclusion - gasnet_ratomic.h : Remote Atomics API - gasnet_vis.h : Non-contiguous RMA (VIS) API - gasnet_coll.h : Collectives API ======================================================================= ======================================================================= README file for GASNet https://gasnet.lbl.gov This is a user manual for GASNet. Anyone planning on using GASNet (either directly or indirectly) should consult this file for usage instructions. Other documentation: * In this README the "docs directory" means either docs/ in the source directory or ${prefix}/share/doc/gasnet/ in an installation of GASNet. * For GASNet licensing and usage terms, see license.txt. * For documentation on a particular GASNet conduit, see the README file in the conduit directory (also installed as README- in the docs directory). * For documentation on the communication-independent GASNet-tools library, see README-tools. * Additional information, including the GASNet specification and our bug tracking database, is available from https://gasnet.lbl.gov * Anyone planning to modify or add to the GASNet code base should also read the developer documents, available in the GASNet git repository, which can be browsed online: https://bitbucket.org/berkeleylab/gasnet/src/develop + README-devel: GASNet design information and coding standards + README-git: Rules developers are expected to follow when committing + template-conduit: A fill-in-the-blanks conduit code skeleton Contents of this file: * Introduction * Building and Installing GASNet * Basic Usage Information * Conduit Status * Single-node Development Options * Supported Platforms * Recognized Environment Variables * GASNet exit * GASNet tracing & statistical collection * GASNet Collectives * GASNet debug malloc services * GASNet inter-Process SHared Memory (PSHM) * MPI Interoperability * Contact Info and Support Introduction ============ GASNet is a language-independent, low-level networking layer that provides network-independent, high-performance communication primitives tailored for implementing parallel global address space SPMD languages and libraries such as UPC, UPC++, Co-Array Fortran, Legion, Chapel, and many others. The interface is primarily intended as a compilation target and for use by runtime library writers (as opposed to end users), and the primary goals are high performance, interface portability, and expressiveness. GASNet stands for "Global-Address Space Networking". The GASNet API is defined in the specification, which is included with this archive in docs/, and the definitive version is located on the GASNet webpage: https://gasnet.lbl.gov/ This README accompanies the GASNet source distribution, which includes implementations of the GASNet API for various popular HPC and general-purpose network hardwares. We use the term "conduit" to refer to any complete implementation of the GASNet API which targets a specific network device or lower-level networking layer. A conduit is comprised of any required headers, source files and supporting libraries necessary to provide the functionality of the GASNet API to GASNet clients. This distribution additionally includes a library of communication-independent portability tools called the "GASNet tools", which are used in the conduit implementations and also made available to clients (see README-tools for details). System Requirements =================== GASNet is extremely portable, and runs on most systems that are relevant to HPC production or development (and many that are not). The minimum system requirements are: * A POSIX-like environment, e.g. Linux or another version of Unix. For Mac systems, the free 'XCode command-line tools' from the Apple Store. For Windows systems one needs either of two options: + The free 'Cygwin' toolkit (https://www.cygwin.com/) + Windows 10 Subsystem for Linux, a.k.a. WSL (https://docs.microsoft.com/en-us/windows/wsl/) * GNU make (version 3.79 or newer). * Perl (version 5.005 or newer). * The following standard Unix tools: 'awk', 'sed', 'env', 'basename', 'dirname', and a Bourne-compatible shell (e.g. bash). * A C compiler with at least minimal C99 support. We explicitly support most OS's, architectures and compilers in widespread use today. See the 'Supported Platforms' section for details on systems we've recently validated. Most distributed-memory GASNet conduits have additional requirements, based on their interactions with network hardware and other implementation details. For example: * mpi-conduit requires an MPI-1.1 or newer compliant MPI implementation. * udp-conduit requires POSIX socket libraries and a C++98 or newer compiler. See each conduit README for additional details on system requirements. Building and Installing GASNet ============================== Here are the steps to build GASNet: * Step 0 (optional): ./Bootstrap Runs the autoconf tools to build a configure script (this can be done on any system and may already have been done for you). If you are keeping a copy of the GASNet sources in your own source control repository (CVS, svn, Hg, Git, etc.), then please also see "Source Control and GASNet" in README-devel (in the GASNet git repo). * Step 1: ./configure (options) Generate the Makefiles tailored to your system, creating a build tree in the current working directory. You can run configure from a different directory to place your build files somewhere other than inside the source tree (a nice option when maintaining several build trees on the same source tree). Any compiler flags required for correct operation on your system (e.g. to select the correct ABI) should be included in the values of CC, CXX and MPI_CC. For example to build 32-bit code when your gcc and g++ default to 64-bit: configure CC='gcc -m32' CXX='g++ -m32' MPI_CC='mpicc -m32' or similarly, if you want libgasnet to contain debugging symbols: configure CC='gcc -g' CXX='g++ -g' (however also see --enable-debug, below) Some of the useful configure options: --help - display all available configure options --prefix=/install/path - set the directory where GASNet will be installed --enable-debug - build GASNet in a debugging mode. This turns on C-level debugger options and also enables extensive error and sanity checking system-wide, which is highly recommended for developing and debugging GASNet clients (but should NEVER be used for performance testing). --enable-debug also implies --enable-{trace,stats,debug-malloc}, but these can still be selectively --disable'd. --enable-trace - turn on GASNet tracing (see usage info below) --enable-stats - turn on GASNet statistical collection (see usage info below) --enable-debug-malloc - use GASNet debugging malloc (see usage info below) --enable-segment-{fast,large,everything} - select a GASNet segment configuration (see the GASNet spec for more info) --enable-pshm - Build GASNet with inter-Process SHared Memory (PSHM) support. This feature uses shared memory communication among the processes (aka GASNet nodes) within a single compute node (where the other alternatives are multi-threading via a PAR or PARSYNC build; or use of the conduit's API to perform the communication). Note that not all conduits and operating systems support this feature. For more information, see the section below entitled "GASNet inter-Process SHared Memory (PSHM)". --with-max-segsize= - configure-time default value for GASNET_MAX_SEGSIZE, which is used when GASNET_MAX_SEGSIZE is not set at runtime. See the description of the GASNET_MAX_SEGSIZE environment variable below for details. Configure will detect various interesting features about your system and compilers, including which GASNet conduits are supported. For cross-compilation support, look for an appropriate cross configure script in other/contrib/ and link it into your top-level source directory and invoke it in place of configure. Example for the Cray XC with slurm: cd ln -s other/contrib/cross-configure-cray-aries-slurm . cd

