README/Release Notes OFED 4.8 DAPL Release 2.1.10 Dec 2016 User space libraries/utilities for Direct Access Transport (DAT) v2.0. DAT is a transport-independent, platform-independent Application Programming Interface that supports RDMA (remote direct memory access) devices. Note: v1.2 is no longer supported and will not be included with OFED releases MIC support is provided with the new MCM provider and MPXYD service, since dapl-2.1.0. MCM requires the Intel(R) MPSS 3.x (YOCTO) release for Linux to be installed on your system. MPSS 3.x for Linux can be downloaded from: http://software.intel.com/mic-developer For latest documentation and packages: //www.openfabrics.org/downloads/dapl/ ================= 1.0 Release Notes ================= dapl-2.1.10 changes include bug fixes, CCL Proxy post_send optimizations, dtest/dtest_suite improvements dapl-2.1.9-2 changes include bug fixes, MFO fixes, GID query attributes dapl-2.1.9-1 changes include dtestcm improvement, bug fixes for MIC, Fix SCM interop issue with MTU dapl-2.1.8 changes include dtest improvement, adding Intel OPA support, and MTU negotiation - Add dtest -D option for data check/validation via pingpong test - Add dtest -q option for open/query only option - Add support for new hfi (Intel OPA) driver, all providers CCL Proxy can support hfi in MFO mode via mcm provider - Active MTU is now set as default QP MTU, CM mechanism added to negotiate and fallback to smaller size of endpoints. Is backward compatible with older providers running defaults of 2K and can be adjusted with pre-existing DAPL_IB_MTU environment setting. - Bug fixes, including CCL Data corruption (scif_writeto ordering requirements). - Provide new test/dtest/dtest_suite.sh script for automated host to mic test/validation. dapl-2.1.7 changes include dtest improvement and CCL Proxy P2P inline support: - Add dtest -W option for rdma write pinpong test, new options with -W include -a (all data sizes) -i (incremental size) - CCL Proxy small message latency improvment with Proxy2Proxy inline support for message sizes < 96 bytes, reduces MPI pingpong single byte latency for MFO devices by 27%. dapl-2.1.6 changes include MIC support for full offload mode - Add support for Truescale qib devices with no CCL Direct verbs support on MIC. - Enhancement for inside the box transfers without IB adapter via ibscif. - Add DAPL_NETWORK_NODES, DAPL_NETWORK_PPN environment variables. dapl-2.1.5 changes include improvements for large scale UD communication management: - AH caching, reduced memory footprint (grows as needed) - Port space increased to 24 bits - Hash table for port space, CM object management - Optimized CM wire protocol for fast index lookup Tested on 1200n 28ppn cluster, AlltoAll Intel MPI, UD mode. Both static and dynamic modes, over 500m UD QP connections. ========== 2.0 BUILD: ========== # NON_DEBUG build/install example for x86_64, OFED targets ./configure --prefix /usr --sysconf=/etc --libdir /usr/lib64 LDFLAGS=-L/usr/lib64 CPPFLAGS="-I/usr/include" make install # DEBUG build/install example for x86_64, using OFED targets ./configure --enable-debug --prefix /usr --sysconf=/etc --libdir /usr/lib64 LDFLAGS=-L/usr/lib64 CPPFLAGS="-I/usr/include" make install # COUNTERS build/install example for x86_64, using OFED targets ./configure --prefix /usr --sysconf=/etc --libdir /usr/lib64 LDFLAGS=-L/usr/lib64 CPPFLAGS="-I/usr/include -DDAPL_COUNTERS" make install ========================================================= 3.0 Provider descriptions and CM results (cma, scm, ucm): ========================================================= 1. CMA - uses OFA rdma_cm to setup QP's. IPoIB, ARP, and SA queries required. Provider name: ofa-v2-ib0 PROs: OFA rdma_cm has the most testing across many applications. Supports both iWARP and IB. CONs: Serialization of conn processing with kernel based CM service Requires IPoIB ARP for name resolution, storms Requires SA for path record queries for IB fabrics. Conn Request private data limited to 52 bytes. Settings for larger clusters (512+ cores): setenv DAPL_CM_ROUTE_TIMEOUT_MS 20000 setenv DAPL_CM_ARP_TIMEOUT_MS 10000 2. SCM - uses sockets to exchange QP information. IPoIB, ARP, and SA queries NOT required. Provider name (connectx): ofa-v2-mlx4_0-1 PROs: Each rank has own instance of socket cm. More private data with requests. Doesn't require path-record lookup. CONs: Socket resources grow with scale-out, serialization of connections with kernel based tcp sockets, Competes for MPI socket resources/port space and other TCP applications. Sockets remain in TIMEWAIT state for minutes after closure. Requires ARP for name resolution. Doesn't support iWARP devices. Settings for larger clusters (512+ cores): setenv DAPL_ACK_RETRY 7 /* IB RC Ack retry count */ setenv DAPL_ACK_TIMER 20 /* IB RC Ack retry timer */ 3. UCM - use's IB UD QP to exchange QP info. Sockets, ARP, IPoIB, and SA queries NOT required. Provider name (connectx): ofa-v2-mlx4_0-1u PROs: Each rank has own instance of CM in user process Resources fixed per rank regardless of scale-out size No serialization of user or kernel resources establishing connections, Simple 3-way msg handsake, CM messages fit in inline data for lowest message latency, Supports alternate paths No address resolution required. No path resolution required. CONs: New provider with limited testing, a little tougher to debug. Doesn't support iWARP Settings for larger clusters (512+ cores): setenv DAPL_UCM_REP_TIME 10000 /* REQUEST timer, waiting for REPLY in millisecs */ setenv DAPL_UCM_RTU_TIME 10000 /* REPLY timer, waiting for RTU in millisecs */ setenv DAPL_UCM_CQ_SIZE 2000 /* CM completion queue */ setenv DAPL_UCM_QP_SIZE 2000 /* CM message queue */ setenv