AMUDP Documentation =================== Author: Dan Bonachea and Dan Hettena Contact email: gasnet-devel@lbl.gov Home page: https://gasnet.lbl.gov/amudp AMUDP is an implementation of the AM-2 specification over UDP/IP. AMUDP is a portable implementation of the AM-2 specification that runs on UDP, a standard component of the TCP/IP protocol suite that is ubiquitous across platforms and operating systems. The intent is a fully-portable implementation that will work on any POSIX-like system. We don't expect to achieve latency performance competitive with a native implementation of AM optimized for special purpose hardware, instead we seek to provide a compatibility layer that will allow AM-based systems to quickly get up and running on virtually any platform. The motivation for choosing UDP over other protocols (such as TCP) is that it typically provides the lowest overhead access to the network with little or no internal buffering, and the connectionless model is best suited for scaling up the number of distributed processors. Because UDP occasionally drops packets, we add a thin reliability layer that provides the guaranteed delivery required by AM-2, hopefully providing this fault tolerance with better performance than full-blown TCP. Design documentation for the original AMUDP implementation is available from the website above. The current version of AMUDP strictly targets the BSD/POSIX sockets layer - it no longer supports UFXP or native MS Windows sockets, although Windows support is available via the Cygwin POSIX environment for Windows (http://cygwin.com). AMUDP includes an extension interface called AMUDP_SPMD that performs job spawning and initialization services, using either ssh or a provided site-specific job spawner. See the paper above and test code for details. AMUDP is used to implement the GASNet communication system (https://gasnet.lbl.gov). Further documentation about using AMUDP is available in the udp-conduit documentation: https://gasnet.lbl.gov/dist/udp-conduit/README Requirements ------------ * C99 and C++98 compilers. The C++ STL is not used in any way. * A POSIX-like environment, including BSD/POSIX sockets and file descriptors * GNU Make and basic UNIX tools for the provided Makefiles Limitations ----------- AMUDP has a few notable departures from the AM-2 specification: * The client must call AM_SetExpectedResources() exactly once on an endpoint after setting up the translation table and before making any calls to the transport functions (AM_Poll, AM_Reply* or AM_Request*). It is also an error to call AM_Map, AM_MapAny or AM_Unmap (which change the translation table) after the call to AM_SetExpectedResources(). * AM_PAR bundle/endpoint access is not implemented - this means AMUDP does not provide thread safety, but can be used in a thread-funnelled mode with locking at the client level. * The nbytes argument to Medium and Long AM handlers has type `size_t` * AM_GetXferM is not implemented. * Clients should not take the address of AM_ entry point "functions", because many are implemented as macros. Change Log ---------- AMUDP 3.20 (12/2023) - Implement new extension AMX_GetTokenInfo() - Improve warning messages about environment variables to more uniformly reference the prefixed variable name. - Improved error reporting from the default AM handler AMUDP 3.19 (09/2022) - amudprun now supports multiple verbosity levels through repeated -v arguments - New envvar AMUDP_SPAWN_VERBOSE now controls spawn-related verbose output, AMUDP_VERBOSEENV now specifically controls envvar-related output. - New envvars SOCKETBUFFER_{INITIAL,MAX} now allow manual control over library requests made for kernel-level socket buffers. AMUDP 3.18 (03/2021) - Add AMUDP_SPMDSetProc and envvar WORKER_RANK to support explicit rank assignment AMUDP 3.17 (10/2020) - Remove BLCR support - This version breaks compatibility of the amudprun spawner with prior versions AMUDP 3.16 (03/2020) - Improved startup error checking and reporting for DNS misconfigurations AMUDP 3.15 (06/2018) - Medium and Long AM handlers change in signature on 64-bit platforms: the nbytes argument now has type size_t instead of int - New AMUDP_LIBRARY_VERSION_{MAJOR,MINOR} defines - Fix compatibility issues with PGI C++ 17+ - Significant rearrangement of internal files and utilities - AM Handler dispatching code is now strictly spec-compliant - AMX_ENV_PREFIX_STR is now the preferred define for setting a client ENV prefix - Several pseudo-public programmatic knobs renamed into the AMX_ namespace - Bug 2774: source_addr to Med/Long injection is now ignored for nbytes=0 - Fix an unreported cosmetic output bug in AMX_RETURN error reporting AMUDP 3.14 (03/2017) - Fix AM_{Get,Set}NumHandlers and AM_MoveEndpoint - Fix bug 3379: workaround a PGI optimizer bug on Darwin - Export AMUDP_enEqual - make run-tests now outputs summary results at the end - Fix some harmless warnings about unused variables - Renamed portable_platform.h - Remove Makefile.darwin (no longer necessary) - Fix compilation errors in verbose debug mode - Update contact email AMUDP 3.13 (10/2016) - Restructure network buffer management to provide scalable buffer memory utilization to thousands of nodes and beyond. - Adjust the set of available environment tuning knobs for better control of transfer performance at scale. - Adjust default transfer parameters for better performance on 1GbE - Numerous misc changes to improve performance - Restructure AMUDP headers to reduce namespace clutter - Expose AMUDP_SetTranslationTag - Fix possible syntax errors with AM expression arguments - Fix a compatibility issue with Cray CNL - Fix a use-after-free bug in argument processing - Fix a data corruption issue on 64-bit OSX on PowerPC - Use realloc when appropriate to manage internal state AMUDP 3.12 (06/2016) - Add BLCR support - Improve stdout/stderr routing to be more robust and use fewer descriptors - Add AMUDP_ENV_CMD and AMUDP_LINEBUFFERSZ - Remove UETH/UFXP support - Remove GLUNIX and REXEC spawn support - Remove support for HPUX, SuperUX, MTX, Catamount, and native Win32 (use cygwin instead) - Remove some obsolete files Fix three significant bugs: - Out-of-order delivery of UDP messages (which can occur on networks with multiple paths) was going undetected and could lead to silent loss and/or erroneous redelivery of AMs in the presence of retransmissions (caused by packet loss or high network load). - On Cygwin 2.5 (2016-04-11 and later), ioctl(FIONREAD) always returns failure, was rendering AMUDP entirely unusable on that platform. - On FreeBSD, ioctl(FIONREAD) may truncate long datagrams to around 600 bytes, was rendering AMUDP unusable on that platform.