/cross-configure-cray-aries-slurm (configure_options) For cross-compilation support on platforms without scripts in other/contrib see the instructions in other/cross-configure-help.c. [NOTE: We currently don't distribute the cross-configure-help.c in our normal distribution. If you think you need it, contact us at gasnet-devel@lbl.gov] On HPE Cray EX (aka "Shasta") systems, we recommend the following configure arguments to use the vendor's compiler wrappers: --with-cc=cc --with-cxx=CC --with-mpi-cc=cc Additionally, exactly one of the following is recommended to ensure that ofi-conduit is built for the appropriate libfabric provider: * HPE Cray EX with Slingshot-10 (100Gbps) NICs: --with-ofi-provider=verbs * HPE Cray EX with Slingshot-11 (200Gbps) NICs: --with-ofi-provider=cxi * HPE Cray EX with BOTH NIC types: --with-ofi-provider=generic On Linux clusters with Intel Omni-Path networks, we recommend the following configure arguments to avoid using ibv-conduit over an emulated libibverbs: --disable-ibv --enable-ofi --with-ofi-provider=psm2 * Step 2: make all Build the GASNet libraries. A number of other useful makefile targets are available from the top-level: make {seq,par,parsync} build the conduit libraries in a given mode make tests-{seq,par,parsync} build all the GASNet tests in a given mode make run-tests-{seq,par,parsync} build and run all the GASNet tests in a given mode make (run-)tests-installed-{seq,par,parsync} use the installed library to build (and run) all the GASNet tests make run-tests run whatever tests are already built in the conduit directories make run-tests TESTS="test1 test2..." run specifically-listed tests that are already built in the conduit directories make DO_WHAT="" build selected makefile target in all supported conduit directories Each conduit directory also has several useful makefile targets: make {seq,par,parsync} build the conduit libraries in a given mode make tests-{seq,par,parsync} build the conduit tests in a given mode make testXXX build just testXXX, in SEQ mode make testXXX-{seq,par,parsync} build just testXXX, in a given mode make run-tests-{seq,par,parsync} build and run the conduit tests in a given mode make (run-)tests-installed-{seq,par,parsync} use the installed library to build (and run) all the GASNet tests make run-tests run whatever tests are already built in the conduit directory make run-tests TESTS="test1 test2..." run specifically-listed tests that are already built in the conduit directory make run-testexit build a script to run the testexit tester and run it Compilation and linker flags for the GASNet libraries and tests can be augmented from the command-line by setting the following variables in the make command, as described in more detail in the next section: make MANUAL_CFLAGS=... Flags to add on the C compile for GASNet libs, clients & tests make MANUAL_CXXFLAGS=... Flags to add on the C++ compile for GASNet libs, clients & tests make MANUAL_MPICFLAGS=... Flags to add on the mpicc compile for GASNet libs, clients & tests make MANUAL_DEFINES=... Flags to add on all compiles for GASNet libs, clients & tests make MANUAL_LDFLAGS=... Linker flags to add for GASNet clients & tests make MANUAL_LIBS=... Linker library flags to add for GASNet clients & tests Note this feature should be used sparingly, as some flags can invalidate the results of tests performed at configure time. The preferred way to add arbitrary flags is in the $CC, $CXX and $MPI_CC variables passed to configure. The following misc make variables can also be set to affect GASNet compilation: make SEPARATE_CC=1 Build libgasnet using separate C compiler invocations, rather than one big one. make KEEPTMPS=1 Keep temporary files generated by the C compiler, if supported. * Step 3 (optional): make install Install GASNet to the directory chosen at configure time. This will create an include directory with a sub-directory for each supported conduit, and a lib directory containing a library file for each supported conduit, as well as any supporting libraries. GASNet may also be used directly from the build directory, as a convenience to eliminate steps if you are making changes to GASNet or its configuration. Manual control over compile and link flags ========================================== As described in the previous section, the recommended mechanism for passing flags to the compilers used by GASNet is to include them in the definition of the compiler variable itself (e.g. CC='gcc -m64'). These will be followed on the command line by flags chosen by GASNet's configure script instead of using the standard CFLAGS and CXXFLAGS make variables. If there is a need to pass flags that override ones chosen by configure, then one may set MANUAL_CFLAGS, MANUAL_CXXFLAGS and MANUAL_MPICFLAGS on the 'make' command line, and these are guaranteed to appear on the compilation command line after the ones chosen by configure. Additionally, MANUAL_DEFINES is provided to pass flags to all three compilers, but its order on the command line is not guaranteed (since it is intended for position-independent command line options such as '-Dfoo=10'). Note that GASNet does not assume that MPI_CC is the same compiler as CC (though that is recommended), and therefore invokes it using a distinct MANUAL_MPICFLAGS, instead of MANUAL_CFLAGS. The MPI C compiler is used to compile a portion of mpi-conduit (AMMPI) and the MPI-based spawner code used by several conduits; and to link executables for mpi-conduit and any conduit in which the MPI-based spawner is enabled. Similarly, the C++ compiler is used to compile a portion of udp-conduit (AMUDP) and to link udp-conduit executables. Finally, the make variables MANUAL_LDFLAGS and MANUAL_LIBS are provided to add to the link command lines, after the respective configure-detected settings. They are intended for flags such as '-Ldir' and '-lfoo', respectively. Since they do not have compiler-specific variants, their use is limited to compiler-independent flags unless special care is taken with the selection of the make target to ensure that make only invokes one compiler as a linker. All of the MANUAL_* make variables described above are honored by the Makefile infrastructure used to build GASNet's libraries and tests. Additionally, the makefile fragments (next section) use these variables when compiling client code. Basic Usage Information ======================= See the README for each GASNet conduit implementation for specific usage information, but generally client programs should #include (and nothing else), and use the conduit-provided compilation settings. The best way to get the correct compiler flags for your GASNet client is to "include" the appropriate makefile fragment for the conduit and configuration you want in your Makefile, and use the variables it defines in the Makefile rules for your GASNet client code. For example: ---------------------------- include $(gasnet_prefix)/include/mpi-conduit/mpi-seq.mak .c.o: $(GASNET_CC) $(GASNET_CPPFLAGS) $(GASNET_CFLAGS) -c -o $@ $< .cc.o: $(GASNET_CXX) $(GASNET_CXXCPPFLAGS) $(GASNET_CXXFLAGS) -c -o $@ $< myprog: myprog.o $(GASNET_LD) $(GASNET_LDFLAGS) -o $@ $< $(GASNET_LIBS) ---------------------------- See tests/Makefile for another example of compiling GASNet client code. For more fine-grained control, the flags variables break-down as follows: GASNET_CFLAGS is an alias for: $(GASNET_OPT_CFLAGS) $(GASNET_MISC_CFLAGS) GASNET_CPPFLAGS is an alias for: $(GASNET_MISC_CPPFLAGS) $(GASNET_DEFINES) $(GASNET_INCLUDES) GASNET_CXXFLAGS is an alias for: $(GASNET_OPT_CXXFLAGS) $(GASNET_MISC_CXXFLAGS) GASNET_CXXCPPFLAGS is an alias for: $(GASNET_MISC_CXXCPPFLAGS) $(GASNET_DEFINES) $(GASNET_INCLUDES) The content of these variables follow these general guidelines: * GASNET_[CC,CXX] contain the configure-detected C and C++ compilers used to build GASNet and guaranteed to work with any compiler-specific flags embedded in the corresponding variables. Clients are strongly advised to compile all modules using these same compilers to ensure object compatibility. * GASNET_INCLUDES contains only and all -I preprocessor flags * GASNET_DEFINES contains only and all -D or -U preprocessor flags * GASNET_OPT_* contain compiler-specific flags controlling the debug/opt level * GASNET_MISC_* contain other needed compiler-specific compile-time flags * GASNET_LD contains one of the configure-detected compilers (CC, CXX or MPI_CC) which should be used to link the GASNet client executable, in order to guarantee satisfaction of conduit library dependencies (eg on C++ or MPI). * GASNET_LIBS contains only and all -L or -l link-time flags * GASNET_LDFLAGS contains other needed link-time flags, which may be GASNET_LD-specific This summary is provided for informational purposes only - altering or omitting any of the flags contained in these variables could result in a non-functional build environment. Using GASNet with pkg-config ---------------------------- As a convenience, the same variables described in the previous section are also available via the UNIX pkg-config utility. For example, if pkg-config is installed on your system and the GASNet .pc files are in your PKG_CONFIG_PATH, then you can retrieve the value for GASNET_CC for udp-conduit in GASNET_SEQ mode with a command like: pkg-config gasnet-udp-seq --variable=GASNET_CC Note the pkg-config --cflags and --libs arguments retrieve the appropriate subset of GASNet variables, but pkg-config does not offer aliases corresponding to GASNET_CC, GASNET_LD or the GASNET_*CXX* variables. Use the `pkg-config --variable=` syntax to retrieve these. Here is a complete example of using pkg-config with GASNet in a Makefile: PKG_CONFIG_PATH = $(gasnet_prefix)/lib/pkgconfig pkg = gasnet-udp-seq .c.o: `pkg-config $(pkg) --variable=GASNET_CC` `pkg-config $(pkg) --cflags` -c -o $@ $< .cc.o: `pkg-config $(pkg) --variable=GASNET_CXX` `pkg-config $(pkg) --variable=GASNET_CXXCPPFLAGS` \ `pkg-config $(pkg) --variable=GASNET_CXXFLAGS` -c -o $@ $< myprog: myprog.o `pkg-config $(pkg) --variable=GASNET_LD` -o $@ $< `pkg-config $(pkg) --libs` Conduit Status ============== The GASNet distribution includes multiple complete implementations of the GASNet API targeting particular lower-level networking layers. Each of these implementations is called a 'conduit'. In some cases the lower-level layer is a proprietary or hardware-specific network API, whereas in other cases the target API is