Updated: 25 Aug 2015 Introduction ============ Mellanox's Messaging Accelerator (VMA) is dynamically linked user space Linux library for transparently enhancing the performance of networking-heavy applications. It boosts performance for message-based and streaming applications such as those found in financial services market data environments and Web2.0 clusters. It allows application written over standard socket API to run over Infiniband and/or Ethernet from user-space with full network stack bypass. The result is a reduction in latency by as much as 300%, an increase in application throughput by as much as 200%, higher packets rates and better CPU utilization as compared to applications running on standard Ethernet or InfiniBand interconnect networks. Build libvma from source ======================== Prerequisites: 1. MLNX_OFED as described in the "Pre Installation" step of next section. 2. Or, upstream kernel and userspace verbs libraries (libibverbs, libmlx4, librdmacm) 3. Autoconf, Automake, libtool, unzip, patch, libnl-devel (netlink 1 or 3) Build: 1. ./autogen.sh 2. ./configure --with-ofed=/usr --prefix=/usr --libdir=/usr/lib64 --includedir=/usr/include/mellanox --docdir=/usr/share/doc/libvma --sysconfdir=/etc 3. make You will find libvma.so in path_to_vma_dir/src/vma/.libs/libvma.so. Install: 1. sudo make install Tip: ./install.sh can do the build and install steps for you. Install libvma from rpm or debian ================================= Pre Installation: 1. If possible, install latest MLNX_OFED with the --vma option. This will also install libvma, and you can skip to "Running" step. 2. If installing over existing MLNX_OFED, add the following to /etc/modprobe.d/mlnx.conf: options ib_uverbs disable_raw_qp_enforcement=1 options mlx4_core fast_drop=1 options mlx4_core log_num_mgm_entry_size=-1 And restart the driver: /etc/init.d/openibd restart Installing: Install the package as any other rpm or debian package [rpm -i libvma.X.Y.Z-R.rpm]. The installation copies the VMA library to: /usr/lib[64]/libvma.so The VMA monitoring utility is installed at: /usr/bin/vma_stat The VMA extra socket API is located at: /usr/include/mellanox/vma_extra.h The installation location of the README.txt and version information file (VMA_VERSION), are as follows: - Redhat: /usr/share/doc/libvma-X.Y.Z-R/ - SuSE: /usr/share/doc/packages/libvma-X.Y.Z-R/ Post Installation: When working over Infiniband, we recommend to manually add persistence for the following system parameters: 1. Force IPoIB to work in 'datagram' mode (disabling IPoIB 'connected' mode) Modify "SET_IPOIB_CM=no" in file "/etc/infiniband/openib.conf" 2. Force IGMP to work in V2 on IB interfaces Add "net.ipv4.conf.all.force_igmp_version = 2" in file "/etc/sysctl.conf" Upgrading: Use rpm update procedure: # rpm -U libvma.X.Y.Z-R.rpm You can upgrade by uninstalling (rpm -e) the previously installed package before starting to install the new VMA rpm. Uninstalling: When uninstalling remember to uninstall (rpm -e) the package before you uninstall ofed. Running: Set the env variable LD_PRELOAD to libvma.so and run your application. Example: # LD_PRELOAD=libvma.so iperf -uc 224.22.22.22 -t 5 Configuration Values ==================== On default startup the VMA library logs to stderr the VMA version, the modified configuration parameters being used and their values. Please notice that except VMA_TRACELEVEL, VMA logs just those parameters whose value != default. Example: VMA INFO: --------------------------------------------------------------------------- VMA INFO: VMA_VERSION: 7.0.11-0 VMA INFO: Cmd Line: sockperf sr VMA INFO: Current Time: Tue Dec 22 16:17:53 2015 VMA INFO: Pid: 5797 VMA INFO: OFED Version: MLNX_OFED_LINUX-3.1-1.5.5: VMA INFO: Architecture: x86_64 VMA INFO: Node: r-aa-bob06.mtr.labs.mlnx VMA INFO: --------------------------------------------------------------------------- VMA INFO: Log Level DETAILS [VMA_TRACELEVEL] VMA DETAILS: Log Details 0 [VMA_LOG_DETAILS] VMA DETAILS: Log Colors Enabled [VMA_LOG_COLORS] VMA DETAILS: Log File [VMA_LOG_FILE] VMA DETAILS: Stats File [VMA_STATS_FILE] VMA DETAILS: Stats shared memory directory /tmp/ [VMA_STATS_SHMEM_DIR] VMA DETAILS: Stats FD Num (max) 100 [VMA_STATS_FD_NUM] VMA DETAILS: Conf File /etc/libvma.conf [VMA_CONFIG_FILE] VMA DETAILS: Application ID VMA_DEFAULT_APPLICATION_ID [VMA_APPLICATION_ID] VMA DETAILS: Polling CPU idle usage Disabled [VMA_CPU_USAGE_STATS] VMA DETAILS: SigIntr Ctrl-C Handle Disabled [VMA_HANDLE_SIGINTR] VMA DETAILS: SegFault Backtrace Disabled [VMA_HANDLE_SIGSEGV] VMA DETAILS: Ring allocation logic TX 0 (Ring per interface) [VMA_RING_ALLOCATION_LOGIC_TX] VMA DETAILS: Ring allocation logic RX 0 (Ring per interface) [VMA_RING_ALLOCATION_LOGIC_RX] VMA DETAILS: Ring migration ratio TX 100 [VMA_RING_MIGRATION_RATIO_TX] VMA DETAILS: Ring migration ratio RX 100 [VMA_RING_MIGRATION_RATIO_RX] VMA DETAILS: Ring limit per interface 0 (no limit) [VMA_RING_LIMIT_PER_INTERFACE] VMA DETAILS: TCP max syn rate 0 (no limit) [VMA_TCP_MAX_SYN_RATE] VMA DETAILS: Tx Mem Segs TCP 1000000 [VMA_TX_SEGS_TCP] VMA DETAILS: Tx Mem Bufs 200000 [VMA_TX_BUFS] VMA DETAILS: Tx QP WRE 3000 [VMA_TX_WRE] VMA DETAILS: Tx Max QP INLINE 220 [VMA_TX_MAX_INLINE] VMA DETAILS: Tx MC Loopback Enabled [VMA_TX_MC_LOOPBACK] VMA DETAILS: Tx non-blocked eagains Disabled [VMA_TX_NONBLOCKED_EAGAINS] VMA DETAILS: Tx Prefetch Bytes 256 [VMA_TX_PREFETCH_BYTES] VMA DETAILS: Rx Mem Bufs 200000 [VMA_RX_BUFS] VMA DETAILS: Rx QP WRE 16000 [VMA_RX_WRE] VMA DETAILS: Rx QP WRE BATCHING 64 [VMA_RX_WRE_BATCHING] VMA DETAILS: Rx Byte Min Limit 65536 [VMA_RX_BYTES_MIN] VMA DETAILS: Rx Poll Loops 100000 [VMA_RX_POLL] VMA DETAILS: Rx Poll Init Loops 0 [VMA_RX_POLL_INIT] VMA DETAILS: Rx UDP Poll OS Ratio 100 [VMA_RX_UDP_POLL_OS_RATIO] VMA DETAILS: Rx UDP HW TS Conversion 3 [VMA_RX_UDP_HW_TS_CONVERSION] VMA DETAILS: Rx Poll Yield Disabled [VMA_RX_POLL_YIELD] VMA DETAILS: Rx Prefetch Bytes 256 [VMA_RX_PREFETCH_BYTES] VMA DETAILS: Rx Prefetch Bytes Before Poll 0 [VMA_RX_PREFETCH_BYTES_BEFORE_POLL] VMA