VMA Configuration

<program-name|*>

Define the program name (not including the path) to which the control statements appearing below this statement apply.

Wildcards with the same semantics as "ls" are supported (* and ?).

For example:

• db2* matches any program with a name starting with db2.

• t?cp matches ttcp, etc.

<user-defined-id|*>

Specify the process ID to which the control statements appearing below this statement apply.

transport

Define the mode of transport:

• vma – VMA should be used.

• os – the socket should be handled by the OS network stack. In this mode, the sockets are not offloaded.

The default is vma.

role

Specify one of the following roles:

• tcp_server – for listen sockets. Accepted sockets follow listen sockets. Defined by local_ip:local_port.

• tcp_client – for connected sockets. Defined by remote_ip:remote_port:local_ip:local_port

• udp_sender – for TX flows. Defined by remote_ip:remote_port

• udp_receiver – for RX flows. Defined by local_ip:local_port

• udp_connect – for UDP connected sockets. Defined by remote_ip:remote_port:local_ip:local_port

address

You can specify the local address the server is bind to or the remote server address the client connects to.

The syntax for address matching is:

<IPv4 address>[/<prefix_length>]|*

• IPv4 address – [0-9]+\.[0-9]+\.[0-9]+\.[0-9]+ each sub number < 255

• prefix_length – [0-9]+ and with value <= 32. A prefix_length of 24 # matches the subnet mask 255.255.255.0 . A prefix_length of 32 requires matching of the exact IP.

port range

Define the port range as:

            

            
start-port[-end-port]
        

Port range: 0-65536

VMA_TRACELEVEL

PANIC = 0 – Panic level logging.

This trace level causes fatal behavior and halts the application, typically caused by memory allocation problems. PANIC level is rarely used.

ERROR = 1 – Runtime errors in VMA.

Typically, this trace level assists you to identify internal logic errors, such as errors from underlying OS or InfiniBand verb calls, and internal double mapping/unmapping of objects.

WARN = WARNING = 2– Runtime warning that does not disrupt the application workflow.

A warning may indicate problems in the setup or in the overall setup configuration. For example, address resolution failures (due to an incorrect routing setup configuration), corrupted IP packets in the receive path, or unsupported functions requested by the user application.

INFO = INFORMATION = 3– General information passed to the user of the application.

This trace level includes configuration logging or general information to assist you with better use of the VMA library.

DETAILS – Greater general information passed to the user of the application.

This trace level includes printing of all environment variables of VMA at start up.

DEBUG = 4 – High-level insight to the operations performed in VMA.

In this logging level all socket API calls are logged, and internal high-level control channels log their activity.

FINE = FUNC = 5 – Low-level runtime logging of activity.

This logging level includes basic Tx and Rx logging in the fast path. Note that using this setting lowers application performance. We recommend that you use this level with the VMA_LOG_FILE parameter.

FINER = FUNC_ALL = 6 – Very low-level runtime logging of activity. This logging level drastically lowers application performance. We recommend that you use this level with the VMA_LOG_FILE parameter.

VMA_LOG_DETAILS

Provides additional logging details on each log line.

0 = Basic log line

1 = With ThreadId

2 = With ProcessId and ThreadId

3 = With Time, ProcessId, and ThreadId (Time is the amount of milliseconds from the start of the process)

Default: 0

For VMA_TRACELEVEL >= 4, this value defaults to 2.

VMA_LOG_FILE

Redirects all VMA logging to a specific user-defined file.

This is very useful when raising the VMA_TRACELEVEL.

The VMA replaces a single '%d' appearing in the log file name with the pid of the process loaded with VMA. This can help when running multiple instances of VMA, each with its own log file name.

Example: VMA_LOG_FILE=/tmp/vma_log.txt

VMA_CONFIG_FILE

Sets the full path to the VMA configuration file.

Example: VMA_CONFIG_FILE=/tmp/libvma.conf

Default: /etc/libvma.conf

LOG_COLORS

Uses a color scheme when logging; red for errors and warnings, and dim for very low level debugs.

VMA_LOG_COLORS is automatically disabled when logging is done directly to a non-terminal device (for example, when VMA_LOG_FILE is configured).

Default: 1 (Enabled)

VMA_CPU_USAGE_STATS

Calculates the VMA CPU usage during polling hardware loops. This information is available through VMA stats utility.

