DOKK / manpages / debian 11 / libfabric-dev / fi_sockets.7.en
fi_sockets(7) @VERSION@ fi_sockets(7)

fi_sockets - The Sockets Fabric Provider

The sockets provider is being deprecated in favor of the tcp, udp, and utility providers. Further work on the sockets provider will be minimal. Most applications should instead use the tcp provider instead.

The sockets provider is a general purpose provider that can be used on any system that supports TCP sockets. The provider is not intended to provide performance improvements over regular TCP sockets, but rather to allow developers to write, test, and debug application code even on platforms that do not have high-performance fabric hardware. The sockets provider supports all libfabric provider requirements and interfaces.

The sockets provider supports all the features defined for the libfabric API. Key features include:

The provider supports all endpoint types: FI_EP_MSG, FI_EP_RDM, and FI_EP_DGRAM.
The following data transfer interface is supported for a all endpoint types: fi_msg. Additionally, these interfaces are supported for reliable endpoints (FI_EP_MSG and FI_EP_RDM): fi_tagged, fi_atomic, and fi_rma.
The sockets provider supports all operational modes including FI_CONTEXT and FI_MSG_PREFIX.
Sockets provider supports both FI_PROGRESS_AUTO and FI_PROGRESS_MANUAL, with a default set to auto. When progress is set to auto, a background thread runs to ensure that progress is made for asynchronous requests.

Sockets provider attempts to emulate the entire API set, including all defined options. In order to support development on a wide range of systems, it is implemented over TCP sockets. As a result, the performance numbers are lower compared to other providers implemented over high-speed fabric, and lower than what an application might see implementing to sockets directly.

Does not support FI_ADDR_STR address format.

The sockets provider checks for the following environment variables -

An integer value that specifies how many milliseconds to spin while waiting for progress in FI_PROGRESS_AUTO mode.
An integer value that specifies how many milliseconds to wait for one connection establishment.
An integer value that specifies the number of socket connection retries before reporting as failure.
An integer to specify the default connection map size.
An integer to specify the default address vector size.
An integer to specify the default completion queue size.
An integer to specify the default event queue size.
An integer value to specify the drop rate of dgram frame when endpoint is FI_EP_DGRAM. This is for debugging purpose only.
If specified, progress thread is bound to the indicated range(s) of Linux virtual processor ID(s). This option is currently not supported on OS X. The usage is - id_start[-id_end[:stride]][,].
A boolean to enable the keepalive support.
An integer to specify the idle time in seconds before sending the first keepalive probe. Only relevant if FI_SOCKETS_KEEPALIVE_ENABLE is enabled.
An integer to specify the time in seconds between individual keepalive probes. Only relevant if FI_SOCKETS_KEEPALIVE_ENABLE is enabled.
An integer to specify the maximum number of keepalive probes sent before dropping the connection. Only relevant if FI_SOCKETS_KEEPALIVE_ENABLE is enabled.
The prefix or the name of the network interface (default: any)

For large scale runs one can use these environment variables to set the default parameters e.g. size of the address vector(AV), completion queue (CQ), connection map etc. that satisfies the requirement of the particular benchmark. The recommended parameters for large scale runs are FI_SOCKETS_MAX_CONN_RETRY, FI_SOCKETS_DEF_CONN_MAP_SZ, FI_SOCKETS_DEF_AV_SZ, FI_SOCKETS_DEF_CQ_SZ, FI_SOCKETS_DEF_EQ_SZ.

fabric(7), fi_provider(7), fi_getinfo(3)

OpenFabrics.

2019-05-30 Libfabric Programmer's Manual