a portable standard (although in some cases this distinction is blurred). The corresponding GASNet conduits can be loosely categorized as either 'native' or 'portable' conduits. Below is the list of conduits in the current distribution, and their high-level status. Many conduits are supported on multiple platforms (CPU architectures, operating systems and compilers) - see the 'Supported Platforms' section for more details on platforms. For more detailed info about each conduit, please consult the corresponding conduit README. Portable conduits: ----------------- smp-conduit: SMP loopback (shared memory) The conduit of choice for GASNet operation within a single shared-memory node. Rigorously tested and supported on all current platforms. udp-conduit: UDP/IP (part of the TCP/IP protocol suite) The conduit of choice for GASNet over Ethernet, supported on any TCP/IP-compliant network. Rigorously tested over Ethernet and supported on most current platforms (see below). mpi-conduit: MPI (Message Passing Interface) A portable implementation of GASNet over MPI-1.1 or later. Intended as a reference implementation for systems lacking native conduit support. Rigorously tested and supported on most current platforms (see below). ucx-conduit: Unified Communication X framework [EXPERIMENTAL] GASNet over the Unified Communication X framework (UCX). This conduit is currently experimental, and is not yet carefully tuned for performance. It has only been validated on Mellanox InfiniBand devices starting from ConnectX-5. ofi-conduit: Open Fabrics Interfaces (most providers) GASNet over the Open Fabrics Interface framework (libfabric). This conduit is functionally complete but not yet carefully tuned for performance. With the exception of HPE Slingshot and Intel Omni-Path networks (where ofi-conduit is either the best or only option and considered "native"), users are advised to use other conduits compatible with their hardware. Native, high-performance conduits: --------------------------------- aries-conduit: Cray XC Aries The conduit of choice for GASNet on the Aries networks in Cray XC systems. Rigorously tested and supported. ibv-conduit: InfiniBand Verbs GASNet over the OpenFabrics Verbs API. Rigorously tested and supported over InfiniBand hardware on all supported systems (see below). Believed to also work on other hardware offering a standard-compliant Verbs layer. ofi-conduit: Open Fabrics Interfaces (select providers/networks) GASNet over the Open Fabrics Interface framework (libfabric). This conduit is functionally complete but not yet carefully tuned for performance. On HPE Slingshot and Intel OmniPath networks, ofi-conduit is either the best or only option available. On only those systems, ofi-conduit is categorized as a native, high-performance conduit. Single-node Development Options =============================== GASNet supports hardware configurations ranging from HPC supercomputers and clusters to individual workstations and laptops. Often the easiest way to develop and debug code is in a single-node environment, so this section provides conduit recommendations for development. It should also be noted that configure option --enable-debug is highly recommended for finding bugs when developing GASNet client code. Option 1. smp-conduit ---------------------- This is a pure shared-memory implementation and should be the fastest and easiest to use option. GASNet's shared memory support should work "out of the box" (no special configuration options required) on common laptop, desktop, and workstation environments, including Linux, macOS, Windows with Cygwin, Windows 10 Subsystem for Linux (WSL), and Solaris. Option 2. udp-conduit ---------------------- By default this will be a shared-memory implementation within your single node, indistinguishable from smp-conduit except for small issues like job launch and stdin/out/err handling. However, one can disable shared memory (by passing --disable-pshm at configure time or by setting the environment variable GASNET_SUPERNODE_MAXSIZE=1 at runtime). This may be more realistic for testing of the expected behaviors in multi-node systems. Option 3. mpi-conduit ---------------------- If you have a MPI installed (and in your $PATH when GASNet is configured) then everything said for udp-conduit holds true for mpi-conduit, including the ability to disable shared-memory support. Additionally, since MPI will very likely use shared-memory internally, you can get multi-node like isolation at the GASNet level with the performance of shared-memory communication (much better than udp). Supported Platforms =================== Platforms where GASNet and Berkeley UPC have been successfully tested include: OS/Architecture/compiler/ABI: network conduits ---------------------------------------------- * Linux/x86-Ethernet/{gcc,clang}32: smp, mpi, udp * Linux/x86-Mellanox/gcc/32: smp, mpi, udp, ibv * Linux/x86/IntelC/32: smp, mpi, udp * Linux/x86/PortlandGroupC/32: smp, mpi, udp * Linux/x86_64-Ethernet/{gcc,clang,PGI}/{32,64}: smp, mpi, udp, ofi * Linux/x86_64-Mellanox/{gcc,clang,PathScale,NVHPC,IntelC,Intel oneAPI}/64: smp, mpi, udp, ibv * Linux/x86_64-OmniPath/gcc/64: smp, mpi, udp, ofi * Linux/x86_64/x86-Open64/64: smp, udp, # * Linux/PowerPC-Ethernet/{gcc,clang}/{32,64}: smp, mpi, udp * Linux/PowerPC-HFI/gcc/32: smp, mpi, udp & * Linux/PPC64le/{gcc,clang,pgi,NVHPC,xlc}/64: smp, mpi, udp, ibv # * Linux/MIPS/gcc/{32,n32,64}: smp, udp, # * Linux/MIPS64el/gcc/{32,n32,64}: smp, udp, # * FreeBSD/{x86,amd64}/{gcc,clang}/{32,64}: smp, mpi, udp * OpenBSD/{x86,amd64}/{gcc,clang}/{32,64}: smp, mpi, udp * NetBSD/{x86,amd64}/{gcc,clang}/{32,64}: smp, mpi, udp * Solaris10/SPARC/gcc/{32,64}: smp, udp, mpi * Solaris10/x86/gcc/{32,64}: smp, udp, mpi * OpenSolaris/x86/gcc/{32,64}: smp, udp, # * Solaris11Express/x86/gcc/{32,64}: smp, udp, mpi, ibv * Solaris11/x86/gcc/{32,64}: smp, udp, mpi, ibv * MSWindows-Cygwin/{x86,x86_64}/{gcc,clang}/{32,64}: smp, udp, mpi (OpenMPI) * MSWindows10-Linux(WSL)/{gcc,clang}/64: smp, udp, mpi * macOS/{x86,x86_64}/icc/{32,64}: smp, udp, mpi & * macOS/{x86,x86_64}/{gcc,clang}/{32,64}: smp, udp, mpi * macOS/{x86,x86_64}/PGI/{32,64}: smp, udp, # * macOS/AARCH64/{gcc,clang}/64: smp, udp, # * CNL/Cray-XT/{gcc,PGI,PathScale,Intel}/64: smp, mpi & * CNL/Cray-XE/{gcc,PGI,PathScale,Intel,Cray}/64: smp, mpi & * CNL/Cray-XK/{gcc,PGI,PathScale,Intel,Cray}/64: smp, mpi & * CNL/Cray-XC/{gcc,Intel,Cray}/64: smp, mpi, aries, ofi * CNL/HPE-Cray-EX/{gcc,Cray}/64: smp, mpi, ofi * Linux/Cray-XD1/{gcc,PGI}/64: smp, mpi & * ucLinux/Microblaze/gcc/32: smp, udp, #%& * Linux/ARM/{gcc,clang}/32: smp, udp, # * Linux/AARCH64/{gcc,clang}/64: smp, udp, mpi # = We have not tested MPI but have no reasons to doubt that mpi-conduit would work. % = System lacks pthreads or they are broken & = System has not been tested in recent releases due to lack of access. Reports of success or failure on this system are strongly encouraged. BETA = Support for this system is in beta state, please report your experiences. This list is not meant to be exhaustive. Other combinations of the platforms above are likely to also work, these are just the systems we've personally tested. Several of the systems listed using a vendor-specific C compiler can also use gcc as the underlying C compiler, although we generally recommend the vendor C compiler for performance reasons. The following compilers are believed to work on platforms listed above, with the provided minimum version: Gnu (gcc 3.0+), LLVM (clang 3.6+), Apple (Xcode 7.1+), PGI (pgcc 11.0 to 20.4), NVHPC (20.9+), Intel (icc 16+), Intel oneAPI (icx 21+), IBM XL (xlc 13+), Cray (CCE 8.6+) Recognized Environment Variables ================================ Users of language- or application-specific wrappers for job launch should also consult the wrapper's documentation. Such wrappers often have options to set these environment variables while also enabling any corresponding language- or application-specific support. In the following descriptions, a "Boolean" setting is one which accepts "1", "y" and "yes" as TRUE and "0", "n" and "no" as FALSE (all case-insensitive). While the descriptions use "0" or "1" to be concrete, one may substitute any of these aliases. * GASNET_VERBOSEENV: Boolean setting to output information about environment variable settings read by the conduit that affect conduit behavior. * GASNET_FREEZE: Boolean setting to make GASNet pause and wait for a debugger to attach on startup * GASNET_FREEZE_ON_ERROR: Boolean setting to make GASNet pause and wait for a debugger to attach on any fatal errors or fatal signals * GASNET_FREEZE_SIGNAL: set to a signal name (e.g. "SIGINT" or "SIGUSR1") to specify a signal that will cause the process to freeze and await debugger attach. * GASNET_TRACEFILE, GASNET_TRACEMASK, GASNET_STATSFILE, GASNET_STATSMASK, GASNET_TRACEFLUSH, GASNET_TRACELOCAL, GASNET_TRACENODES, GASNET_STATSNODES: control tracing & statistical features, if enabled at configure time. See usage information below. * GASNET_TEST_POLITE_SYNC: Boolean setting to enable polite-mode synchronization for the GASNet tests (only), for running with overcommitted CPUs. * GASNET_MALLOC_* : control the