DETAILS: Rx CQ Drain Rate Disabled [VMA_RX_CQ_DRAIN_RATE_NSEC] VMA DETAILS: GRO max streams 32 [VMA_GRO_STREAMS_MAX] VMA DETAILS: TCP 3T rules Disabled [VMA_TCP_3T_RULES] VMA DETAILS: ETH MC L2 only rules Disabled [VMA_ETH_MC_L2_ONLY_RULES] VMA DETAILS: Select Poll (usec) 100000 [VMA_SELECT_POLL] VMA DETAILS: Select Poll OS Force Disabled [VMA_SELECT_POLL_OS_FORCE] VMA DETAILS: Select Poll OS Ratio 10 [VMA_SELECT_POLL_OS_RATIO] VMA DETAILS: Select Skip OS 4 [VMA_SELECT_SKIP_OS] VMA DETAILS: CQ Drain Interval (msec) 10 [VMA_PROGRESS_ENGINE_INTERVAL] VMA DETAILS: CQ Drain WCE (max) 10000 [VMA_PROGRESS_ENGINE_WCE_MAX] VMA DETAILS: CQ Interrupts Moderation Enabled [VMA_CQ_MODERATION_ENABLE] VMA DETAILS: CQ Moderation Count 48 [VMA_CQ_MODERATION_COUNT] VMA DETAILS: CQ Moderation Period (usec) 50 [VMA_CQ_MODERATION_PERIOD_USEC] VMA DETAILS: CQ AIM Max Count 560 [VMA_CQ_AIM_MAX_COUNT] VMA DETAILS: CQ AIM Max Period (usec) 250 [VMA_CQ_AIM_MAX_PERIOD_USEC] VMA DETAILS: CQ AIM Interval (msec) 250 [VMA_CQ_AIM_INTERVAL_MSEC] VMA DETAILS: CQ AIM Interrupts Rate (per sec) 5000 [VMA_CQ_AIM_INTERRUPTS_RATE_PER_SEC] VMA DETAILS: CQ Poll Batch (max) 16 [VMA_CQ_POLL_BATCH_MAX] VMA DETAILS: CQ Keeps QP Full Enabled [VMA_CQ_KEEP_QP_FULL] VMA DETAILS: QP Compensation Level 256 [VMA_QP_COMPENSATION_LEVEL] VMA DETAILS: Offloaded Sockets Enabled [VMA_OFFLOADED_SOCKETS] VMA DETAILS: Timer Resolution (msec) 10 [VMA_TIMER_RESOLUTION_MSEC] VMA DETAILS: TCP Timer Resolution (msec) 100 [VMA_TCP_TIMER_RESOLUTION_MSEC] VMA DETAILS: TCP control thread 0 (Disabled) [VMA_TCP_CTL_THREAD] VMA DETAILS: TCP timestamp option 0 [VMA_TCP_TIMESTAMP_OPTION] VMA DETAILS: Exception handling mode -1 (just log debug message) [VMA_EXCEPTION_HANDLING] VMA DETAILS: Avoid sys-calls on tcp fd Disabled [VMA_AVOID_SYS_CALLS_ON_TCP_FD] VMA DETAILS: Delay after join (msec) 0 [VMA_WAIT_AFTER_JOIN_MSEC] VMA DETAILS: Internal Thread Affinity -1 [VMA_INTERNAL_THREAD_AFFINITY] VMA DETAILS: Internal Thread Cpuset [VMA_INTERNAL_THREAD_CPUSET] VMA DETAILS: Internal Thread Arm CQ Disabled [VMA_INTERNAL_THREAD_ARM_CQ] VMA DETAILS: Internal Thread TCP Handling 0 (deferred) [VMA_INTERNAL_THREAD_TCP_TIMER_HANDLING] VMA DETAILS: Thread mode Multi spin lock [VMA_THREAD_MODE] VMA DETAILS: Buffer batching mode 1 (Batch and reclaim buffers) [VMA_BUFFER_BATCHING_MODE] VMA DETAILS: Mem Allocate type 1 (Contig Pages) [VMA_MEM_ALLOC_TYPE] VMA DETAILS: Num of UC ARPs 3 [VMA_NEIGH_UC_ARP_QUATA] VMA DETAILS: UC ARP delay (msec) 10000 [VMA_NEIGH_UC_ARP_DELAY_MSEC] VMA DETAILS: Num of neigh restart retries 1 [VMA_NEIGH_NUM_ERR_RETRIES] VMA DETAILS: IPOIB support Enabled [VMA_IPOIB] VMA DETAILS: BF (Blue Flame) Enabled [VMA_BF] VMA DETAILS: fork() support Enabled [VMA_FORK] VMA DETAILS: close on dup2() Enabled [VMA_CLOSE_ON_DUP2] VMA DETAILS: MTU 0 (follow actual MTU) [VMA_MTU] VMA DETAILS: MSS 0 (follow VMA_MTU) [VMA_MSS] VMA DETAILS: TCP CC Algorithm 0 (LWIP) [VMA_TCP_CC_ALGO] VMA DETAILS: Suppress IGMP ver. warning Disabled [VMA_SUPPRESS_IGMP_WARNING] VMA INFO: --------------------------------------------------------------------------- VMA_TRACELEVEL Logging level the VMA library will be using. Default is info Example: # VMA_TRACELEVEL=debug none Print no log at all panic Panic level logging, this would generally cause fatal behavior and an exception will be thrown by the VMA library. Typically, this is caused by memory allocation problems. This level is rarely used. error Runtime ERRORs in the VMA. Typically, these can provide insight for the developer of wrong internal logic like: Errors from underlaying OS or Infiniband verbs calls. internal double mapping/unmapping of objects. warn Runtime warning that do not disrupt the workflow of the application but might warn of a problem in the setup or the overall setup configuration. Typically, these can be address resolution failure (due to wrong routing setup configuration), corrupted ip packets in the receive path or unsupported functions requested by the user application info General information passed to the user of the application. Bring up configuration logging or some general info to help the user better use the VMA library details Complete VMA's configuration information. Very high level insight of some of the critical decisions done in VMA. debug High level insight to the operations done in the VMA. All socket API calls are logged and internal high level control channels log there activity. fine Low level run time logging of activity. This logging level includes basic Tx and Rx logging in the fast path and it will lower application performance. It is recommended to use this level with VMA_LOG_FILE param. finer Very low level run time logging of activity! This logging level will DRASTICALLY lower application performance. It is recommended to use this level with VMA_LOG_FILE param. all today this level is identical to finer VMA_LOG_DETAILS Add details on each log line. 0 = Basic log line 1 = ThreadId 2 = ProcessId + ThreadId 3 = Time + ProcessId + ThreadId [Time is in milli-seconds from start of process] Default value is 0 VMA_LOG_COLORS Use color scheme when logging. Red for errors, purple for warnings and dim for low level debugs. VMA_LOG_COLORS is automatically disabled when logging is direct to a non terminal device (e.g. VMA_LOG_FILE is configured). Default value is 1 (Enabled) VMA_LOG_FILE Redirect all VMA logging to a specific user defined file. This is very useful when raising the VMA_TRACELEVEL VMA will replace a single '%d' appearing in the log file name with the pid of the process loaded with VMA. This can help in running multiple instances of VMA each with it's own log file name. Example: VMA_LOG_FILE=/tmp/vma_log.txt VMA_STATS_FILE Redirect socket statistics to a specific user defined file. VMA will dump each socket's statistics into a file when closing the socket. Example: VMA_STATS_FILE=/tmp/stats VMA_STATS_SHMEM_DIR Set the directory path for vma to create the shared memory files for vma_stats. No files will be created when setting this value to empty string "". Default