DAPL_UCM_RETRY 7 /* REQUEST and REPLY retries */ setenv DAPL_ACK_RETRY 7 /* IB RC Ack retry count */ setenv DAPL_ACK_TIMER 20 /* IB RC Ack retry timer */ CM Performance: CPS profile for cma, scm, and ucm v2 uDAPL providers: ----------------------------------------------------------------------- Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz (IVT) Mellanox MLX4 IB FDR, no switch. dtestcm (server/client): cma: Connections: 313.10 usec, CPS 3193.83 Total 0.31 secs, poll_cnt=6300, Num=1000 scm: Connections: 167.65 usec, CPS 5964.92 Total 0.17 secs, poll_cnt=2394, Num=1000 ucm: Connections: 71.85 usec, CPS 13918.06 Total 0.07 secs, poll_cnt=2360, Num=1000 dapl_cm_bw: MPI uDAPL/CM profiling application (all-to-all connections, all ranks) CMA 2 Connect times (10): Total 0.0049 per 0.0005 CPS=2051.38 4 Connect times (40): Total 0.0151 per 0.0004 CPS=2650.16 8 Connect times (240): Total 0.0548 per 0.0002 CPS=4380.59 16 Connect times (1120): Total 4.0356 per 0.0036 CPS=277.53 32 Connect times (4800): Total 4.4704 per 0.0009 CPS=1073.72 SCM 2 Connect times (10): Total 0.0029 per 0.0003 CPS=3441.31 4 Connect times (40): Total 0.0060 per 0.0002 CPS=6635.97 8 Connect times (240): Total 0.0194 per 0.0001 CPS=12383.47 16 Connect times (1120): Total 0.0649 per 0.0001 CPS=17246.93 32 Connect times (4800): Total 1.0193 per 0.0002 CPS=4708.95 UCM 2 Connect times (10): Total 0.0014 per 0.0001 CPS=6993.91 4 Connect times (40): Total 0.0045 per 0.0001 CPS=8837.87 8 Connect times (240): Total 0.0155 per 0.0001 CPS=15477.13 16 Connect times (1120): Total 0.0630 per 0.0001 CPS=17765.12 32 Connect times (4800): Total 0.2632 per 0.0001 CPS=18236.54 =================================================================================================== 4.0 BKM for installing new DAPL library on your cluster without any impact on existing OFED install: ==================================================================================================== Note: example for user /home/user1, (assumes /home/user1 is exported) and MLX4 adapter, port 1 Download latest 2.1.x package: http://www.openfabrics.org/downloads/dapl/dapl-2.1.x.tar.gz untar in /home/user1 cd /home/user1/dapl-2.1.x ./configure LDFLAGS=-L/usr/lib64 CPPFLAGS="-I/usr/include" make Create /home/user1/dat.conf with following 3 lines. (entries with path to new libraries): ofa-v2-mlx4_0-1u u2.0 nonthreadsafe default /home/user1/dapl-2.1.x/dapl/udapl/.libs/libdaploucm.so.2 dapl.2.0 "mlx4_0 1" "" ofa-v2-mlx4_0-1m u2.0 nonthreadsafe default /home/user1/dapl-2.1.x/dapl/udapl/.libs/libdaplomcm.so.2 dapl.2.0 "mlx4_0 1" "" ofa-v2-mlx4_0-1 u2.0 nonthreadsafe default /home/user1/dapl-2.1.x/dapl/udapl/.libs/libdaploscm.so.2 dapl.2.0 "mlx4_0 1" "" ofa-v2-ib0 u2.0 nonthreadsafe default /home/user1/dapl-2.1.x/dapl/udapl/.libs/libdaplcma.so.1 dapl.2.0 "ib0 0" "" Run uDAPL application or Intel MPI that uses uDAPL, with (assuming mlx4_0 adapters) following: setenv DAT_OVERRIDE=/home/user1/dat.conf setenv LD_LIBRARY_PATH=/home/user1/dapl-2.1.5/dapl/udapl/.libs:$LD_LIBRARY_PATH If running Intel MPI and uDAPL IB UD cm, set the following (recommended): setenv I_MPI_DAPL_PROVIDER=ofa-v2-mlx4_0-1u If running Intel MPI and uDAPL IB mcm with MIC, set the following: setenv I_MPI_DAPL_PROVIDER=ofa-v2-mlx4_0-1m If running Intel MPI and uDAPL socket cm, set the following: setenv I_MPI_DAPL_PROVIDER=ofa-v2-mlx4_0-1 If running Intel MPI and uDAPL rdma_cm, set the following: setenv I_MPI_DAPL_PROVIDER=ofa-v2-ib0 ============================================================ 5.0 MCM Provider, MPXYD Daemon (CCL-proxy) Build and Install ============================================================ MCM is a new uDAPL provider that is an extension to standard DAT 2.0 libraries. The purpose of this service is to proxy RDMA writes from the MIC to the HOST to improve large IO performance. The provider will support MIC to MIC, HOST to HOST, and MIC to HOST environments. The mcm client will NOT use MPXYD when running on the host. It requires a new MPXYD daemon service when clients are running on a MIC KNC adapter. This package installs all the host side libraries and daemon service. The MIC libraries must be built and moved over to MIC adapter. This verion is currently included with MPSS and all libraries and services will be installed by default. Current release package: dapl-2.1.9.tar.gz * Sample host build from source package (ofed must be installed) ./configure --enable-mcm --prefix=/usr --libdir=/usr/lib64 --sysconfdir=/etc make make install * Sample host rpmbuild/update from release tarball, /root: rpmbuild -ta dapl-2.1.9.tar.gz rpm -U /root/rpmbuild/RPMS/x86_64/dapl* * Sample MIC build from source package for MPSS 3.x KNC (MPSS must be installed) * Assume /opt is nfs mounted across cluster source /opt/mpss/3.x/environment-setup-k1om-mpss-linux ./configure --enable-mcm --prefix /opt/dapl/mic --host=x86_64-k1om-linux make make install copy /opt/dapl/mic/* files out to all MIC cards * Cluster deployment (1) Build once on the head or on one of the nodes (with MPSS) as described in the above steps. (2) HOST: Install dapl libraries and mpxyd service, "rpm -U" all dapl RPM files on host nodes: (3) MIC: Setup dapl overlay for new package (/opt/intel/dapl): Create /etc/mpss/conf.d/dapl.conf with following entry: Overlay Filelist /opt/dapl /opt/dapl/dapl.filelist on Create /opt/dapl/dapl.filelist with following entries: file /etc/dat.conf mic/etc/dat.conf 755 0 0 file /usr/bin/dtest mic/bin/dtest 755 0 0 file /usr/bin/dtestx mic/bin/dtestx 755 0 0 file /usr/bin/dtestcm mic/bin/dtestcm 755 0 0 file /usr/bin/dapltest mic/bin/dapltest 755 0 0 file /usr/lib64/libdat.so.2.0.0 mic/lib/libdat.so.2.0.0 755 0 0 file /usr/lib64/libdaplofa.so.2.0.0 mic/lib/libdaplofa.so.2.0.0 