GASNet debug malloc features, if enabled at configure time. See usage information below. * GASNET_MAX_SEGSIZE - control the upper limit for FAST/LARGE segment size on most conduits This setting defaults to the value passed to configure: --with-max-segsize= In FAST and LARGE segment configurations, GASNet probes each compute node at startup to determine an upper-limit on the available space for use in the GASNet segment (and some other large internal objects). This value provides one upper-limit to that probe, which also has the effect of limiting the space available for client segments (as reported by gasnet_getMaxLocalSegmentSize()). has the following format: size_spec ( / opt_suffix ) where 'size_spec' is either an absolute memory size: [0-9]+{KB,MB,GB} or a fraction of compute node physical memory: 0.85 and 'opt_suffix' is one of the following: (or empty, which means "P") "P" : means the limit is per-process and EXCLUDES internal GASNet objects "H" : means the limit is host-wide and INCLUDES internal GASNet objects Examples: "0.85/H" : limit host-wide use at 85% of physical memory (this is also the default) "4GB/P" : try to ensure 4GB per process of GASNet shared segment space The default behavior of this option has grown considerably smarter over time, so it's anticipated that most clients will never need to set this. * GASNET_DISABLE_MUNMAP - Boolean setting to request the use of mallopt() on glibc systems to disable mmap-based allocation for satisfying malloc. This can be used to work-around a known bug in firehose (bug 495) that could lead to incorrect behavior after free()ing out-of-segment memory areas previously used for communication. Note that on some systems (32-bit Linux in particular) the disable is only partly effective because once the sbrk()-controlled heap reaches the bottom of shared libraries, glibc will use mmap() used to obtain memory regardless of any options one can control. The default is conduit-dependent. * GASNET_MAX_THREADS - per-node limit on the number of GASNet client pthreads (in PAR and PARSYNC modes) that can simultaneously be live on each GASNet node. This is subject to the hard limit established by configure --with-max-pthreads-per-node, and pthread limits that may be imposed by specific conduits (see conduit README). * GASNET_SUPERNODE_MAXSIZE - limit on size of a GASNet "supernode". This is the maximum number of processes (GASNet "nodes") that will be grouped into a shared-memory "supernode", as reported by gasnet_getNodeInfo() and used by shared-memory communication (PSHM). A value of zero means no limit. * GASNET_PSHM_BARRIER_HIER - Boolean setting to enable/disable hierarchical shared-memory barrier. When shared-memory communication (PSHM) is enabled, the default behavior of most of GASNet's barrier implementations is to use a two-stage barrier which coordinates within each supernode before communicating across the network. This variable can be set to "0" to disable this optimization. * GASNET_CATCH_EXIT - Boolean setting, where a "0" prevents GASNet from forcing global job termination (via atexit() or on_exit()) when a process calls exit() or returns from main(). GASNet's default behavior helps prevent orphaned processes that can occur in some systems after an incomplete job termination, but may interfere with some profiling tools that write output inside atexit handlers. Setting this variable may allow those tools to operate, but the client code (or other entity) must assume responsibility for ensuring no orphan processes are left behind. Note this variable does not affect the behavior of explicit calls to gasnet_exit() (either directly or indirectly via calls like upc_global_exit(), returning from UPC main(), or reaching the GASNet default fatal signal handler) which will still bypass atexit handlers and kill the job. The recommended method to ensure the execution of atexit handlers is to run with GASNET_CATCH_EXIT=0 and collectively invoke libc exit(). * GASNET_NO_CATCH_SIGNAL - specify a comma separated list of signals to exclude from GASNet default signal handling. Formats "SIGSEGV", "SEGV", "sigsegv", "segv", and "11" are accepted. If the value of this environment variable is "*" (alone, with no leading or trailing whitespace), GASNet will not register the default handler for any signal. The default handler provides GASNet's backtrace support and ensures clean exits when fatal signals are received. Disabling this handler may allow use of other tools for debugging of signals, but is not intended for production use. * GASNET_BACKTRACE - Boolean setting to request the generation of stack backtraces on most fatal errors. The format/content of these backtraces varies by platform. On some platforms no backtrace support is available and this variable will be ignored. Backtraces are sent to stderr and to the trace file if tracing is active (see below). WARNING: Some fatal errors may involve memory corruption or other abnormal conditions that could cause the backtrace code to hang. For this reason we do not recommend setting GASNET_BACKTRACE by default (though there is no performance penalty for doing so). When reporting bugs, one is strongly encouraged to include a backtrace if possible. The backtrace is almost always more detailed if GASNet is built with debugging enabled, but may still be useful to a GASNet developer in a non-debug build. If tracing is active (see below) then a copy of the backtrace will be sent to the trace file. This file may provide developers with potentially useful information about activities prior to the error. * GASNET_BACKTRACE_NODES - if enabled by GASNET_BACKTRACE then this provides an optional list of nodes on which to permit backtraces. The list 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 may generate backtraces. * GASNET_BACKTRACE_SIGNAL: set to a signal name (e.g. "SIGINT" or "SIGUSR1") to specify a signal that will cause the process to generate an immediate backtrace, and then continue executing. This is useful for getting a convenient backtrace for a "hung" process. * GASNET_BACKTRACE_TYPE - set to a comma-delimited, ordered list of mechanisms (i.e. different debugger tools) to try when generating a backtrace for GASNET_BACKTRACE. The default value (visible via GASNET_VERBOSEENV) includes all mechanisms detected as supported on the current platform. * GASNET_BACKTRACE_MT - Boolean setting requesting multi-threaded backtraces when supported by the backtrace mechanism. This overrides the default, which is generally "1" for thread-safe builds and SEQ builds with conduit-internal threads. Otherwise, the default is "0". * GASNET_DISABLE_ENVDECODE/GASNET_DISABLE_ARGDECODE - Boolean setting to disable the automatic decoding of environment variable values/command-line arguments. Some GASNet spawners automatically encode shell meta-characters passing through non-GASNet spawn scripts, in to order to ensure their safe delivery to the GASNet client program. * GASNET_SPAWN_VERBOSE - Boolean setting to enable console debugging output of operations related to job creation and teardown. Details vary by configuration. * GASNET_BARRIER - select the communication algorithm for use in GASNet barriers. The following values are available on all conduits: AMDISSEM - uses Active Messages to implement the Dissemination barrier algorithm as described in section 3.3 of John M. Mellor-Crummey and Michael L. Scott. "Algorithms for scalable synchronization on shared-memory multiprocessors." ACM ToCS, 9(1):21 65, 1991. RDMADISSEM - uses Put operations to implement the Dissemination algorithm. DISSEM - auto-selects either AMDISSEM or RDMADISSEM The AMDISSEM or RDMADISSEM algorithm is selected automatically based on conduit-specific criteria. In general RDMADISSEM is favored when the GASNet Extended API is implemented natively. AMCENTRAL - uses Active Messages to manipulate a single centralized counter. This is an inherently non-scalable barrier which does not honor the setting of GASNET_PSHM_BARRIER_HIER to enable shared-memory optimizations. Therefore, this choice is available only for debugging purposes. In addition to those choices, many conduits have additional network-specific barrier algorithms documented in the corresponding conduit READMEs. The default is DISSEM, unless the conduit README documents another default. * GASNET_PSHM_BARRIER_RADIX - set radix for intra-node barrier algorithm For configurations using PSHM (a function of OS and conduit) the GASNet barrier is performed in intra-node and inter-node stages. This environment variable is the radix of the tree-based intra-node (shared-memory) barrier. If zero (default) then radix = size - 1, resulting in a "flat tree" (linear time) If positive, then the given value is the out-degree of an N-ary tree. If negative, then a tree is built with the processes in groups of size = -radix. The first process in each group is the parent of the others in that group. The rank==0 process is the parent of the other group-representatives (in addition to being the parent of the others in its own group). The default is 0 (linear) on most platforms. * GASNET_PSHM_NETWORK_DEPTH - set depth of the intra-node AM network For configurations using