Default: 0 (Disabled)

VMA_APPLICATION_ID

Specifies a group of rules from libvma.conf for VMA to apply.

Example: VMA_APPLICATION_ID=iperf_server

Default: VMA_DEFAULT_APPLICATION_ID (match only the '*' group rule)

VMA_HANDLE_SIGINTR

When enabled, the VMA handler is called when an interrupt signal is sent to the process.

VMA also calls the application's handler, if it exists.

Range: 0 to 1

Default: 0 (Disabled)

VMA_HANDLE_SIGSEGV

When enabled, a print backtrace is performed, if a segmentation fault occurs.

Range: 0 to 1

Default: 0 (Disabled)

VMA_STATS_FD_NUM

Maximum number of sockets monitored by the VMA statistics mechanism.

Range: 0 to 1024

Default: 100

VMA_STATS_FILE

Redirects socket statistics to a specific user-defined file.

VMA dumps each socket's statistics into a file when closing the socket.

Example: VMA_STATS_FILE=/tmp/stats

VMA_STATS_SHMEM_DIR

Sets the directory path for VMA to create the shared memory files for vma_stats.

If this value is set to an empty string: “ “, no shared memory files are created.

Default: /tmp/

VMA_VMAD_NOTIFY_DIR

Sets the directory path for VMA to write files used by vmad.

Default value is /tmp/vma

Note: when used vmad must be run with --notify-dir directing the same folder.

VMA_TCP_MAX_SYN_RATE

Limits the number of TCP SYN packets that VMA handles per second for each listen socket.

Example: by setting this value to 10, the maximal number of TCP connection accepted by VMA per second for each listen socket will be 10.

Set this value to 0 for VMA to handle an unlimited number of TCP SYN packets per second for each listen socket.

Value range is 0 to 100000.

Default value is 0 (no limit)

VMA_TX_SEGS_TCP

Number of TCP LWIP segments allocation for each VMA process.

Default: 1000000

VMA_TX_BUFS

Number of global Tx data buffer elements allocation.

Default: 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: 3000

The size of the Tx buffers is determined by the VMA_MTU parameter value (see below).

If this value is raised, the packet rate peaking can be better sustained; however, this increases memory usage. A smaller number of data buffers gives a smaller memory footprint, but may not sustain peaks in the data rate.

VMA_TX_WRE_BATCHING

Controls the number of aggregated Work Requests Elements before receiving a completion signal (CQ entry) from the hardware. Previously this number was hard coded as 64.

The new update allows a better control of the jitter encountered in the Tx completion handling.

Valid value range: 1-64

Default: 64

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: 220

VMA_TX_MC_LOOPBACK

Sets the initial value used internally by the VMA to control multicast loopback packet behavior during transmission. An application that calls setsockopt() with IP_MULTICAST_LOOP overwrites the initial value set by this parameter.

Range: 0 - Disabled, 1 - Enabled

Default: 1

VMA_TX_NONBLOCKED_EAGAINS

Returns value 'OK' on all send operations that are performed on a non-blocked udp socket. This is the OS default behavior. The datagram sent is silently dropped inside the VMA or the network stack.

When set to Enabled (set to 1), VMA returns with error EAGAIN if it was unable to accomplish the send operation, and the datagram was dropped.

In both cases, a dropped Tx statistical counter is incremented.

Default: 0 (Disabled)

VMA_TX_PREFETCH_BYTES

Accelerates an offloaded send operation by optimizing the cache. Different values give an optimized send rate on different machines. We recommend that you adjust this parameter to your specific hardware.

Range: 0 to MTU size

Disable with a value of 0

Default: 256 bytes

VMA_RX_BUFS

The number of Rx data buffer elements allocated for the processes. These data buffers are used by all QPs on all HCAs, as determined by the VMA_QP_LOGIC.

Default: 200000

VMA_RX_WRE

The number of Work Request Elements allocated in all received QPs.

Default: 16000

VMA_RX_WRE_BATCHING

Number of Work Request Elements and RX buffers to batch before recycling.

Batching decreases the latency mean, but might increase latency STD.

Valid value range: 1-1024

Default: 64

VMA_RX_BYTES_MIN

The minimum value in bytes used per socket by the VMA when applications call to setsockopt(SO_RCVBUF).