value is /tmp/ VMA_STATS_FD_NUM Max number of sockets monitored by VMA statistic mechanism. Value range is 0 to 1024. Default value is 100 VMA_CONFIG_FILE Sets the full path to the VMA configuration file. Default values is: /etc/libvma.conf Example: VMA_CONFIG_FILE=/tmp/libvma.conf VMA_APPLICATION_ID Specify a group of rules from libvma.conf for VMA to apply. Example: 'VMA_APPLICATION_ID=iperf_server'. Default is "VMA_DEFAULT_APPLICATION_ID" (match only the '*' group rule) VMA_CPU_USAGE_STATS Calculate VMA CPU usage during polling HW loops. This information is available through VMA stats utility. Default value is 0 (Disabled) VMA_HANDLE_SIGINTR When Enabled, VMA handler will be called when interrupt signal is sent to the process. VMA will also call to application's handler if exist. Value range is 0 to 1 Default value is 0 (Disabled) VMA_HANDLE_SIGSEGV When Enabled, print backtrace if segmentation fault happens. Value range is 0 to 1 Default value is 0 (Disabled) VMA_TX_SEGS_TCP Number of TCP LWIP segments allocation for each VMA process. Default value is 1000000 VMA_TX_BUFS Number of global Tx data buffer elements allocation. Default value is 200000 VMA_TX_WRE Number of Work Request Elements allocated in all transmit QP's. The number of QP's can change according to the number of network offloaded interfaces. Default value is 3000 VMA_TX_MAX_INLINE Max send inline data set for QP. Data copied into the INLINE space is at least 32 bytes of headers and the rest can be user datagram payload. VMA_TX_MAX_INLINE=0 disables INLINEing on the tx transmit path. In older releases this parameter was called: VMA_MAX_INLINE Default VMA_TX_MAX_INLINE is 220 VMA_TX_MC_LOOPBACK This parameter sets the initial value used by VMA internally to controls the multicast loopback packets behavior during transmission. An application that calls setsockopt() with IP_MULTICAST_LOOP will run over the initial value set by this parameter. Read more in 'Multicast loopback behavior' in notes section below Default value is 1 (Enabled) VMA_TX_NONBLOCKED_EAGAINS Return value 'OK' on all send operation done on a non-blocked udp sockets. This is the OS default behavior. The datagram sent is silently dropped inside VMA or the network stack. When set Enabled (set to 1), VMA will return with error EAGAIN if it was unable accomplish the send operation and the datagram was dropped. In both cases a dropped Tx statistical counter is incremented. Default value is 0 (Disabled) VMA_TX_PREFETCH_BYTES Accelerate offloaded send operation by optimizing cache. Different values give optimized send rate on different machines. We recommend you tune this for your specific hardware. Value range is 0 to MTU size Disable with a value of 0 Default value is 256 bytes VMA_RING_ALLOCATION_LOGIC_TX VMA_RING_ALLOCATION_LOGIC_RX Ring allocation logic is used to separate the traffic to different rings. By default all sockets use the same ring for both RX and TX over the same interface. Even when specifing the logic to be per socket or thread, for different interfaces we use different rings. The logic options are: 0 - Ring per interface 10 - Ring per socket (using socket fd as separator) 20 - Ring per thread (using the id of the thread in which the socket was created) 30 - Ring per core (using cpu id) 31 - Ring per core - attach threads : attach each thread to a cpu core Default value is 0 VMA_RING_MIGRATION_RATIO_TX VMA_RING_MIGRATION_RATIO_RX Ring migration ratio is used with the "ring per thread" logic in order to decide when it is beneficial to replace the socket's ring with the ring allocated for the current thread. Each VMA_RING_MIGRATION_RATIO iterations (of accessing the ring) we check the current thread ID and see if our ring is matching the current thread. If not, we consider ring migration. If we keep accessing the ring from the same thread for some iterations, we migrate the socket to this thread ring. Use a value of -1 in order to disable migration. Default value is 100 VMA_RING_LIMIT_PER_INTERFACE Limit the number of rings that can be allocated per interface. For example, in ring allocation per socket logic, if the number of sockets using the same interface is larger than the limit, then several sockets will be sharing the same ring. [Note:VMA_RX_BUFS might need to be adjusted in order to have enough buffers for all rings in the system. Each ring consume VMA_RX_WRE buffers.] Use a value of 0 for unlimited number of rings. Default value is 0 (no limit) VMA_RX_BUFS Number Rx data buffer elements allocation for the processes. These data buffers will be used by all QPs on all HCAs as determined by the VMA_QP_LOGIC. Default value is 200000 VMA_RX_WRE Number of Work Request Elements allocated in all receive QP's. The number of QP's can change according to the VMA_QP_LOGIC. Default value is 16000 VMA_RX_WRE_BATCHING Number of Work Request Elements and RX buffers to batch before recycling. Batching decrease latency mean, but might increase latency STD. Value range is 1-1024. Default value is 64 VMA_RX_BYTES_MIN Minimum value in bytes that will be used per socket by VMA when applications call to setsockopt(SO_RCVBUF). If application tries to set a smaller value then configured in VMA_RX_BYTES_MIN, VMA will force this minimum limit value on the socket.VMA offloaded socket's receive max limit of ready bytes count. If the application does not drain a sockets and the byte limit is reached, new received datagrams will be dropped. Monitor of the applications socket's usage of current, max and dropped bytes and packet counters can be done with vma_stats. Default value is 2000000 VMA_RX_POLL The number of times to poll on Rx path for ready packets before going to sleep (wait for interrupt in blocked mode) or return -1 (in non-blocked mode). This Rx polling is done when the application is working with direct blocked calls to read(), recv(), recvfrom() & recvmsg(). When Rx path has successfull poll hits (see performace monitoring) the latency is improved dramatically. This comes on account of CPU utilization. Value range