755 0 0 file /usr/lib64/libdaplomcm.so.2.0.0 mic/lib/libdaplomcm.so.2.0.0 755 0 0 file /usr/lib64/libdaploscm.so.2.0.0 mic/lib/libdaploscm.so.2.0.0 755 0 0 file /usr/lib64/libdaploucm.so.2.0.0 mic/lib/libdaploucm.so.2.0.0 755 0 0 slink /usr/lib64/libdat.so libdat.so.2.0.0 777 0 0 slink /usr/lib64/libdat.so.2 libdat.so.2.0.0 777 0 0 slink /usr/lib64/libdaplofa.so libdaplofa.so.2.0.0 777 0 0 slink /usr/lib64/libdaplofa.so.2 libdaplofa.so.2.0.0 777 0 0 slink /usr/lib64/libdaplomcm.so libdaplomcm.so.2.0.0 777 0 0 slink /usr/lib64/libdaplomcm.so.2 libdaplomcm.so.2.0.0 777 0 0 slink /usr/lib64/libdaploscm.so libdaploscm.so.2.0.0 777 0 0 slink /usr/lib64/libdaploscm.so.2 libdaploscm.so.2.0.0 777 0 0 slink /usr/lib64/libdaploucm.so libdaploucm.so.2.0.0 777 0 0 slink /usr/lib64/libdaploucm.so.2 libdaploucm.so.2.0.0 777 0 0 Reboot or restart MPSS and ofed-mic services Check for dapl overlay micctrl --config * Setup for non-root CCL Proxy testing, MPXYD running as process with different service port from your /home directory: Using build instructions above, change prefix as follow and "make install": Build MIC: --prefix=/home/username/ccl-proxy-mic Build host: --prefix=/home/username/ccl-proxy-host edit /home/username/ccl-proxy-host/etc/mpxyd.conf and change the following entries: log_file /var/log/mpxyd.log to log_file /tmp/username/mpxyd.log lock_file /var/log/mpxyd.pid to lock_file /tmp/username/mpxyd.log scif_port_id 68 to scif_port_id 1068 start the mpxyd process on each node ssh node1-hostname /home/username/ccl-proxy-host/sbin/mpxyd -P -O /home/username/ccl-proxy-host/etc/mpxyd.conf& Note: override default port id using following environment variable: export DAPL_MCM_PORT_ID=1068 * Notes (1) Modify "/etc/mpxyd.conf" to change the settings for the proxy. Especially, try different values of "buffer_segment_size" for performance tuning. Use a smaller value for "buffer_pool_mb" to reduce the memory foorprint of mpxyd. Use a larger value for "scif_listen_qlen" to run more MPI ranks per card. Also modify mcm_affinity_base to the desired CPU_id to insure socket to adapter affinity. Best performance when HCA, MIC, and CPU are on same socket. Default settings are on CPU socket 0. (2) By default, only writes originated from MIC is proxied. However, it is also possible to proxy host-originated writes (e.g. for debugging purpose). To do this, set the environment variable "DAPL_MCM_ALWAYS_PROXY=1". This variable applies to the provider, not the proxy. (3) Use the MCM provider with Intel MPI 5.1 or greater for best out of box experiences with MIC. Recommended settings: export I_MPI_MIC=1 export I_MPI_DEBUG=2 export I_MPI_FALLBACK=0 (4) Server Platform Notes: For optimal CCL Proxy performance, installed on Intel® Xeon® Processor E5-26xx v2 and v3 platforms, please update to the latest platform BIOS and set the following Processor, Power, and Memory settings. https://downloadcenter.intel.com/default.aspx [BIOS::Advanced::Processor Configuration] Intel(R) QPI Frequency Select=Auto Max Intel(R) Turbo Boost Technology=Enabled Enhanced Intel SpeedStep(R) Tech=Enabled Processor C3=Disabled Processor C6=Enabled Intel(R) Virtualization Technology=Disabled Intel(R) VT for Directed I/O=Disabled Direct Cache Access (DCA)=Enabled Extended ATR=0x01 [BIOS::Advanced::Power & Performance] CPU Power and Performance Policy=Performance [BIOS::Advanced::Memory Configuration::Memory RAS and Performance Configuration] Select Memory RAS Configuration=Maximum Performance NUMA Optimized=Enabled ==================================================================== 6.0 Mellanox Fabric Collective Accelerator (FCA) build for Intel MPI ==================================================================== 1) Download latest package: http://www.openfabrics.org/downloads/dapl/dapl-2.x.x.tar.gz 2) Build/Install with collectives configured for fca type: ./configure --enable-coll-type=fca --disable-mcm --prefix=/usr LDFLAGS=-L/opt/mellanox/fca/lib CPPFLAGS=-I/opt/mellanox/fca/include make install If an uDAPL provider with collective support is installed as an additional package, while another uDAPL provider also exists in a system, LD_LIBRARY_PATH environment variable should be updated to include the path to the newly installed DAPL package. 3) Set following MPI variables: export FCA_MGR_HOME=/opt/mellanox/fca export FCA_HOME=/opt/mellanox/fca export I_MPI_DAPL_COLLECTIVES=all export I_MPI_FABRICS=dapl export I_MPI_FALLBACK_DEVICE=off ============================= 7.0 Environment Variables ============================= - IB UD options using UCM provider, large scale settings (Xeon) export DAPL_NETWORK_NODES= /* set to active nodes on network for CM */ export DAPL_NETWORK_PPN= /* set to active processes per node for CM */ /* The following will be adjusted by provider based on NODES, PPN */ export DAPL_UCM_REP_TIME=8000 /* REQUEST timer, waiting on REPLY, msecs, default = 800 */ export DAPL_UCM_RTU_TIME=8000 /* REPLY timer, waiting for RTU in msecs, default=400 */ export DAPL_UCM_RETRY=7 /* REQUEST & REPLY retries, default = 7 */ export DAPL_UCM_QP_SIZE=4000 /* CM req/reply work queue size, default = 500 entries */ export DAPL_UCM_CQ_SIZE=4000 /* CM req/reply completion queue size, default = 500 entries */ export DAPL_UCM_TX_BURST=100 /* CM signal rate on send messages */ export DAPL_UCM_ENTRY_BITS=11 /* default = 11-bit, 2KB entries, allocation blocks */; export DAPL_UCM_ARRAY_BITS=18 /* default = 18 bit, 256KB total */ - IB RC options using SCM provider export DAPL_SCM_NETDEV=ib0 /* default is first non-loopback netdev */ - Other IB settings for all providers: export DAPL_MAX_INLINE=64 /* IB RC inline optimization, best small msg latency, def=64 */ export DAPL_ACK_RETRY=7 /* IB RC Ack retry count, default 7 */ export DAPL_ACK_TIMER=20 /* IB RC Ack retry timer, 5 bits, 4.096us*2^ack_timer. 