If the application tries to set a smaller value than configured in VMA_RX_BYTES_MIN, VMA forces this minimum limit value on the socket.

VMA offloaded sockets receive the maximum amount of ready bytes. If the application does not drain sockets and the byte limit is reached, newly received datagrams are dropped.

The application's socket usage of current, max,dropped bytes and packet counters, can be monitored using vma_stats.

Default: 65536

VMA_RX_POLL

The number of times to unsuccessfully poll an Rx for VMA packets before going to sleep.

Range: -1, 0 … 100,000,000

Default: 100,000

This value can be reduced to lower the load on the CPU. However, the price paid for this is that the Rx latency is expected to increase.

Recommended values:

• 10000 – when CPU usage is not critical and Rx path latency is critical.

• 0 – when CPU usage is critical and Rx path latency is not critical.

• -1 – causes infinite polling.

Once the VMA has gone to sleep, if it is in blocked mode, it waits for an interrupt; if it is in non-blocked mode, it returns -1.

This Rx polling is performed when the application is working with direct blocked calls to read(), recv(), recvfrom(), and recvmsg().

When the Rx path has successful poll hits, the latency improves dramatically. However, this causes increased CPU utilization. For more information, see Debugging, Troubleshooting, and Monitoring.

VMA_RX_POLL_INIT

VMA maps all UDP sockets as potential Offloaded-capable. Only after ADD_MEMBERSHIP is set, the offload starts working and the CQ polling starts VMA.

This parameter controls the polling count during this transition phase where the socket is a UDP unicast socket and no multicast addresses were added to it.

Once the first ADD_MEMBERSHIP is called, the VMA_RX_POLL (above) takes effect.

Value range is similar to the VMA_RX_POLL (above).

Default: 0

VMA_RX_UDP_POLL_OS_RATIO

Defines the ratio between VMA CQ poll and OS FD poll.

This will result in a single poll of the not-offloaded sockets every VMA_RX_UDP_POLL_OS_RATIO offloaded 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 offloaded sockets are polled.

This parameter replaces the two old parameters:

• VMA_RX_POLL_OS_RATIO and

• VMA_RX_SKIP_OS

Disable with 0

Default: 10

VMA_HW_TS_CONVERSION

Defines timestamp conversion method.

The value of VMA_HW_TS_CONVERSION is determined by all devices, that is, if the hardware of one device does not support the conversion, then it will be disabled for the other devices.

Currently only UDP RX flow is supported.

Options = [0,1,2,3,4]:

• 0 – Disabled

• 1 – Raw-HW time

  Only convert the timestamp to seconds.nano_seconds time units (or disable if hardware does not supports).

• 2 – Best possible – Raw-HW or system time.

  Sync to system time, then Raw hardware time. Disable if none of them are supported by hardware.

• 3 – Sync to system time

  Convert the timestamp to seconds.nano_seconds time units. Comparable to UDP receive software timestamp. Disable if hardware does not supports.

• 4 – PTP Sync

  Convert the timestamp to seconds.nano_seconds time units. In case it is not supported – will apply option 3 (or disable if hardware does not supports).

Default value: 3 (Sync to system time)

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 VMA (read, readv, recv, and recvfrom) to yield the CPU to another polling thread so as not to starve them from processing incoming packets.

Default: 0 (Disabled)

VMA_RX_PREFETCH_BYTES

The size of the receive buffer to prefetch into the 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 user payload.

Increasing this size can help improve performance for larger user payloads.

Range: 32 bytes to MTU size

Default: 256 bytes

VMA_RX_CQ_DRAIN_RATE_NSEC

Socket's receive path CQ drain logic rate control.

When disabled (default), the socket's receive path attempts to return a ready packet from the socket's receive ready packet queue. If the ready receive packet queue is empty, the socket checks the CQ for ready completions for processing.

When enabled, even if the socket's receive ready packet queue is not empty, this parameter checks the CQ for ready completions for processing. This CQ polling rate is controlled in nanosecond resolution to prevent CPU consumption due to over CQ polling. This enables improved 'real-time' monitoring of the socket ready packet queue.

Recommended value is 100-5000 (nsec)

Default: 0 (Disabled)

VMA_RX_POLL_ON_TX_TCP

Enables TCP RX polling during TXP TX operation for faster TCP ACK reception

Default: 0 (Disabled)

VMA_GRO_STREAMS_MAX

Controls the number of TCP streams to perform GRO (generic receive offload) simultaneously.