is -1, 0 to 100,000,000 Where value of -1 is used for infinite polling Default value is 100000 VMA_RX_POLL_INIT VMA maps all UDP sockets as potential offloaded capable. Only after the ADD_MEMBERSHIP does the offload start to work and the CQ polling kicks in VMA. This parameter control the polling count during this transition phase where the socket is a UDP unicast socket and no multicast addresses where added to it. Once the first ADD_MEMBERSHIP is called the above VMA_RX_POLL takes effect. Value range is similar to the above VMA_RX_POLL Default value is 0 VMA_RX_UDP_POLL_OS_RATIO The above param will define the ratio between VMA CQ poll and OS FD poll. This will result in a signle poll of the not-offloaded sockets every VMA_RX_UDP_POLL_OS_RATIO offlaoded socket (CQ) polls. No matter if the CQ poll was a hit or miss. No matter if the socket is blocking or non-blocking. When disabled, only offlaoded sockets are polled. This parameter replaces the two old parameters: VMA_RX_POLL_OS_RATIO and VMA_RX_SKIP_OS Disable with 0 Default value is 100 VMA_RX_UDP_HW_TS_CONVERSION The above param will define the udp hardware receive time stamp conversion method. Experimental verbs is required for converting the time stamp from hardware time (Hz) to system time (seconds.nano_seconds). Hence, hardware support is not guaranteed. The value of VMA_RX_UDP_HW_TS_CONVERSION determined by all devices - i.e if the hardware of one device does not support the conversion, then it will be canceled for the other devices. Disable with 0 Raw HW time with 1 - only convert the time stamp to seconds.nano_seconds time units (or disable if hardware does not supports). Use best sync possible with 2 - Sync to system time, then Raw hardware time - disable if none of them are supported by hardware. Sync to system time with 3 - convert the time stamp to seconds.nano_seconds time units. comparable to UDP receive software timestamp. disable if hardware does not supports. Default value 3 VMA_RX_POLL_YIELD When an application is running with multiple threads, on a limited number of cores, there is a need for each thread polling inside the VMA (read, readv, recv & recvfrom) to yield the CPU to other polling thread so not to starve them from processing incoming packets. Default value is 0 (Disable) VMA_RX_PREFETCH_BYTES Size of receive buffer to prefetch into cache while processing ingress packets. The default is a single cache line of 64 bytes which should be at least 32 bytes to cover the IPoIB+IP+UDP headers and a small part of the users payload. Increasing this can help improve performance for larger user payload sizes. Value range is 32 bytes to MTU size Default value is 256 bytes VMA_RX_PREFETCH_BYTES_BEFORE_POLL Same as the above VMA_RX_PREFETCH_BYTES, only that prefetch is done before acutally getting the packets. This benefit low pps traffic latency. Disable with 0. Default value is 0 VMA_RX_CQ_DRAIN_RATE_NSEC Socket's receive path CQ drain logic rate control. When disabled (Default) the socket's receive path will first try to return a ready packet from the socket's receive ready packet queue. Only if that queue is empty will the socket check the CQ for ready completions for processing. When enabled, even if the socket's receive ready packet queue is not empty it will still check the CQ for ready completions for processing. This CQ polling rate is controls in nano-second resolution to prevent CPU consumption because of over CQ polling. This will enable a more 'real time' monitoring of the sockets ready packet queue. Recommended value is 100-5000 (nsec) Default value is 0 (Disable) VMA_GRO_STREAMS_MAX Control the number of TCP streams to perform GRO (generic receive offload) simultaneously. Disable GRO with a value of 0. Default value is 32 VMA_TCP_3T_RULES Use only 3 tuple rules for TCP, instead of using 5 tuple rules. This can improve performance for a server with listen socket which accept many connections. VMA_ETH_MC_L2_ONLY_RULES Use only L2 rules for Ethernet Multicast. All loopback traffic will be handled by VMA instead of OS. VMA_SELECT_POLL The duration in micro-seconds (usec) in which to poll the hardware on Rx path before going to sleep (pending an interrupt blocking on OS select(), poll() or epoll_wait(). The max polling duration will be limited by the timeout the user is using when calling select(), poll() or epoll_wait(). When select(), poll() or epoll_wait() path has successfull receive poll hits (see performace monitoring) the latency is improved dramatically. This comes on account of CPU utilization. Value range is -1, 0 to 100,000,000 Where value of -1 is used for infinite polling Where value of 0 is used for no polling (interrupt driven) Default value is 100000 VMA_SELECT_POLL_OS_FORCE This flag forces to poll the OS file descriptors while user thread calls select(), poll() or epoll_wait() even when no offloaded sockets are mapped. Enabling this flag causes VMA to set VMA_SELECT_POLL_OS_RATIO and VMA_SELECT_SKIP_OS to 1. This will result in VMA_SELECT_POLL number of times VMA will poll the OS file descriptors, along side with offloaded sockets, if such sockets exists. Note that setting VMA_SELECT_SKIP_OS and VMA_SELECT_POLL_OS_RATIO directly will override the values these parameters gets while VMA_SELECT_POLL_OS_FORCE is enabled. Enable with 1 Disable with 0 Default value is 0 VMA_SELECT_POLL_OS_RATIO This will enable polling of the OS file descriptors while user thread calls select(), poll() or epoll_wait() and the VMA is busy in the offloaded sockets polling loop. This will result in a signle poll of the not-offloaded sockets every VMA_SELECT_POLL_RATIO offlaoded sockets (CQ) polls. When disabled, only offlaoded sockets are polled. (See VMA_SELECT_POLL for more info) Disable with 0 Default value is 10 VMA_SELECT_SKIP_OS Similar to VMA_RX_SKIP_OS, but in select(), poll() or epoll_wait() this will force the VMA to check the non offloaded fd even though an offloaded socket has ready packets found while polling. Default value is 4 VMA_PROGRESS_ENGINE_INTERVAL VMA Internal