PSHM (the default on most systems) GASNet implements intra-node Active Messages using a shared-memory queue. This variable sets the "network depth" of this implementation in units of maximum-sized messages. A process can send at least this many out-bound intra-node AMs concurrently (to any peer) before the implementation might stall awaiting peer attentiveness to retire AMs. When sending smaller messages, the effective depth is greater due to allocating less space in the queue than would be needed for a maximum-sized message. The default is 32 and the minimum is 4. * GASNET_NODEMAP_EXACT - Boolean setting to enable an exact algorithm for discovery of shared memory nodes. Several GASNet conduits use mmap() and/or conduit-specific memory registration resources to establish the GASNet segment. When multiple GASNet nodes (processes) run on the same O/S node, there is a potential for competition for resources which can be managed by coordinating among the processes. When PSHM support is enabled, multiple processes on the same O/S node must be identified so they can cross-mmap() their GASNet segments. Processes sharing the same O/S node can be discovered using an algorithm that runs in time linear in the number of GASNet nodes, and which is sufficient for all common process layout patterns. However, this approach may fail to discover sharing in unusual cases. Setting this variable to "1" enables an algorithm that is certain to find all sharing of memory, but has an expected running time proportional to N*log(N). The default is currently "1": use of the slower, but safer, exact algorithm. * GASNET_VIS_AMPIPE - Boolean setting to enable packing of most non-contiguous put/gets into AMMediums, with each packet of size approx MaxMedium (the only exception being cases where both sides happen to be fully contiguous, in which case we skip packing). The default is conduit-dependent. * GASNET_VIS_{PUT,GET}_MAXCHUNK - limits the max size of a contiguous chunk which will be packed by AM pipelining in a strided put/get or indexed put/get, respectively. The chunk size may additionally be limited based on the size that will fit in one MaxMedium. The default value is conduit-specific. * GASNET_VIS_MAXCHUNK - Provides a default value for GASNET_VIS_{PUT,GET}_MAXCHUNK, to be used when the more specific knob is unset. * GASNET_VIS_REMOTECONTIG - Boolean setting to enable a pack & RDMA algorithm for gather puts and scatter gets - i.e. cases that are locally non-contiguous but remotely contiguous. The default is conduit-dependent. * GASNET_COLL_SCRATCH_SIZE - Specifies the size of the scratch space allocated on each rank for internal use in collective communications. This is the preferred size allocated for the initial team, and is the value returned from a query using GEX_FLAG_TM_SCRATCH_SIZE_RECOMMENDED. If the size is set too low, then the performance of collectives may suffer. This parameter must be single-valued (same value on all processes). A value of zero is permitted, but any value below some implementation-specific minimum value will be silently increased to that minimum. Defaults to 2MB per rank. * GASNET_COLL_ENABLE_SEARCH - Boolean setting to enable autotuning of collectives * GASNET_COLL_TUNING_FILE - file to read and/or write collective autotuning data For usage information, see the file autotuner.txt in the docs directory. * GASNET_FS_SYNC - Boolean setting to enable a sync() call (or equivalent) at exit time. Default is "0". Try this setting if you experience truncated output. * GASNET_TMPDIR: if set to a valid directory name this is used instead of TMPDIR or "/tmp" as a location for creating temporary files. * GASNET_SD_INIT, GASNET_SD_INITVAL and GASNET_SD_INITLEN These variables are *only* recognized in a build configured with --enable-debug. When using the negotiated-payload AM APIs, GASNet can optionally initialize the start of buffers it allocates at Prepare time with a defined pattern, and verify at Commit time that the client has over-written this "canary". To reduce overhead, only a fixed-length pattern is used. However, to avoid false alarms, this length must be sufficiently long to make the probability of matching the client data very low. GASNET_SD_INIT - Boolean setting which defaults to "1" in debug builds (and is effectively "0" otherwise) Setting to a "0" disables the initialization and checking GASNET_SD_INITVAL - Defaults to "NAN" Specifies the initialization pattern. See GASNET_MALLOC_INITVAL for more info. GASNET_SD_INITLEN - Defaults to 128 Specifies the length of the "canary" used for this checking. If 'nbytes' passed to the Commit call is less than this length, then no checking will be performed. If set too small then the likelihood of random payloads corresponding to the initialization pattern may lead to false positives. The minimum is 1 byte and smaller values will be silently replaced by 1. * OPTIONAL: GASNET_EXITTIMEOUT, GASNET_EXITTIMEOUT_MAX, GASNET_EXITTIMEOUT_MIN, and GASNET_EXITTIMEOUT_FACTOR - control exit-coordination timeout. Some conduits use a timeout to distinguish orderly job exit from uncoordinated failure of one or more nodes. This is important to help avoid leaving "orphan" processes on the nodes. These environment variables allow the user to adjust the length of time that the conduit will wait to establish contact among all nodes before deciding that an exit is uncoordinated. By default, the timeout is computed as GASNET_EXITTIMEOUT = min(GASNET_EXITTIMEOUT_MAX, GASNET_EXITTIMEOUT_MIN + nodes * GASNET_EXITTIMEOUT_FACTOR) Setting GASNET_EXITTIMEOUT provides a specific value, ignoring this formula. Setting one or more of the others will compute a value for GASNET_EXITTIMEOUT using the formula above (and defaults for any variables not set in the environment). Currently most conduits honor these settings. Smp-conduit honors these settings only when PSHM-support is enabled. Default values differ among these conduits. * OPTIONAL: GASNET_THREAD_STACK_MIN, GASNET_THREAD_STACK_PAD Some conduits spawn additional threads internally for various purposes, which may include running client's AM handlers. If such AM handlers have unusually large stack space requirements, a mechanism is required to ensure these internal threads will have large enough stacks. These settings allow one to control the stack sizes of these internal threads. The stack size requested from the system will be computed as stack_size = MAX(GASNET_THREAD_STACK_MIN, GASNET_THREAD_STACK_PAD + default) where "default" denotes the system's default thread stack size, and values of both variables have been rounded-up to a multiple of the page size. The actual stack size may be smaller (e.g. to satisfy rlimits) or larger (for instance, the next power of two in some implementations). The default for both settings is zero, which results in using the system default size for thread stacks. * GASNET_HOST_DETECT To implement gex_System_QueryHostInfo() and to construct shared-memory "nbrhds", GASNet must map the hosts (compute nodes) in a job. This requires a unique identifier for each host. This string-valued setting selects the identifier used. The following are implemented for most conduits: "gethostid" - the 32-bit value returned by POSIX gethostid() "hostname" - a 64-bit hash of the hostname (as reported by POSIX gethostname()) Some conduits support: "conduit" - a network-specific identifier (such as a MAC address) On networks providing the "conduit" option, conduit-specific documentation will describe whether it is supported in addition to the two listed above, or if it is the *only* supported option. The default is determined as follows (from highest to lowest priority): IF "conduit" is supported THEN "conduit" is the default. ELSEIF configured using '--with-host-detect=...' THEN the "..." is the default. ELSEIF gethostid() is available THEN "hostid" is the default. ELSE "hostname" is the default. * see conduit-specific documentation in conduit directories for more settings GASNet exit =========== GASNet clients desiring robustness and portability should not call _exit(). Normal process termination should be done using gasnet_exit(), exit() or return from main(). Use of these paths allows GASNet to ensure proper shutdown. This includes efforts to avoid zombie/orphan processes in the case of non-collective exits, and proper release of network resources. This second point is important because we are aware of multiple network stacks which can permanently leak system resources if they are not explicitly released. We acknowledge that this represents a restriction that may be hard to satisfy. We fully intend to separate library termination from process termination in a future release. GASNet tracing & statistical collection ======================================= GASNet includes an extensive tracing and statistical collection utility for monitoring communication events in GASNet applications. To use, configure with --enable-stats and/or --enable-trace (or --enable-debug, which implies both of these by default), and run your program as usual. In order to see the trace/stats output, you must set the environment variable GASNET_TRACEFILE and/or GASNET_STATSFILE