Disable GRO with a value of 0.

Default: 32

VMA_TCP_3T_RULES

Uses only 3 tuple rules for TCP, instead of using 5 tuple rules.

This can improve performance for a server with a listen socket which accepts many connections from the same source IP.

Enable with a value of 1.

Default: 0 (Disabled)

VMA_UDP_3T_RULES

This parameter is relevant in case the application uses connected UDP sockets. 3 tuple rules are used in hardware flow steering rule when the parameter is enabled, and in 5 tuple flow steering rule when it is disabled. Enabling this option can reduce hardware flow steering resources. However, when it is disabled, the application might see benefits in latency and cycles per packet. Default: 1 (Enable)

VMA_ETH_MC_L2_ONLY_RULES

Uses only L2 rules for Ethernet Multicast.

All loopback traffic will be handled by VMA instead of OS.

Enable with a value of 1.

Default: 0 (Disabled)

VMA_SELECT_POLL

The duration in micro-seconds (usec) in which to poll the hardware on Rx path before blocking for an interrupt (when waiting and also when calling select(), poll(), or epoll_wait()).

Range: -1, 0 … 100,000,000

Default: 100,000

When the selected path has successfully received poll hits, the latency improves dramatically. However, this comes at the expense of CPU utilization. For more information, see Debugging, Troubleshooting, and Monitoring.

VMA_SELECT_POLL_OS_RATIO

This enables polling the OS file descriptors while the user thread calls select(), poll(), or epoll_wait(), and VMA is busy in the offloaded socket polling loop. This results in a single poll of the non-offloaded sockets every VMA_SELECT_POLL_RATIO offloaded socket (CQ) polls.

When disabled, only offloaded sockets are polled.

(See VMA_SELECT_POLL for more information.)

Disable with 0

Default: 10

VMA_SELECT_SKIP_OS

In select(), poll(), or epoll_wait()forces the VMA to check the non-offloaded sockets even though an offloaded socket has a ready packet that was found while polling.

Range: 0 … 10,000

Default: 4

VMA_CQ_POLL_BATCH_MAX

The maximum size of the array while polling the CQs in the VMA.

Default: 16

VMA_PROGRESS_ENGINE_INTERVAL

Internal VMA thread safety which checks that the CQ is drained at least once every N milliseconds. This mechanism allows VMA to progress the TCP stack even when the application does not access its socket (so it does not provide a context to VMA). If the CQ was already drained by the application receive socket API calls, this thread goes back to sleep without any processing.

Disable with 0

Default: 10 milliseconds

VMA_PROGRESS_ENGINE_WCE_MAX

Each time the VMA's internal thread starts its CQ draining, it stops when it reaches this maximum value.

The application is not limited by this value in the number of CQ elements that it can ProcessId from calling any of the receive path socket APIs.

Default: 2048

VMA_CQ_MODERATION_ENABLE

Enable CQ interrupt moderation.

Default: 1 (Enabled)

VMA_CQ_MODERATION_COUNT

Number of packets to hold before generating interrupt.

Default: 48

VMA_CQ_MODERATION_PERIOD_USEC

Period in microseconds for holding the packet before generating interrupt.

Default: 50

VMA_CQ_AIM_MAX_COUNT

Maximum count value to use in the adaptive interrupt moderation algorithm.

Default: 560

VMA_CQ_AIM_MAX_PERIOD_USEC

Maximum period value to use in the adaptive interrupt moderation algorithm.

Default: 250

VMA_CQ_AIM_INTERVAL_MSEC

Frequency of interrupt moderation adaptation.

Interval in milliseconds between adaptation attempts.

Use value of 0 to disable adaptive interrupt moderation.

Default: 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: 5000

VMA_CQ_KEEP_QP_FULL

If disabled (default), the CQ does not try to compensate for each poll on the receive path. It uses a "debt" to remember how many WRE are missing from each QP, so that it can fill it when buffers become available.

If enabled, CQ tries to compensate QP for each polled receive completion. If there is a shortage of buffers, it reposts a recently completed buffer. This causes a packet drop, and is monitored in vma_stats.