16== 268ms, 20==4.2s */ export DAPL_IB_MTU=2048 /* IB MTU size, default = 2048 */ export DAPL_RNR_TIMER=12 /* 5 bits, 12 =.64ms, 28 =163ms, 31 =491ms */ export DAPL_RNR_RETRY=7 /* 3 bits, 7 == infinite */ export DAPL_IB_PKEY= 0 /* override IB partition key, default is pkey index 0 */ export DAPL_IB_SL=0 /* override IB Sevice level, default = 0 */ - Other options: export DAPL_WR_MAX=500 /* used to reduce max qp depth on all IB providers, default = dev attributes */ Debug logging and Counter settings ( --enable-counters) export DAPL_DBG_SYS_MEM=10 /* threshold for low sys memory warning, def = 10 percent */ export DAPL_DBG_TYPE=0x0000003 /* set log, monitor, and error checking, default = warnings and errors */ DAPL_DBG_TYPE bit settings as follow: DAPL_DBG_TYPE_ERR = 0x0001, DAPL_DBG_TYPE_WARN = 0x0002, DAPL_DBG_TYPE_EVD = 0x0004, DAPL_DBG_TYPE_CM = 0x0008, DAPL_DBG_TYPE_EP = 0x0010, DAPL_DBG_TYPE_UTIL = 0x0020, DAPL_DBG_TYPE_CALLBACK = 0x0040, DAPL_DBG_TYPE_DTO_COMP_ERR = 0x0080, DAPL_DBG_TYPE_API = 0x0100, DAPL_DBG_TYPE_RTN = 0x0200, DAPL_DBG_TYPE_EXCEPTION = 0x0400, DAPL_DBG_TYPE_SRQ = 0x0800, DAPL_DBG_TYPE_CNTR = 0x1000, DAPL_DBG_TYPE_CM_LIST = 0x2000, DAPL_DBG_TYPE_THREAD = 0x4000, DAPL_DBG_TYPE_CM_EST = 0x8000, DAPL_DBG_TYPE_CM_WARN = 0x10000, DAPL_DBG_TYPE_EXTENSION = 0x20000, DAPL_DBG_TYPE_CM_STATS = 0x40000, DAPL_DBG_TYPE_CM_ERRS = 0x80000, /* print any cm errors on device close */ DAPL_DBG_TYPE_LINK_ERRS = 0x100000, /* print any link errors on device close */ DAPL_DBG_TYPE_LINK_WARN = 0x200000, /* print any link warning on device close */ DAPL_DBG_TYPE_DIAG_ERRS = 0x400000, /* print any diag_counter errors on dev close */ DAPL_DBG_TYPE_SYS_WARN = 0x800000, /* print low mem warning during alloc, reg_mem */ DAPL_DBG_TYPE_VER = 0x1000000, /* print dapl ver and build date during dev open */ ============================= 9.0 SAMPLE uDAPL APPLICATION: ============================= There are 2 sample programs, with manpages, provided with this package. (dapl/test/dtest/) NAME dtest - simple uDAPL send/receive and RDMA test SYNOPSIS dtest [-P provider] [-b buf size] [-B burst count][-v] [-c] [-p] [-d] [-s] dtest [-P provider] [-b buf size] [-B burst count][-v] [-c] [-p] [-d] [-h HOSTNAME] DESCRIPTION dtest is a simple test used to exercise and verify the uDAPL interfaces. At least two instantia- tions of the test must be run. One acts as the server and the other the client. The server side of the test, once invoked listens for connection requests, until timing out or killed. Upon receipt of a cd connection request, the connection is established, the server and client sides exchange information necessary to perform RDMA writes and reads. OPTIONS -P=PROVIDER use PROVIDER to specify uDAPL interface using /etc/dat.conf (default OpenIB-cma) -b=BUFFER_SIZE use buffer size BUFFER_SIZE for RDMA(default 64) -B=BURST_COUNT use busrt count BURST_COUNT for interations (default 10) -v, verbose output(default off) -c, use consumer notification events (default off) -p, use polling (default wait for event) -d, delay in seconds before close (default off) -s, run as server (default - run as server) -h=HOSTNAME use HOSTNAME to specify server hostname or IP address (default - none) EXAMPLES dtest -P OpenIB-cma -v -s Starts a server process with debug verbosity using provider OpenIB-cma. dtest -P OpenIB-cma -h server1-ib0 Starts a client process, using OpenIB-cma provider to connect to hostname server1-ib0. SEE ALSO dapltest(1) AUTHORS Arlin Davis BUGS /dapl/test/dapltest/ NAME dapltest - test for the Direct Access Programming Library (DAPL) DESCRIPTION Dapltest is a set of tests developed to exercise, characterize, and verify the DAPL interfaces during development and porting. At least two instantiations of the test must be run. One acts as the server, fielding requests and spawning server-side test threads as needed. Other client invo- cations connect to the server and issue test requests. The server side of the test, once invoked, listens continuously for client connection requests, until quit or killed. Upon receipt of a con- nection request, the connection is established, the server and client sides swap version numbers to verify that they are able to communicate, and the client sends the test request to the server. If the version numbers match, and the test request is well-formed, the server spawns the threads needed to run the test before awaiting further connections. USAGE dapltest [ -f script_file_name ] [ -T S|Q|T|P|L ] [ -D device_name ] [ -d ] [ -R HT|LL|EC|PM|BE ] With no arguments, dapltest runs as a server using default values, and loops accepting requests from clients. The -f option allows all arguments to be placed in a file, to ease test automation. The following arguments are common to all tests: [ -T S|Q|T|P|L ] Test function to be performed: S - server loop Q - quit, client requests that server wait for any outstanding tests to complete, then clean up and exit T - transaction test, transfers data between client and server P - performance test, times DTO operations L - limit test, exhausts various resources, runs in client w/o server interaction Default: S [ -D device_name ] Specifies the interface adapter name as documented in the /etc/dat.conf static configura- tion file. This name corresponds to the provider library to open. Default: none [ -d ] Enables extra debug verbosity, primarily tracing of the various DAPL operations as they progress. Repeating this parameter increases debug spew. Errors encountered result in the test spewing some explanatory text and stopping; this flag provides more detail about what lead up