thread safe check that the CQ is drained at least onse every N milliseconds. This mechanism allows VMA to progress the TCP stack even when the application doesn't access its socket (so it doesn't provide a context to VMA). If CQ was already drained by the application receive socket API calls then this thread goes back to sleep without any processing. Disable with 0 Default value is 10 msec VMA_PROGRESS_ENGINE_WCE_MAX Each time VMA's internal thread starts it's CQ draining, it will stop when reach this max value. The application is not limited by this value in the number of CQ elements it can ProcessId form calling any of the receive path socket APIs. Default value is 2048 VMA_CQ_MODERATION_ENABLE Enable CQ interrupt moderation. Default value is 1 (Enabled) VMA_CQ_MODERATION_COUNT Number of packets to hold before generating interrupt. Default value is 48 VMA_CQ_MODERATION_PERIOD_USEC Period in micro-seconds for holding the packet before generating interrupt. Default value is 50 VMA_CQ_AIM_MAX_COUNT Maximum count value to use in the adaptive interrupt moderation algorithm. Default value is 560 VMA_CQ_AIM_MAX_PERIOD_USEC Maximum period value to use in the adaptive interrupt moderation algorithm. Default value is 250 VMA_CQ_AIM_INTERVAL_MSEC Frequency of interrupt moderation adaptation. Intervall in milli-seconds between adaptation attempts. Use value of 0 to disable adaptive interrupt moderation. Default value is 250 VMA_CQ_AIM_INTERRUPTS_RATE_PER_SEC Desired interrupts rate per second for each ring (CQ). The count and period parameters for CQ moderation will change automatically to achieve the desired interrupt rate for the current traffic rate. Default value is 5000 VMA_CQ_POLL_BATCH_MAX Max size of the array while polling the CQs in the VMA Default value is 8 VMA_CQ_KEEP_QP_FULL If disabled (default), CQ will not try to compensate for each poll on the receive path. It will use a "debth" to remember how many WRE miss from each QP to fill it when buffers become avilable. If enabled, CQ will try to compensate QP for each polled receive completion. If buffers are short it will re-post a recently completed buffer. This causes a packet drop and will be monitored in the vma_stats. Default value is 1 (Enabled) VMA_QP_COMPENSATION_LEVEL Number of spare receive buffer CQ holds to allow for filling up QP while full receive buffers are being processes inside VMA. Default value is 256 buffers VMA_OFFLOADED_SOCKETS Create all sockets as offloaded/not-offloaded by default. Value of 1 is for offloaded, 0 for not-offloaded. Default value is 1 (Enabled) VMA_TIMER_RESOLUTION_MSEC Control VMA internal thread wakeup timer resolution (in milli seconds) Default value is 10 (milli-sec) VMA_TCP_TIMER_RESOLUTION_MSEC Control VMA internal TCP timer resolution (fast timer) (in milli seconds). Minimum value is the internal thread wakeup timer resolution (VMA_TIMER_RESOLUTION_MSEC). Default value is 100 (milli-sec) VMA_TCP_CTL_THREAD Do all tcp control flows in the internal thread. This feature should be kept disabled if using blocking poll/select (epoll is OK). Use value of 0 to disable. Use value of 1 for waking up the thread when there is work to do. Use value of 2 for waiting for thread timer to expire. Default value is disabled VMA_TCP_TIMESTAMP_OPTION Currently, LWIP is not supporting RTTM and PAWS mechanisms. see RFC1323 for info. Use value of 0 to disable. Use value of 1 for enable. Use value of 2 for OS follow up. Disabled by default (enabling causing a slight performance degradation of ~50-100 nano sec per half round trip) VMA_EXCEPTION_HANDLING Mode for handling missing support or error cases in Socket API or functionality by VMA. Useful for quickly identifying VMA unsupported Socket API or features Use value of -1 for just handling at DEBUG severity. Use value of 0 to log DEBUG message and try recovering via Kernel network stack (un-offloading the socket). Use value of 1 to log ERROR message and try recovering via Kernel network stack (un-offloading the socket). Use value of 2 to log ERROR message and return API respectful error code. Use value of 3 to log ERROR message and abort application (throw vma_error exception). Default value is -1 (notice, that in the future the default value will be changed to 0) VMA_AVOID_SYS_CALLS_ON_TCP_FD For TCP fd, avoid system calls for the supported options of: ioctl, fcntl, getsockopt, setsockopt. Non-supported options will go to OS. To activate, use VMA_AVOID_SYS_CALLS_ON_TCP_FD=1. Default value is disabled VMA_INTERNAL_THREAD_AFFINITY Control which CPU core(s) the VMA internal thread is serviced on. The cpu set should be provided as *EITHER* a hexidecmal value that represents a bitmask. *OR* as a comma delimited of values (ranges are ok). Both the bitmask and comma delimited list methods are identical to what is supported by the taskset command. See the man page on taskset for additional information. Where value of -1 disables internal thread affinity setting by VMA Bitmask Examples: 0x00000001 - Run on processor 0. 0x00000007 - Run on processors 1,2, and 3. Comma Delimited Examples: 0,4,8 - Run on processors 0,4, and 8. 0,1,7-10 - Run on processors 0,1,7,8,9 and 10. Default value is -1 (Disabled). VMA_INTERNAL_THREAD_CPUSET Select a cpuset for VMA internal thread (see man page of cpuset). The value is the path to the cpuset (for example: /dev/cpuset/my_set), or an empty string to run it on the same cpuset the process runs on. Default value is an empty string. VMA_INTERNAL_THREAD_TCP_TIMER_HANDLING Select the internal thread policy when handling TCP timers Use value of 0 for deferred handling. The internal thread will not handle TCP timers upon timer expiration (once every 100ms) in order to let application threads handling it first Use value of 1 for immediate handling. The internal thread will try locking and handling TCP timers upon timer expiration (once every 100ms). Application threads may be blocked till internal thread finishes handling TCP timers Default value is 0 (deferred handling) VMA_INTERNAL_THREAD_ARM_CQ Wakeup the