as explained below. Note that system performance is likely to be degraded as a result of tracing and statistical collection. This is still true even when output is disabled by not setting this GASNET_TRACEFILE/GASNET_STATSFILE (so production builds should not enable tracing/stats at GASNet configure time). Optional environment variable settings: GASNET_TRACEFILE - specify a file name to receive the trace output may also be "stdout" or "stderr", (or "-" to indicate stderr) each node may have its output directed to a separate file, and any '%' character in the value is replaced by the node number at runtime (e.g. GASNET_TRACEFILE="mytrace-%") unsetting this environment variable (or setting it to empty) disables tracing output (although the trace code still has performance impact) GASNET_TRACENODES - specify an optional list of nodes on which to generate tracing output (if enabled by GASNET_TRACEFILE). List may contain one or more integers or ranges separated by commas, such as "0,2-4,6". If empty or equal to "*" then all nodes may generate tracing output. GASNET_STATSFILE - specify a filename to receive statistical output operates analogously to GASNET_TRACEFILE GASNET_STATSNODES - limit nodes to generate statistical output operates analogously to GASNET_TRACENODES GASNET_TRACEMASK - specify the types of trace messages to report A string containing one or more of the following letters: G - gets P - puts R - remote atomics S - non-blocking synchronization W - collective operations (excluding barriers) B - barriers L - locks A - AM requests/replies (and handler execution, if conduit-supported) X - AMPoll I - informational messages about system status or performance alerts O - Object creation, modification and destruction C - conduit-specific (low-level) messages D - Detailed message data for gets/puts/AMreqrep N - Line number information from client source files H - High-level messages from the client U - Unsuppressable messages, which are always output (use with caution) default: (all of the above) GASNET_STATSMASK - specify the types of statistics to collect and report operates analogously to GASNET_TRACEMASK GASNET_TRACEFLUSH - set this variable to force a file system flush after every write to the tracefile. This seriously degrades tracing performance, but ensures any final trace messages before a crash are flushed into the tracefile. GASNET_TRACELOCAL - set to control whether the PG trace messages for local (i.e. loopback) put/get operations are entered into the trace file or suppressed (they are included by default). This can be used to reduce tracing overhead and trace file size for clients with a large number of local put/get operations which are not of interest. GASNet Collectives ================== GASNet includes interfaces for collective operations, which is still evolving. The design for the interfaces is located under the source directory in docs/collective_notes.txt. Anyone planning to implement a client that uses these collective should contact us (gasnet-devel@lbl.gov) first to determine the completeness and stability of the relevant portions of the implementation. In GASNet 1.14.0 we introduce a mechanism for auto-tuning of collectives in which a variety of algorithms are tried for each collective operation in a user's application and the best choice can be recorded in a file for use in future runs. For more information, see the file autotuner.txt in the docs directory. GASNet debug malloc services ============================ GASNet includes a debug malloc implementation that can be used to find local heap corruption bugs in GASNet itself and also in GASNet client applications (notably Berkeley UPC). To use, configure GASNet with --enable-debug-malloc (or --enable-debug which implies --enable-debug-malloc), and run your program as usual. GASNet will regularly scan the heap for corruption at malloc events and AMPoll's and report any detected problems. If you're a GASNet client writer, you should see the gasnett_debug_malloc functions in gasnet_tools.h to tie your applications into the GASNet debug malloc services. Note the debug malloc implementation imposes some CPU and memory consumption overhead relative to the system malloc implementation (it's implemented as a thin wrapper around the system malloc). This means heap behavior is likely to be somewhat perturbed relative to the normal mode of operation, however the overhead is probably less than it would be with third-party heap corruption tools such as efence or purify (at some cost in lost precision). Optional environment variable settings (recognized only when debugging malloc is enabled): GASNET_MALLOC_INIT: When set to 1, every byte of allocated memory is initialized to the value specified by GASNET_MALLOC_INITVAL. This can be useful for detecting use of uninitialized data. GASNET_MALLOC_CLOBBER: When set to 1, every byte of freed memory is overwritten with the value specified by GASNET_MALLOC_CLOBBERVAL. This can be useful for detecting inadvertent use of freed data. GASNET_MALLOC_INITVAL, GASNET_MALLOC_CLOBBERVAL: The data value to use for allocation init or free clobbering, respectively. The value may be specified as a decimal integer matching the pattern: "-?[0-9]+" or a hexadecimal integer matching the pattern: "0x[0-9A-F]+". If the value is between 0 and 255 (inclusive) it is taken to be an 8-bit value which is used to overwrite every byte in the given region. Otherwise, it is taken to be a 64-bit value which is used to overwrite the given region in 8-byte chunks. The value may also be one of "NAN", "sNAN" or "qNAN", which select a double-precision signaling or quiet NAN, which is likely to propagate through floating point calculations and cause segmentation faults when used as a pointer. Both values default to "NAN". GASNET_MALLOC_LEAKALL: set to 1 to leak all freed objects, ensuring they are not re-allocated during subsequent mallocs. This has an obvious cost in memory consumption, but can be helpful for tracking bugs in conjunction with GASNET_MALLOC_CLOBBER for tracking usage of dead objects. GASNET_MALLOC_SCANFREED: set to 1 to enable scanning of freed objects, to detect write-after-free errors. This option implies GASNET_MALLOC_LEAKALL and GASNET_MALLOC_CLOBBER. GASNET_MALLOC_EXTRACHECK: set to 1 to enable more frequent checking for memory corruption (at a cost in performance) GASNET_MALLOCFILE: specify a file name to receive a report of malloc heap behavior and utilization at process exit. Includes a list of leaked objects. The filename may also be "stdout" or "stderr", (or "-" to indicate stderr) Each node may have its output directed to a separate file, and any '%' character in the value is replaced by the node number at runtime (e.g. GASNET_MALLOCFILE="mallocreport-%"). These output files are human-readable, but can be passed to gasnet_trace for summarization. GASNET_MALLOCNODES: limit which nodes will generate a malloc report. operates analogously to GASNET_TRACENODES GASNet inter-Process SHared Memory (PSHM) ========================================= GASNet's PSHM support provides mechanism to communicate through shared memory among processes on the same compute node (ie sharing a cache-coherent physical memory). Inter-process communication through shared memory is usually the fastest way for processes sharing a compute node to communicate. GASNet clients are free to use any combination of processes communicating through PSHM and client pthreads within processes (in GASNET_PAR mode) to implement on-node parallelism, and all such execution contexts may perform network communication (with an appropriate conduit). GASNet's PSHM support is enabled by default unless one is cross-compiling, or running Cygwin prior to version 2.0. However, most cross-compilation scripts enable it with any additional settings as required. If desired, PSHM can be disabled by passing --disable-pshm at configure time. In addition to configure-time control, the environment variable GASNET_SUPERNODE_MAXSIZE sets a limit on the number of processes that will be grouped into a shared-memory "supernode". By default all co-located processes are grouped into one supernode, but the value can be set to 1 to disable the use of shared-memory for communication. The PSHM support in GASNet can operate via three generic mechanisms: POSIX shared memory, SystemV shared memory, or mmap()ed disk files. POSIX shared memory is recommended as the best option in most cases, but is not always available (many kernels can be configured not to support it), or may not permit large-enough allocations. See "System Settings for POSIX Shared Memory", below, for information on sizing of allocations. In the absence of the POSIX shared memory, users are advised to use the SystemV shared memory as the next-best option. However, see "System Settings for SystemV Shared Memory", below, for information about configuring a system to use SystemV shared memory. In the absence of both POSIX and SystemV shared memory, a user may try using mmap()ed disk files. However, on some systems we see significant performance degradation when using files (apparently due to committing the changes from memory to disk). Unless PSHM is