to the error. Default: zero [ -R BE ] Indicate the quality of service (QoS) desired. Choices are: HT - high throughput LL - low latency EC - economy (neither HT nor LL) PM - premium BE - best effort Default: BE Usage - Quit test client dapltest [Common_Args] [ -s server_name ] Quit testing (-T Q) connects to the server to ask it to clean up and exit (after it waits for any outstanding test runs to complete). In addition to being more polite than simply killing the server, this test exercises the DAPL object teardown code paths. There is only one argument other than those supported by all tests: -s server_name Specifies the name of the server interface. No default. Usage - Transaction test client dapltest [Common_Args] [ -s server_name ] [ -t threads ] [ -w endpoints ] [ -i iterations ] [ -Q ] [ -V ] [ -P ] OPclient OPserver [ op3, Transaction testing (-T T) transfers a variable amount of data between client and server. The data transfer can be described as a sequence of individual operations; that entire sequence is transferred ’iterations’ times by each thread over all of its endpoint(s). The following parameters determine the behavior of the transaction test: -s server_name Specifies the name or IP address of the server interface. No default. [ -t threads ] Specify the number of threads to be used. Default: 1 [ -w endpoints ] Specify the number of connected endpoints per thread. Default: 1 [ -i iterations ] Specify the number of times the entire sequence of data transfers will be made over each endpoint. Default: 1000 [ -Q ] Funnel completion events into a CNO. Default: use EVDs [ -V ] Validate the data being transferred. Default: ignore the data [ -P ] Turn on DTO completion polling Default: off OP1 OP2 [ OP3, ... ] A single transaction (OPx) consists of: server|client Indicates who initiates the data transfer. SR|RR|RW Indicates the type of transfer: SR send/recv RR RDMA read RW RDMA write Defaults: none [ seg_size [ num_segs ] ] : Indicates the amount and format of the data to be transferred. Default: 4096 1 (i.e., 1 4KB buffer) [ -f ] For SR transfers only, indicates that a client’s send transfer completion should be reaped when the next recv completion is reaped. Sends and receives must be paired (one client, one server, and in that order) for this option to be used. Restrictions: Due to the flow control algorithm used by the transaction test, there must be at least one SR OP for both the client and the server. Requesting data validation (-V) causes the test to automatically append three OPs to those specified. These additional operations provide synchronization points during each iteration, at which all user-specified transaction buffers are checked. These three appended operations satisfy the "one SR in each direction" requirement. The transaction OP list is printed out if -d is supplied. Usage - Performance test client dapltest [Common_Args] -s server_name [ -m p|b ] [ -i iterations ] [ -p pipeline ] OP Performance testing (-T P) times the transfer of an operation. The operation is posted ’iterations’ times. The following parameters determine the behavior of the transaction test: -s server_name Specifies the name or IP address of the server interface. No default. -m b|p Used to choose either blocking (b) or polling (p) Default: blocking (b) [ -i iterations ] Specify the number of times the entire sequence of data transfers will be made over each endpoint. Default: 1000 [ -p pipeline ] Specify the pipline length, valid arguments are in the range [0,MAX_SEND_DTOS]. If a value greater than MAX_SEND_DTOS is requested the value will be adjusted down to MAX_SEND_DTOS. Default: MAX_SEND_DTOS OP Specifies the operation as follow: RR|RW Indicates the type of transfer: RR RDMA read RW RDMA write Defaults: none [ seg_size [ num_segs ] ] Indicates the amount and format of the data to be transferred. Default: 4096 1 (i.e., 1 4KB buffer) Usage - Limit test client Limit testing (-T L) neither requires nor connects to any server instance. The client runs one or more tests which attempt to exhaust various resources to determine DAPL limits and exercise DAPL error paths. If no arguments are given, all tests are run. Limit testing creates the sequence of DAT objects needed to move data back and forth, attempting to find the limits supported for the DAPL object requested. For example, if the LMR creation limit is being examined, the test will create a set of {IA, PZ, CNO, EVD, EP} before trying to run dat_lmr_create() to failure using that set of DAPL objects. The ’width’ parameter can be used to control how many of these parallel DAPL object sets are created before beating upon the requested constructor. Use of -m limits the number of dat_*_create() calls that will be attempted, which can be helpful if the DAPL in use supports essentailly unlimited numbers of some objects. The limit test arguments are: [ -m maximum ] Specify the maximum number of dapl_*_create() attempts. Default: run to object creation failure [ -w width ] Specify the number of DAPL object sets to create while initializing. Default: 1 [ limit_ia ] Attempt to exhaust dat_ia_open() [ limit_pz ] Attempt to exhaust dat_pz_create() [ limit_cno ] Attempt to exhaust dat_cno_create() [ limit_evd ] Attempt to exhaust dat_evd_create() [ limit_ep ] Attempt to exhaust dat_ep_create() [ limit_rsp ] Attempt to exhaust dat_rsp_create() [ limit_psp ] Attempt to exhaust dat_psp_create() [ limit_lmr ] Attempt to exhaust dat_lmr_create(4KB) [ limit_rpost ] Attempt to exhaust dat_ep_post_recv(4KB) [ limit_size_lmr ] Probe maximum size dat_lmr_create() Default: run all tests EXAMPLES dapltest -T S -d -D OpenIB-cma Starts a server process with debug verbosity. dapltest -T T -d -s host1-ib0 -D