Default: 1 (Enabled)

VMA_QP_COMPENSATION_LEVEL

The number of spare receive buffer CQ holds that can be allowed for filling up QP while full receive buffers are being processed inside VMA.

Default: 256 buffers

VMA_OFFLOADED_SOCKETS

Creates all sockets as offloaded/not-offloaded by default.

• 1 is used for offloaded

• 0 is used for not-offloaded

Default: 1 (Enabled)

VMA_TIMER_RESOLUTION_MSEC

Control VMA internal thread wakeup timer resolution (in milliseconds).

Default: 10 (milliseconds)

VMA_TCP_TIMER_RESOLUTION_MSEC

Controls VMA internal TCP timer resolution (fast timer) (in milliseconds). Minimum value is the internal thread wakeup timer resolution (VMA_TIMER_RESOLUTION_MSEC).

Default: 100 (milliseconds)

VMA_TCP_CTL_THREAD

Does all TCP control flows in the internal thread.

This feature should be disabled if using blocking poll/select (epoll is OK).

• Use value of 0 to disable

• Use value of 1 to wake up the thread when there is work to be done

• Use value of 2 to wait for thread timer to expire

Default: 0 (disabled)

VMA_TCP_TIMESTAMP_OPTION

Currently, LWIP is not supporting RTTM and PAWS mechanisms.

See RFC1323 for info.

• Use value of 0 to disable (enabling causing a slight performance degradation of ~50-100 nano sec per half round trip).

• Use value of 1 for enable.

• Use value of 2 for OS follow up.

Default: 0 (disabled)

VMA_TCP_NODELAY

If set, it disables the Nagle algorithm option for each TCP socket during initialization. Meaning that TCP segments are always sent as soon as possible, even if there is only a small amount of data.

For more information on TCP_NODELAY flag refer to TCP manual page.

Valid Values are:

• 0 to disable.

• 1 to enable (default)

VMA_TCP_QUICKACK

If set, it disables the delayed acknowledge ability. Meaning that TCP will respond after every packet.

For more information on TCP_QUICKACK flag refer to TCP manual page.

Valid Values are:

• 0 to disable.

• 1 to enable (default)

VMA_EXCEPTION_HANDLING

Handles missing support or error cases in Socket API or functionality by VMA.

It quickly identifies VMA unsupported Socket API or features.

• Use value of -1 to handle 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: -1

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: 0 (disabled)

VMA_THREAD_MODE

By default VMA is ready for multi-threaded applications, meaning it is thread-safe.

If the user application is single threaded, use this configuration parameter to help eliminate VMA locks and improve performance.

Values:

• 0 – Single-threaded application

• 1 – Multi-threaded application with spin lock

• 2 – Multi-threaded application with mutex lock

• 3 – Multi-threaded application with more threads than cores using spin lock

Default: 1 (Multi with spin lock)

VMA_BUFFER_BATCHING_MODE

Enables batching of returning Rx buffers and pulling Tx buffers per socket.

• If the value is 0, then VMA will not use buffer batching

• If the value is 1, then VMA will use buffer batching and will try to periodically reclaim unused buffers

• If the value is 2, then VMA will use buffer batching with no reclaim

Default: 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 override OFED use: (MLX_QP_ALLOC_TYPE, MLX_CQ_ALLOC_TYPE).

Default: 1 (Contiguous pages)

VMA_FORK

Controls VMA fork support. 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 later, which support the MADV_DONTFORK flag for madvise().

You should use an OFED stack version that supports fork() with huge pages (MLNX_OFED 1.5.3 to 3.2 and 4.0-2.0.0.0 and later). VMA allocates huge pages (VMA_HUGETBL) by default.

For limitations of using fork() with VMA, please refer to the Release Notes.

Default: 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 interface 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: 0 (following interface actual MTU)

VMA_MSS

Defines the max TCP payload size that can be sent without IP fragmentation.

Value of 0 will set VMA's TCP MSS to be aligned with VMA_MTU configuration (leaving 40 bytes of 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: 0 (following VMA_MTU)

VMA_CLOSE_ON_DUP2

When this parameter is enabled, VMA handles the duplicated file descriptor (oldfd), as if it is closed (clear internal data structures) and only then forwards the call to the OS.

This is, in effect, a very rudimentary dup2 support. It supports only the case where dup2 is used to close file descriptors.