disabled, the default behavior of the configure step is to probe for support via exactly one preferred mechanism. For Solaris the preferred mechanism is SystemV, while for all other platforms it is POSIX. If the preferred mechanism is not supported (or is explicitly disabled) the then there is no automatic fallback to any other mechanism. So, if one wishes to use another mechanism, one should explicitly disable the default (POSIX or SysV) support and enable the desired mechanism: Common Usage Summary (flags to pass to the configure script): ---------------------------------------------------------- OFF: --disable-pshm POSIX: no flags required (default unless cross-compiling) SYSV: --disable-pshm-posix --enable-pshm-sysv FILE: --disable-pshm-posix --enable-pshm-file Solaris Usage Summary (as above except that SysV is default) ---------------------------------------------------------- OFF: --disable-pshm POSIX: --disable-pshm-sysv --enable-pshm-posix SYSV: no flags required (default) FILE: --disable-pshm-sysv --enable-pshm-file There are also implementations which are specific to certain platforms: For Cray XE, XK and XC systems, PSHM-over-XPMEM is enabled automatically by the corresponding cross-configure scripts. However, one can also optionally enable PSHM-over-XPMEM on an SGI Altix as follows: XPMEM: --enable-pshm --disable-pshm-posix --enable-pshm-xpmem For Linux systems with libhugetlbfs and a hugetlbfs mount, there is also experimental support for mmap()ed files on hugetlbfs, which can be enabled with: HUGETLBFS: --disable-pshm-posix --enable-pshm-hugetlbfs System Settings for POSIX Shared Memory: --------------------------------------- On most systems (all but Solaris and some cross-compiled platforms), the default implementation of PSHM uses POSIX shared memory. On many operating systems the amount of available POSIX shared memory is controlled by the sizing of a pseudo-filesystem that consumes space in memory (and sometimes swap space) rather than stable storage. If this filesystem is not large enough, it can limit the amount of POSIX shared memory which can be allocated for the GASNet segment. Insufficient available POSIX shared memory may either lead to failures at start-up, or to SIGBUS or SIGSEGV later in a run (if the OS has permitted allocation of more virtual address space than is actually available). If one encounters either of these failure modes when GASNet is configured to use POSIX shared memory, then one should check the space available (see below) and may need to increase the corresponding system settings. Setting these parameters is system-specific and requires administrator privileges. On most modern Linux distribution, POSIX shared memory allocations reside in the /dev/shm filesystem (of type 'tmpfs'), though /var/shm and /run/shm are also used in some cases. The mechanism for sizing of this filesystem varies greatly between distributions. Please consult the documentation for your specific Linux distribution for instructions to resize. In particular, be advised that distributions may mount this filesystem early in the boot process, without regards to any entry in /etc/fstab. On macOS, Cygwin and Solaris, there are no settings (known to us at this time) required to increase the space available for allocation as POSIX shared memory. On FreeBSD, POSIX shared memory allocations reside in the /tmp filesystem, which is normally of type 'tmpfs'. This filesystem often defaults to having no limit on its size other than the sum of physical memory and swap space. However, the size can be set in /etc/fstab. See the fstab(5) and tmpfs(5) manpages for more information. On NetBSD, POSIX shared memory allocations reside in the /var/shm filesystem (of type 'tmpfs'), which can be sized in /etc/fstab. See the fstab(5) and mount_tmpfs(8) manpages for more information. On OpenBSD, POSIX shared memory allocations reside in the /tmp filesystem, which is normally a filesystem of type 'mfs' and can be sized in /etc/fstab. See the fstab(5) and mount_mfs(8) manpages for more information. System Settings for SystemV Shared Memory: ----------------------------------------- SystemV shared memory is the second most widely used implementation of PSHM after POSIX shared memory. It is used by default on Solaris, and is recommended on most other systems when POSIX does not work for any reason. On most operating systems the amount of available SystemV shared memory and the number of shared memory segments is controlled by the kernel parameters: shmmax, shmall and shmmni. shmmax = largest size of a shared memory segment (in bytes) shmall = total amount of memory allocatable as shared (in pages) shmmni = maximum number of shared memory segments Insufficient amount of SystemV shared memory will lead to failures at start-up of any application using a runtime configured to use PSHM over SystemV. Setting these parameters is system-specific and requires administrator privileges. * Examples of configuration: These examples are for 64-bit systems, and essentially allow unlimited amounts of memory to be used for GASNet segments. These assume you are running no more than 127 processes per node. If you have more cores (or cpu threads) then increase the shmmni value to the maximum number of processes you wish to be be able to run, plus 1. Some examples below do not include a setting for shmmni because the system defaults are large enough in general, and often cannot be reduced. For 32-bit systems, use 2147483647 for the shmmax value to avoid a potential overflow. - Linux: Add the following three lines to /etc/sysctl.conf (sudo required): kernel.shmmax=1099511627776 kernel.shmall=268435456 kernel.shmmni=128 To activate the new settings reboot, or run the following: sudo /sbin/sysctl -p /etc/sysctl.conf - macOS: Add the following three lines to /etc/sysctl.conf (sudo required): kern.sysv.shmmax=1099511627776 kern.sysv.shmall=268435456 kern.sysv.shmmni=128 To activate the new settings reboot, or run the following: grep ^kern.sysv.shm /etc/sysctl.conf | sudo xargs /usr/sbin/sysctl -w - FreeBSD: Add the following two lines to /etc/sysctl.conf (sudo required): kern.ipc.shmmax=1099511627776 kern.ipc.shmall=268435456 To activate the new settings reboot, or run the following: sudo /sbin/sysctl -f /etc/sysctl.conf - NetBSD: Add the following two lines to /etc/sysctl.conf (sudo required): kern.ipc.shmmax=1099511627776 kern.ipc.shmmaxpgs=268435456 To activate the new settings reboot, or run the following: sudo /sbin/sysctl -f /etc/sysctl.conf - OpenBSD: Add the following two lines to /etc/sysctl.conf (sudo required): kern.shminfo.shmmax=2147483647 kern.shminfo.shmall=268435456 To activate the new settings reboot, or run the following: grep ^kern.shminfo /etc/sysctl.conf | sudo xargs /sbin/sysctl -w - Solaris: Depends on version. Please see the Sun/Oracle documentation. However, on recent releases the defaults are sufficient. - Cygwin: If you have not done so yet, please Read the Cygwin documentation on "cygserver-config" to create an initial /etc/cygsever.conf and start the server as a Windows service (optional). You may then edit /etc/cygserver to edit kern.ipc.shmmaxpgs kern.ipc.shmmni kern.ipc.shmseg Except for shmmaxpgs, the defaults are often large enough. These configuration values are only read when cygserver starts. So, read the Cygwin documentation to determine if/how to restart the cygserver service. Under Cygwin-1.5 you may also need to add "server" to the value of the CYGWIN environment variable. Again, you should see the Cygwin documentation for more information on this subject. Warning to users of macOS: ------------------------- We have evidence which suggests that there is a kernel bug in all releases of macOS though at least 10.12 (Sierra) which leads to a small leak of kernel memory (perhaps a few 10s of bytes) each time POSIX or SystemV shared memory is used. This may, after tens of thousands of runs of GASNet applications using PSHM, lead to a low memory condition contributing to slow performance and eventually to total memory exhaustion. In normal usage, it is reasonable to expect that a Mac laptop or desktop system will be rebooted for software updates frequently enough that this leak will not impact normal users. Also note that --enable-pshm-file is NOT recommended for use on macOS, where it has been seen to cause VERY slow startup times. IMPORTANT: SYSTEM CLEANING OF PSHM OBJECTS: ------------------------------------------ If a GASNet application using PSHM is terminated before ending the initialization phase, there is a possibility that the shared memory objects will remain in the system. A large amount of memory or disk space can remain allocated, preventing users from fully utilizing all available hardware resources. In the SystemV case, the allocated (but not released) shared memory segments can be listed via the "ipcs" command, and can be removed via the "ipcrm" command. Note that on the systems with a batch scheduler, the "ipcs" and "ipcrm" instructions need to be run on the compute nodes. In the mmap()ed file case, the allocated but not released shared memory files can be found in the directory pointed by the TMPDIR environment variable (default: /tmp). These files are named with the prefix GASNT (the lack of an 'E' is not a typo), and can be deleted using the "rm" command. The case of mmap()ed file on hugetlbfs is like the regular mmap()ed file case except that the directory will be the hugetlbfs mount point. This is typically something like /dev/hugepages. If uncertain of the mount point, the command "mount|grep hugetlbfs" should locate it for you. The case of POSIX shared memory is OS-specific, but in most cases the shared memory objects are visible in the file system. On systems we use for development, we have observed the following defaults: * Linux: located in /dev/shm with a filename prefix of "GASNT" * Cygwin: located in /dev/shm with a filename prefix of "GASNT" * Solaris: located in /tmp with a filename prefix of ".SHMDGASNT" * OpenBSD: located in /tmp with long random names suffixed by ".shm" * NetBSD: located in /var/shm with a filename prefix of ".shmobj_GASNT" MPI Interoperability ==================== The Message Passing Interface (MPI) is a ubiquitous portable interface for communication in scientific computing and is used both directly and indirectly by various applications and libraries. In most cases GASNet is implemented natively on the various SAN interconnects, bypassing the MPI layer in order to provide the best possible performance. This means that in general GASNet does NOT use MPI for performing communication (the only notable exception being mpi-conduit, which is a portable implementation of GASNet over MPI). In some cases GASNet can use MPI to assist in parallel job creation and termination (thereby taking advantage of the existing MPI infrastructure), but in those cases it will not use MPI during the steady state of operation - only at startup and shutdown. Although GASNet is not implemented using MPI, it is *compatible* with MPI - GASNet and MPI can be used together within the same network and even within the same program. However, there are a few subtle issues to be dealt with when using both communication systems in the same process, because in those cases GASNet and MPI must both be initialized properly and cooperate to "share" the resources of the network and physical memory. This section describes issues that arise when combining the use of MPI-based communication and GASNet communication within a single process. The information described here applies to any case where a single process attempts to access the network hardware using both GASNet and MPI, even when such usage occurs at different levels within a layered application (for example, a UPC application which uses GASNet via the Berkeley UPC compiler and makes calls to a helper library that communicates using MPI). MPI requires exactly one initialization call before MPI communication can be used within a process. The GASNet implementation may or may not automatically initialize the MPI library, depending on the conduit and various configuration settings. Therefore, processes using both communication libraries must be prepared to correctly handle either situation - the recommended way to arbitrate this is using the MPI_Initialized() call which reports whether or not the MPI layer has been initialized. Here is some example code for accomplishing this: int isMPIinit; int main(int argc, char **argv) { size_t segment_size = 64*1024*1024; /* want 64MB segment in this example */ size_t segment_max; gasnet_init(&argc, &argv); segment_max = gasnet_getMaxGlobalSegmentSize(); if (segment_size > segment_max) segment_size = segment_max; gasnet_attach(NULL, 0, segment_size, 0); gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS); gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS); if (MPI_Initialized(&isMPIinit) != MPI_OK) { /* test if MPI already init */ fprintf(stderr, "Error calling MPI_Initialized()\n"); abort(); } if (!isMPIinit) MPI_Init(argc, argv); /* MPI not init, so do it */ /* ... use MPI as usual ... */ MPI_Barrier(MPI_COMM_WORLD); /* ... use GASNet as usual ... */ gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS); gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS); ... if (!isMPIinit) MPI_Finalize(); return 0; } The MPI_Initialized call checks to make sure MPI hasn't already been initialized by GASNet, which will be the case on mpi-conduit and possibly on other conduits that use MPI for job startup. Any MPI usage internal to GASNet uses a private MPI communicator, preventing any potential interference with MPI use by the GASNet client. The code which initializes the MPI layer is also responsible for finalizing it before exit. An extensive example is available in tests/testmpi.c, which implements various interoperability tests between GASNet and MPI. There are a few other important caveats to be aware of when mixing GASNet and MPI: * GASNet must be configured with MPI_CC set to the exact *same* MPI installation that will be used when building any GASNet client code using MPI. Note that many MPI implementations are not binary-compatible across versions, so upgrading the MPI compiler usually requires a fresh configure/build/install of GASNet with the new MPI implementation if you intend to mix GASNet and MPI. * When a node is inside a blocking MPI call, most conduits will be prevented from servicing many GASNet requests from remote nodes. Similarly, GASNet blocking calls will generally not ensure progress of MPI communication. This means that arbitrarily interleaving blocking MPI calls with GASNet calls (e.g. UPC shared accesses) can easily lead to deadlock. In order to ensure safety and portably correct operation across networks, GASNet communication operations and user MPI calls should be isolated from each other by barriers, by strictly adhering to the following "Time-Phasing" protocol: - When the application starts, the first MPI or GASNet call issued from any node should be considered to put the application in 'MPI' or 'GASNet' mode, respectively. - When an application is in 'MPI' mode, and needs to switch to using GASNet, it should collectively execute an 'MPI_Barrier()' as the last MPI call before issuing any GASNet communication calls. Once any GASNet communication has occurred from any node, the program should be considered to have switched to 'GASNet' mode. - When an application is in 'GASNet' mode, and an MPI call that may cause network traffic is needed, a collective call to 'gasnet_barrier_notify()' followed by a 'gasnet_barrier_wait()' should be executed as the last GASNet communication before any MPI calls are made. Once any MPI functions have been called from any thread, the program should be considered to be in 'MPI' mode. If this simple construct--GASNet code must be followed by a GASNet barrier, and MPI code must be followed by an MPI_Barrier--is followed, this deadlock should not occur. This protocol is demonstrated in the code above. * GASNet and MPI both number processes with a unique non-negative index (i.e. gasnet_mynode() and MPI_Comm_rank(MPI_COMM_WORLD)). Every attempt is made to ensure this numbering matches between layers, but it might not always be possible to ensure this. To maximize portability, apps using both layers should use a GASNet or MPI collective (e.g. AllGather) to build a table of ids and translate rank identifiers as needed. * Note that GASNet language clients such as UPC or Titanium might include compilation modes where multiple language-level threads are implemented as pthreads within a single process, which will map to a single GASNet node id (and hence a single MPI rank). Clients using such threading should ensure they are using a pthread-safe implementation of MPI (GASNet is thread-safe by virtue of PAR mode), and be aware that MPI messages sent to a particular node might be received by any pthread at the target (unless steps are taken to avoid this, e.g. using an MPI tag). * There are various network-specific issues that may arise when combining GASNet with specific MPI implementations, most notably related to job creation and spawning mechanisms. Please see the README file in each conduit directory for network-specific discussion of issues with MPI interoperability. Contact Info and Support ======================== For the latest GASNet downloads, publications, and specifications, please visit: https://gasnet.lbl.gov For bug reports and feature requests, please submit a ticket in the GASNet Bugzilla: https://gasnet-bugs.lbl.gov You may find an instant solution to your problem by searching the bug database! GASNet has several mailing lists for support: gasnet-users@lbl.gov General questions or inquiries regarding the installation or use of GASNet. Any users can join the list from the web site above. gasnet-devel@lbl.gov GASNet developers list (multi-institution) gasnet-announce@lbl.gov GASNet release announcements. Join on the web site above. -------------------------------------------------------------------------- The canonical version of this document is located here: https://bitbucket.org/berkeleylab/gasnet/src/master/README For more information, please email: gasnet-users@lbl.gov or visit the GASNet home page at: https://gasnet.lbl.gov --------------------------------------------------------------------------