OpenIB-cma -i 100 client SR 4096 2 server SR 4096 2 Runs a transaction test, with both sides sending one buffer with two 4KB segments, one hundred times. dapltest -T P -d -s host1-ib0 -D OpenIB-cma -i 100 SR 4096 2 Runs a performance test, with the client sending one buffer with two 4KB segments, one hundred times. dapltest -T Q -s host1-ib0 -D OpenIB-cma Asks the server to clean up and exit. dapltest -T L -D OpenIB-cma -d -w 16 -m 1000 Runs all of the limit tests, setting up 16 complete sets of DAPL objects, and creating at most a thousand instances when trying to exhaust resources. dapltest -T T -V -d -t 2 -w 4 -i 55555 -s linux3 -D OpenIB-cma client RW 4096 1 server RW 2048 4 client SR 1024 4 server SR 4096 2 client SR 1024 3 -f server SR 2048 1 -f Runs a more complicated transaction test, with two thread using four EPs each, sending a more complicated buffer pattern for a larger number of iterations, validating the data received. ============================= 9.0 Summary of Fixes/Changes: ============================= Release 2.1.7 (OFED 3.18-1 GA) dtest: add -a -i options, all data sizes, incremental size dapl: Fix segfault while freeing qp mpxyd: add P2P inline support for data size <= 96 bytes dtest: change rdma_write_ping_pong so client is always last receiver ucm: add DAPL_NETWORK_PROCESS_NUM option for total ranks ucm: fca create group incorrectly using IB addr instead of socket address. ucm: fca_comm_destroy called with NULL dtest: add -W option for rdma write pinpong, similiar to ib_write_lat docs: update release notes for collective build mpxyd: reduce log level for rcv message flush dapltest: dapltest with no argument not working in ppc64 arch Release 2.1.6 (OFED 3.18-1) ucm: add cluster size environments to adjust CM timers mpxyd: proxy_in data transfers can improperly start before RTU received mcm: forward open/query for MFO devices in query only mode mpxyd: byte swap incorrect on WRC wr_len dtest: remove ERR message from flush QP function dapltest: Quit command with "-n port" number will core dump config: update dat.conf for MFO qib devices, 2 adapters/ports mpxyd: add MFO support on proxy side mcm: add MFO proxy commands, device, and CM support mcm: add MFO support to openib_common code base mcm: add full offload (MFO) mode to provider to support qib on MIC dtest: pre-allocated buffer too small for RMR, DTO ops timeout mpxyd: fix buffer initialization when no-inline support is active mpxyd: reduce log level on qp_flush to CM level mcm: intra-node proxy missing LID setup on rejects mcm: add intra-node support via ibscif device and mcm provider mcm: provide MIC address info with proxy device open mcm: add device info to non-debug log common: add DAPL_DTO_TYPE_EXTENSION_IMM for rdma_write_imm DTO type checking mpxyd: fix up some of the PI logging dtest: modify rdma_write_with_msg to support uni-direction streaming mcm,mpxyd: fix dreq processing to defer QP flush when proxy WRs still pending mpxyd: update byte_len and comp_cnt for PO to remote HST communications mcm: bug fixes for non-inline devices mcm: return CM_rej with CM_req_in errors mpxyd,mcm: RDMA write with immed data not signaled on request side mcm: add WC opcode and wc_flags in debug log message mpxyd: set options bug fix for mcm_ib_inline Update release notes with latest CM times Release 2.1.5 (OFED 3.18 RC3) update release notes, readme dat.conf: update comments regarding versions dtest: add logging of provider private data size with -v scm: remove use of msg.resv field for process id logging cma: report correct CM req private data size on query mpxyd: memset ib_wr structure before post_send on WC and WR requests mcm: add HST side provider support for device without inline data capability ucm: CM changes for UD extended port space and indexer ucm: add device support for new port space hash table ucm: allocate/free AH hash table for UD endpoint types ucm: check for AH caching when destroying via UD extension ucm: optimizations for large scale UD communication management mpxyd: use wr opcode instead of wc opcode to support logging on error cases mcm: HST->MXS mode, using RDMA_WRITE_WITH_IMM, fails with dtest -w dapl: aarch64 support for linux dapltest: add scripts to dist, set default device to IPoIB mpxyd: add wc_flags to proxy work completions Release 2.1.4 (OFED 3.18 RC1) mpxyd: fix typo in configuration file cma: RR attributes moved to common ib_cm struct mpxyd: tx thread incorrectly sleeps with negative pi_rw_cnt value dat.conf: add entries for True Scale qib device mpxyd: add support for devices without inline data support ucm: long disconnect times with many-to-one applications openib: add inline data support check during device open cleanup ib/cm attribute management across openib providers dapltest: fix -Werror=format-security issue with printf Release 2.1.3 (targeting OFED 3.18) dapl: mpxyd service changes to support multi-thread single-core option dapl: add rdma_write_imm and write only option to dtest ucm: add time wait override capability for CM services common: dapl_ep_free must serialize CM object destroy dtestx: allow scale up to 1000 EP's ucm: RTU not retransmitted in TIMEWAIT state mpxyd: increase max open files for service mpxyd: DTO completion ERR: status 12, op RDMA_WRITE running MPI alltoall test mcm: HST->MXS mode incorrectly signals multiple fragments per WR mcm: add segmentation to HST->MXS mode for improved performance mpxyd: set global seg_sz to 128KB for proxy data service openib: add port_num to provider named attributes mcm: provide CPU family/model attribute on both host and mic sides dtestx: update IB extension example test with new v2.0.9 features dtest: add dtestsrq for SRQ example and provider testing common: add srq support