Default: 1 (Enabled)

VMA_INTERNAL_THREAD_AFFINITY

Controls which CPU core(s) the VMA internal thread is serviced on. The CPU set should be provided as either a hexadecimal 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.

The -1 value disables the 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: -1.

VMA_INTERNAL_THREAD_CPUSET

Selects a CPUSET for VMA internal thread (For further information, see man page of cpuset).

The value is either 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.

VMA_INTERNAL_THREAD_ARM_CQ

Wakes up the internal thread for each packet that the CQ receives.

Polls and processes the packet and brings it to the socket layer.

This can minimize latency for a busy application that is not available to receive the packet when it arrives.

However, this might decrease performance for high pps rate applications.

Default: 0 (Disabled)

VMA_INTERNAL_THREAD_TCP_TIMER_HANDLING

Selects 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_WAIT_AFTER_JOIN_MSEC

This parameter indicates the time of delay the first packet is send after receiving the multicast JOINED event from the SM.

This is helpful to overcome loss of first few packets of an outgoing stream due to SM lengthy handling of MFT configuration on the switch chips.

Default: 0 (milli-sec)

VMA_NEIGH_UC_ARP_QUATA

VMA will send UC ARP in case neigh state is NUD_STALE.

If 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.

Default: 3

VMA_NEIGH_UC_ARP_DELAY_MSEC

This parameter indicates number of msec to wait between every UC ARP.

Default: 10000

VMA_NEIGH_NUM_ERR_RETRIES

Indicates number of retries to restart NEIGH state machine if NEIGH receives ERROR event.

Default: 1

VMA_BF

Enables/disables BlueFlame usage of the card.

Default: 1 (Enabled)

VMA_SOCKETXTREME

When this parameter is enabled, VMA operates in SocketXtreme mode. SocketXtreme mode brings down latency, eliminating copy operations and increasing throughput, thus allowing applications to further utilize true kernel bypass architecture.

An application should use a socket extension API named SocketXtreme.

Default: 0 (Disabled)

VMA_TRIGGER_DUMMY_SEND_GETSOCKNAME

This parameter triggers dummy packet sent from getsockname() to warm up the caches.

For more information see section "Dummy Send" to Improve Low Message Rate Latency.

Default: 0 (Disable)

VMA_SPEC

VMA predefined specification profile for latency:

• latency – Latency profile spec – optimized latency on all use cases. System is tuned to keep balance between Kernel and VMA.

  Warning

  It may not reach the maximum bandwidth.

  Example: VMA_SPEC=latency

• multi_ring_latency – Multi ring latency spec – optimized for use cases that are keen on latency where two applications communicate using send-only and receive-only TCP sockets

  Example: VMA_SPEC=multi_ring_latency

VMA_RING_ALLOCATION_LOGIC_TX

VMA_RING_ALLOCATION_LOGIC_RX

Ring allocation logic is used to separate the traffic into different rings.

By default, all sockets use the same ring for both RX and TX over the same interface. For different interfaces, different rings are used, even when specifying the logic to be per socket or thread.

Using different rings is useful when tuning for a multi-threaded application and aiming for HW resource separation.

The logic options are:

• 0 – Ring per interface

• 1 – Ring per IP address (using IP address)

• 10 – Ring per socket (using socket ID 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: 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 iteration (of accessing the ring), the current thread ID is checked to see whether the ring matches the current thread.

If not, ring migration is considered. If the ring continues to be accessed from the same thread for a certain iteration, the socket is migrated to this thread ring.

Use a value of -1 in order to disable migration.

Default: 100

VMA_RING_LIMIT_PER_INTERFACE

Limits 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, several sockets will share the same ring.

Use a value of 0 for an unlimited number of rings.

Default: 0 (no limit)

VMA_RING_DEV_MEM_TX

VMA can use the on-device-memory to store the egress packet if it does not fit into the BF inline buffer. This improves application egress latency by reducing the PCI transactions.

Using VMA_RING_DEV_MEM_TX, enables the user to set the amount of the on-device-memory buffer allocated for each TX ring.

The total size of the on-device-memory is limited to 256k for a single port HCA and to 128k for dual port HCA.

Default value is 0

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.

Use value of 2 in order to disable the congestion algorithm.

Default: 0 (LWIP).