internal thread for each packet that the CQ recieve. Poll and process the packet and bring it to the socket layer. This can minimize latency in case of a busy application which is not available to recieve the packet when it arrived. However, this might decrease performance in case of high pps rate application. Default value is 0 (Disabled) VMA_WAIT_AFTER_JOIN_MSEC This parameter indicates the time of delay the first packet send after receiving the multicast JOINED event from the SM This is helpful to over come loss of first few packets of an outgoing stream due to SM lengthy handling of MFT configuration on the switch chips Default value is 0 (milli-sec) VMA_THREAD_MODE By default VMA is ready for multi-threaded applications, meaning it is thread safe. If the users application is a single threaded one, then using this configuration parameter you can help eliminate VMA locks and get even better performance. Single threaded application value is 0 Multi threaded application using spin lock value is 1 Multi threaded application using mutex lock value is 2 Multi threaded application with more threads than cores using spin lock value is 3 Default value is 1 (Multi with spin lock) VMA_BUFFER_BATCHING_MODE Batching of returning Rx buffers and pulling Tx buffers per socket. In case the value is 0 then VMA will not use buffer batching. In case the value is 1 then VMA will use buffer batching and will try to periodically reclaim unused buffers. In case the value is 2 then VMA will use buffer batching with no reclaim. [future: other values are reserved] Default value is 1 VMA_MEM_ALLOC_TYPE This replaces the VMA_HUGETBL parameter logic. VMA will try to allocate data buffers as configured: 0 - "ANON" - using malloc 1 - "CONTIG" - using contiguous pages 2 - "HUGEPAGES" - using huge pages. OFED will also try to allocate QP & CQ memory accordingly: 0 - "ANON" - default - use current pages ANON small ones. "HUGE" - force huge pages "CONTIG" - force contig pages 1 - "PREFER_CONTIG" - try contig fallback to ANON small pages. "PREFER_HUGE" - try huge fallback to ANON small pages. 2 - "ALL" - try huge fallback to contig if failed fallback to ANON small pages. To overrive OFED use: (MLX_QP_ALLOC_TYPE, MLX_CQ_ALLOC_TYPE) Default value is 1 (Contiguous pages) The following VMA neigh parameters are for advanced users or Mellanox support only: VMA_NEIGH_UC_ARP_QUATA VMA will send UC ARP in case neigh state is NUD_STALE. In case that neigh state is still NUD_STALE VMA will try VMA_NEIGH_UC_ARP_QUATA retries to send UC ARP again and then will send BC ARP. VMA_NEIGH_UC_ARP_DELAY_MSEC This parameter indicates number of msec to wait betwen every UC ARP. VMA_NEIGH_NUM_ERR_RETRIES This number inidcates number of retries to restart neigh state machine in case neigh got ERROR event. Deafult value is 1 VMA_BF This flag enables / disables BF (Blue Flame) usage of the ConnectX Deafult value is 1 (Enabled) VMA_FORK Control whether VMA should support fork. Setting this flag on will cause VMA to call ibv_fork_init() function. ibv_fork_init() initializes libibverbs's data structures to handle fork() function calls correctly and avoid data corruption. If ibv_fork_init() is not called or returns a non-zero status, then libibverbs data structures are not fork()-safe and the effect of an application calling fork() is undefined. ibv_fork_init() works on Linux kernels 2.6.17 and higher which support the MADV_DONTFORK flag for madvise(). Note that VMA's default with huge pages enabled (VMA_HUGETBL) you should use an OFED stack version that support fork()ing of with huge pages (OFED 1.5 and higher). Default value is 0 (Disabled) VMA_CLOSE_ON_DUP2 When this parameter is enabled, VMA will handle the dupped fd (oldfd), as if it was closed (clear internal data structures) and only then, will forward the call to the OS. This is, in practice, a very rudimentary dup2 support. It only supports the case, where dup2 is used to close file descriptors, Default value is 1 (Enabled) VMA_MTU Size of each Rx and Tx data buffer (Maximum Transfer Unit). This value sets the fragmentation size of the packets sent by the VMA library. If VMA_MTU is 0 then for each inteterface VMA will follow the actual MTU. If VMA_MTU is greater than 0 then this MTU value is applicable to all interfaces regardless of their actual MTU Default value is 0 (following interface actual MTU) VMA_MSS VMA_MSS define the max TCP payload size that can sent without IP fragmentation. Value of 0 will set VMA's TCP MSS to be aligned with VMA_MTU configuration (leaving 40 bytes room for IP + TCP headers; "TCP MSS = VMA_MTU - 40"). Other VMA_MSS values will force VMA's TCP MSS to that specific value. Default value is 0 (following VMA_MTU) VMA_TCP_CC_ALGO TCP congestion control algorithm. The default algorithm coming with LWIP is a variation of Reno/New-Reno. The new Cubic algorithm was adapted from FreeBsd implementation. Use value of 0 for LWIP algorithm. Use value of 1 for the Cubic algorithm. Default value is 0 (LWIP). VMA_TCP_MAX_SYN_RATE Limit the number of TCP SYN packets that VMA will handle per second per listen socket. For example, in case you use 10 for this value than VMA will accept at most 10 (could be less) new connections per second per listen socket. Use a value of 0 for un-limiting the number of TCP SYN packets that can be handled. Value range is 0 to 100000. Default value is 0 (no limit) VMA_IPERF Support iperf server default test which is multithreaded. In you run the iperf server with the additional flag -U and you feel that the VMA can do better, you can disable this function (VMA_IPERF=0) Default value is 1 (Enabled) VMA_SUPPRESS_IGMP_WARNING Use VMA_SUPPRESS_IGMP_WARNING=1 to suppress the warnings about igmp version not forced to be 2. Default value is 0 (Disabled) VMA Monitoring & Performance Counters ===================================== The VMA internal performance counters include information per user sockets and a global view on select() and epoll_wait() usage by the application. Use the 'vma_stats' included utility to view the per