for openib verbs providers openib: add IB UD cm_free/ah_free extension support in UCM provider openib: add new TIMEWAIT state for CM extension: add IB UD extensions to reduce provider CM and AH memory footprint mpxyd/mcm: add provider specific attribute DAT_IB_PROXY_VERSION mpxyd: log warning if running in COMPAT mode add provider and proxy support for GUID across platform common: return appropriate handles with affiliated EP and EVD async events Release 2.1.2 (OFED 3.12-1) mpxyd: add global routing support for proxy connections mcm: only call mix_get_attr if running on MIC openib: modify check for link_layer to handle unspecified dapl: add support for the s390x platform dtest server exchange connection info with client mpxyd: 2 MICs in same numa_node will overlap CPU affinity, don't reset base mcm: implement proxy mix_prov_attr function, add fields CPU model and family mpxyd: tx thread may not be signaled on small segment writes Release 2.1.1 (OFED 3.12-1 RC1) common: add provider name to log messages mpxyd: log warning message if numa_node invalid include debuginfo with build build: include debuginfo with build mpxyd: tx thread doesn't sleep during no pending IO state mpxyd: change MIC cpu_mask to per numa node instead of adapter mpxyd: set to MXS mode if device numa_node is invalid (-1) mpxyd: MXS based alltoall benchmark hangs or returns post_send timeout mpxyd: add IO profile capabilities to help debug alltoall stall cases mpxyd: retry stalled inline post_send, init m_idx only when signaled Release 2.1.0 (OFED 3.12-1, MIC support added) build: add missing NEWS file update autogen.sh add MCM provider and MPXYD service to build mpxyd: service startup script and configuration file add readme for MCM provider and MPXYD service update Copyright dates add new MIC RDMA proxy service daemon (MPXYD) add new dapl MIC provider (MCM) to support MIC RDMA proxy services MCM: new MIC provider and proxy service definitions cleanup build warnings common: add CQ,QP,MR abstractions for new MIC provider and data proxy service openib: cleanup, use inet_ntop for GIDs, remove some logs, destroy pipes on release common: new dapls_evd_cqe_to_event call, cqe to event common: init ring_buffer, assign hd/tl pos in range allow log level changes during device open ucm: fix cm rbuf setup, include grh pad on initialization ucm: remove duplicate async_event code, use common async event call new lightweight open_query/close_query IB extension for fast attribute query dtestcm: add more detailed debug during disconnect phase cma: long delays when opening cma provider with no IPoIB configured common: new debug levels for low system memory, IA stats, and package info build: remove library check for mverbs with --enable-fca IB extension: segfault in create collective group with non-vector type IA handle" build: change configure help to correctly state collective default=none Release 2.0.42 fixes (OFED 3.12 GA) dapltest: increase DTO evd size to prevent CQ overflow on limit_rpost test dapltest: RSP limit test fails. Creation of reserved SP moves EP state to DAT_EP_STATE_RESERVED in error cases. dapl: fix string bug in dapls_dto_op_str Release 2.0.41 fixes (OFED 3.12 RC1) dapltest: change server port, from 45278 to 62000, out of registered IANA range dat: lower log level on load errors of provider library dat: dat_ia_open should close provider after failure dapltest: set default limit max to 1000 openib: add new provider specific attributes dapltest: update scripts for regression testing purposes dapltest: Add final send/recv "sync" for transaction tests. Release 2.0.40 fixes (OFED 3.12) dist: ib collective extension include files missing dapltest: the quit command is missing changes for -n option dat.conf: remove v1, add Mellanox Connect-IB and Intel Xeon Phi MIC NULL undefined on Fedora, incorrectly using kernel stddef.h Release 2.0.39 fixes (OFED 3.5-2 GA) dapltest: fix endian swap issue with performance test scm: getifaddrs modfications for better out of the box experience ucm, scm: UD mode triggers list_head assert with large scale alltoall test Release 2.0.38 dapltest: add -n parameter to override default server port number (45278) ucm,scm: UD mode creates many CR objects per EP that needs cleaned up cma: add DAPL_CM_TOS environment variable to enable passing a TOS to the RDMA CM Release 2.0.37 common: add support for ia name during dat_ia_query common: dapl_os_atomic_inc/dec() not working as expected on ppc64 machines. dapltest: ppc64 endian issue with exchanged mem handle and address Release 2.0.36 scm: increase ACK timeout to 20 for a default value to match other providers. common: allow qp modify in init state common: check for valid states during ep posting dat.conf: keep list of providers in order for backward compatibility ucm: record and silently drop a duplicate reject CM message windows: new version of getlocalipaddr not portable dapltest: DFLT_QLEN is defined in multiple tests Release 2.0.35 config/build: remove post/postun hacking used to modify dat.conf config: clean up help option displays with ext-type options windows: Provide auto-detect between RoCE and Infiniband for Windows. ucm: update UD cm provider to support new CM stat and error counters scm: update socket cm provider to support new CM stat and error counters commom: add cm, link, and diag event counters in IB extended builds scm: use ioctl SIOCIFCONF to get complete list of configured netdev interfaces ucm: UD send failures at scale, ucm_send ERR: get_smsg(hd=149,tl=150) scm: fix retry count on connection pending timeout ucm: cleanup debug message, ntohl on p_size is incorrect cma, scm, ucm: allow EP (QP) creation without EVD (CQ) common: add DAPL_DBG_TYPE_CM_STATS (0x40000) to debug log options common: dapls_ep_flush_cq will segfault when no CQ is attached to EP common: ep_create should allow max_request_iov attribute setting of zero common: add check for NULL handle on ext calls, SRQ free, and helper functions common: add missing sub-types to dat_strerror() common: extended CR event processing missing rejects on errors ucm: incorrectly sends user reject during CR callback errors common: change dbg level on CR callback if not listening on SP scm: incorrectly sends user reject during CR callback errors dat: add check for NULL handle on IA calls cma,scm,ucm: extra reference on EP, with RSP, causes dat_ep_free() to hang common: RSP service points incorrectly freed during CR callback common: clean up dat_rsp_create log message common: cleanup debug message on EVD overflows scm: return correct event error code when remote host refuses requests dapltest: server CR EVD is too small for multi-client configurations. Common: CR EVD overflow causes segfault. Release 2.0.34 scm: change debug message level for listen/bind errors common: increase default IB ack timer from 16 to 20 common: remote ia address null pointer creates seg fault common: posting events on full queue returns wrong error code common: dat_ep_modify seg faults with null ep_param ptr common: dat_evd_free seg faults with resized software EVD common: remove assert for incorrect events during cm_request dat: dat_cno_query with NULL cno_handle causes segmentation fault scm: dat_psp_create returns wrong error code on bind/listen failure scm: socket connect request count is reset improperly on retry scm: when hostname has loopback addr assigned, default to eth0 instead of failing scm: add port number to error log during hca_open failures common: query calls return incorrect IA handle to consumer common: srq create asserts with !dapl_llist_is_empty(head) failed Release 2.0.33 scm,ucm: fix compatibility issues and set minimum protocol support build: link librdmacm dependency to ib_acm usage for ucm and scm providers build: add selective enable/disable-xxx build switch for each provider build: add extended header files to EXTRA_DIST and fix missing backslash build: set IB extended coll-type to none by default common: change errno mapping of EINVAL to DAT_INVALID_PARAMETER build: add IB collective and FCA provider to dapl build package as an option common: add new dapls_evd_post_event_ext call for extended events ucm: add support for IB collective providers scm: add support for IB collective providers cma: add support for IB collective providers common: add supported collective types in named attributes for query common: add collective call mappings via standard dapli_post_ext() common: new debug bitmask definition for extension logging common: new IB collective provider for Mellanox Fabric Collective Agent dat: add definitions for MPI offloaded collectives in IB transport extensions common: cleanup debug messages when building with ibacm feature Release 2.0.32 fixes (OFED 1.5.3 GA): cma: reduce output log level in disconnect from WARN to CM_WARN ucm: delay freeing of active side UD cm object in case RTU is dropped ucm: cm object needs to be on work queue before req sent on wire ucm,scm: remove use of usec_sleep delays and use events for disc and destroy common: reduce default max inline data size because of performance anomaly common: dapls_evd_dto_wait() dbg message should print status and not errno ucm, scm: exchange max_qp_rd_atom and limit outstanding requests scm: retry socket connect on ECONNREFUSED under heavy load common: qp modify RTR using wrong ep attribute parameter for dest_rd_atomic Release 2.0.31 fixes (OFED 1.5.3 RC1): common: clean up build warning for unused variable event_ptr scm, ucm: set RAI_NOROUTE flag with rdma_getaddrinfo() call to avoid blocking. cma: definition for dapl_sp_remove_ep() is missing in cm.c libdat: static provider entries created for local SR database not freed libdat: memory leak in static registration during parsing common: increase default IB inline send threshold to 400 common cq: a mixup of errno and the -1 return from poll in dapls_wait_comp_channel ucm: release UD cm objects after AH is exchanged to avoid duplicate request drops ucm: decrease timeout retry count for disconnect requests ucm: hold lock when sending cm_msgs to sync timer start with packet send ucm: add debugging to include process id for better scale up debug aids cma: disconnect can block for excessive times waiting for rdma_cm DREP timeout ucm: configure the recv channel FD to non-blocking windows: Missing librdmacm include path for build debug build: only timestamp if sending to stdout to avoid performance hit common: print out errors on free build and not just debug builds cma: fix debug build issue scm, ucm: MPI spawn test on oversubcribed server taking excessive time to complete common: add high resolution time stamps and thread id to sdtout debug logs common: modify debug in dat_evd_dequeue to reduce noise, only output on non-empty cma: rdma_destroy_id called twice during device open bind error common: dat_evd_dequeue (poll_cq) fails with invalid parameter after EP (qp) free ucm: allow configuration of CM burst (signal) threshold on posting cma: fix debug build windows: debug version of windows does not build. Allow DAPL out of band connection models to use ibacm to obtain path record data. ucm: add missing map file for UCM provider ibal: delay QP transition during disconnect phase Revert "ibal: delay QP transition during disconnect phase" ibal: delay QP transition during disconnect phase common: restructure EVD processing to handle EP destruction phase ibal: sync QP destruction and device close ucm: remove unnecessary debug warning in async callback