socket information and performance counters during run time. Usage: vma_stats [-p pid] [-k directory] [-v view] [-d details] [-i interval] Defaults: find_pid=enabled, directory="/tmp/", view=1, details=1, interval=1, Options: -p, --pid=<pid> Show VMA statistics for proccess with pid: <pid> -k, --directory=<directory> Set shared memory directory path to <directory> -n, --name=<application> Show VMA statistics for application: <application> -f, --find_pid Find and show statistics for VMA instance running (default) -F, --forbid_clean By setting this flag inactive shared objects would not be removed -i, --interval=<n> Print report every <n> seconds -c, --cycles=<n> Do <n> report print cycles and exit, use 0 value for infinite (default) -v, --view=<1|2|3|4|5> Set view type:1- basic info,2- extra info,3- full info,4- mc groups,5- similar to 'netstat -tunaep' -d, --details=<1|2> Set details mode:1- to see totals,2- to see deltas -z, --zero Zero counters -l, --log_level=<level> Set VMA log level to <level>(1 <= level <= 7) -D, --details_level=<level> Set VMA log details level to <level>(0 <= level <= 3) -s, --sockets=<list|range> Log only sockets that match <list> or <range>, format: 4-16 or 1,9 (or combination) -V, --version Print version -h, --help Print this help message Use VMA_STATS_FILE to get internal VMA statistics like vma_stats provide. If this parameter is set and the user application performed transmite or receive activity on a socket, then these values will be logs once the sockets are closed. Below is a logout example of a socket performance counters. Below the logout example there is some explanations about the numbers. VMA: [fd=10] Tx Offload: 455 KB / 233020 / 0 / 3 [bytes/packets/drops/errors] VMA: [fd=10] Tx OS info: 0 KB / 0 / 0 [bytes/packets/errors] VMA: [fd=10] Rx Offload: 455 KB / 233020 / 0 / 0 [bytes/packets/eagains/errors] VMA: [fd=10] Rx byte: max 200 / dropped 0 (0.00%) / limit 2000000 VMA: [fd=10] Rx pkt : max 1 / dropped 0 (0.00%) VMA: [fd=10] Rx OS info: 0 KB / 0 / 0 [bytes/packets/errors] VMA: [fd=10] Rx poll: 0 / 233020 (100.00%) [miss/hit] Looking good :) - No errors on transmite or receive on this socket (user fd=10) - All the traffic was offloaded. No packets transmitted or receive via the OS. - Just about no missed Rx polls (see VMA_RX_POLL & VMA_SELECT_POLL), meaning the receiving thread did not get to a blocked state to cause a contexts switch and hurt latency. - No dropped packets caused by socket receive buffer limit (see VMA_RX_BYTES_MIN) IGMP ==== To participate in a multicast that spans multiple networks, a host must inform local routers. The local routers contact other routers, passing on the membership information and establishing routes. Routers and hosts that implement multicast use the Internet Group Management Protocol (IGMP) to communicate group membership information, later the local router can propagate multicast membership information to other routers through the Internet. In order to enable this behavior at applications running over IPoIB networks please use IGMPv2 on the network: 'echo 2 > /proc/sys/net/ipv4/conf/ibX/force_igmp_version' Interrupt Moderation ==================== The basic idea behind interrupt moderation is that the HW will not generate interrupt for each packet, but instead only after some amount of packets received or after the packet was held for some time. The adaptive interrupt moderation change this packet count and time period automatically to reach a desired rate of interrupts. 1. Use VMA_RX_POLL=0 and VMA_SELECT_POLL=0 to work in interrupt driven mode. 2. Control the period and frame count parameters with: VMA_CQ_MODERATION_COUNT - hold #count frames before interrupt VMA_CQ_MODERATION_PERIOD_USEC - hold #usec before interrupt 3. Control the adaptive algorithm with the following: VMA_CQ_AIM_MAX_COUNT - max possible #count frames to hold VMA_CQ_AIM_MAX_PERIOD_USEC - max possible #usec to hold VMA_CQ_AIM_INTERRUPTS_RATE_PER_SEC - desired interrupt rate VMA_CQ_AIM_INTERVAL_MSEC - frequency of adaptation 4. Disable CQ moderation with VMA_CQ_MODERATION_ENABLE=0 5. Disable Adaptive CQ moderation with VMA_CQ_AIM_INTERVAL_MSEC=0 Notes ===== * Multicast loopback behavior: There is a different behavior between IPoIB and VMA when dealing with multicast packets at the same machine: - When sending from VMA application to ipoib application on the same machine the packet will never be accepted by the ipoib side (even when the loopback is enabled) - When sending from ipoib application to VMA application on the same machine, the packet will always be accepted by the VMA side (even when the loop is disabled Troubleshooting =============== * High log level: VMA WARNING: ************************************************************* VMA WARNING: * VMA is currently configured with high log level * VMA WARNING: * Application performance will decrease in this log level! * VMA WARNING: * This log level is recommended for debugging purposes only * VMA WARNING: ************************************************************* This warning message means that you are using VMA with high log level: VMA_TRACELEVEL variable value is set to 4 or more. In order to fix it - set VMA_TRACELEVEL to it's default value: 3 * Ethernet RAW_PACKET_QP limited to privilege users VMA WARNING: ****************************************************************************** VMA WARNING: * Verbs RAW_PACKET QP type creation is limited for root user access * VMA WARNING: * Working in this mode might causes VMA malfunction over Ethernet interfaces * VMA WARNING: * WARNING: the following steps will restart your network interface! * VMA WARNING: * 1. "echo options ib_uverbs disable_raw_qp_enforcement=1 > /etc/modprobe.d/ib_uverbs.conf" * VMA WARNING: * 2. "/etc/init.d/openibd restart" * VMA WARNING: * Read the RAW_PACKET QP root access enforcement section in the VMA's User Manual for more information * VMA WARNING: ****************************************************************************** This warning message means that VMA tried to create a HW QP resource over Eth interface while the kernel requires this operation to be done only by privileged users. root can enable this for regular users as well by: 1. "echo options ib_uverbs disable_raw_qp_enforcement=1 > /etc/modprobe.d/ib_uverbs.conf" 2. "/etc/init.d/openibd restart" * IGMP not forced to V2: VMA WARNING: ************************************************************************ VMA WARNING: IGMP Version flag is not forced to IGMPv2 for interface ib2! VMA WARNING: Working in this mode might causes VMA functionality degradation VMA WARNING: Please "echo 2 > /proc/sys/net/ipv4/conf/ib2/force_igmp_version" VMA WARNING: before loading your application with VMA library VMA WARNING: Please refer to the IGMP section in the VMA's User Manual for more information VMA WARNING: ************************************************************************ This warning message means that you are using on IB interfaces IGMP version other then 2 which is supported by VMA. You can disabled VMA_IGMP=0 if you do not need to receive multicast packets from the Ethernet to the InfiniBand fabric or if you are using a Ethernet switch that does not do IGMP snooping. If you do expect to receive multicast packets from the Ethernet to the InfiniBand fabric with VMA or if your Ethernet switch does IGMP snooping then you should force IGMP working mode to version 2 in all your hosts as well as your routers. You can also do: "echo 2 > /proc/sys/net/ipv4/conf/all/force_igmp_version" * IGMP Forced Version info missing: VMA WARNING: ************************************************************************ VMA WARNING: Error in reading IGMP Version flag for interface 192.168.0.10! VMA WARNING: Working in this mode most probably causes VMA performance degradation VMA WARNING: Please refer to the IGMP section in the VMA's User Manual for more information VMA WARNING: ************************************************************************ This warning message means that you are using VMA with an older version of OFED which does not support user space IGMP. If you do expect to receive multicast packets from the Ethernet to the InfiniBand fabric with VMA then you need to upgrade your OFED based network stack. * Huge pages out of resource: VMA WARNING: *************************************************************** VMA WARNING: * NO IMMEDIATE ACTION NEEDED! * VMA WARNING: * Not enough hugepage resources for VMA memory allocation. * VMA WARNING: * VMA will continue working with regular memory allocation. * VMA INFO : * Optional: 1. Switch to a different memory allocation type * VMA_INFO : * (VMA_MEM_ALLOC_TYPE= 0 or 1) * VMA INFO : * 2. Restart process after increasing the number of * VMA INFO : * hugepages resources in the system: * VMA INFO : * "cat /proc/meminfo | grep -i HugePage" * VMA INFO : * "echo 1000000000 > /proc/sys/kernel/shmmax" * VMA INFO : * "echo 800 > /proc/sys/vm/nr_hugepages" * VMA WARNING: * Please refer to the memory allocation section in the VMA's * VMA WARNING: * User Manual for more information * VMA WARNING: *************************************************************** This warning message means that you are using VMA with hugepages memory allocation, but not enough huge pages resources are available in the system. If you want VMA to take full advantage of the performance benefits of huge pages then you should restart the application after adding more hugepages resources in your system similar to the details in the warning message above or trying to free unused hupge pages shared memory segments with the below script. NOTE: Use 'ipcs -m' and 'ipcrm -m shmid' to check and clean unused shared memory segments. Below is a short script to help you release VMAs unused huge pages resources: for shmid in `ipcs -m | grep 0x00000000 | awk '{print $2}'`; do echo 'Clearing' $shmid; ipcrm -m $shmid; done; * Not supported Bonding Configuration: VMA WARNING: ****************************************************************************** VMA WARNING: VMA doesn't support current bonding configuration of bond0. VMA WARNING: The only supported bonding mode is "802.3ad(#4)" or "active-backup(#1)" VMA WARNING: with "fail_over_mac=1" or "fail_over_mac=0". VMA WARNING: The effect of working in unsupported bonding mode is undefined. VMA WARNING: Read more about Bonding in the VMA's User Manual VMA WARNING: ****************************************************************************** This warning message means that VMA has detected bonding device which is configured to work in mode which is not supported by VMA, this means that VMA will not support high avilability events for that interface. VMA currently supports just active-backup(#1) or 802.3ad(#4) and fail_over_mac = 1 or 0 mode. In order to fix this issue please change the bonding configuration. Example: Lets assume that the bonding device is bond0, which has two slaves: ib0 and ib1. Shut down the bond0 interface: #ifconfig bond0 down Find all the slaves of bond0: #cat sys/class/net/bond0/bonding/slaves ib0 ib1 Free all the slaves: #echo -ib0 > /sys/class/net/bond0/bonding/slaves #echo -ib1 > /sys/class/net/bond0/bonding/slaves Change the bond mode: #echo active-backup > /sys/class/net/bond0/bonding/mode Change the fail_over_mac mode: #echo 1 > /sys/class/net/bond0/bonding/fail_over_mac Enslave the interfaces back: #echo +ib0 > /sys/class/net/bond0/bonding/slaves #echo +ib1 > /sys/class/net/bond0/bonding/slaves Bring up the bonding interface: #ifconfig bond0 up OR #ifconfig bond0 <ip> netmask <netmask> up * Not supported Bonding & VLAN Configuration: VMA WARNING: ****************************************************************** VMA WARNING: bond0.10: vlan over bond while fail_over_mac=1 is not offloaded VMA WARNING: ****************************************************************** This warning message means that VMA has detected bonding device which is configured with VLAN over it while fail_over_mac=1. This means that the bond will not be offloaded. In order to fix this issue please change the bonding configuration.