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

fi_mrail - The Multi-Rail Utility Provider

The mrail provider (ofi_mrail) is an utility provider that layers over an underlying provider to enable the use of multiple network ports (rails). This increases the total available bandwidth of an underlying provider. The current status of mrail provider is experimental - not all libfabric features are supported and performance is not guaranteed.

mrail provider requires the underlying provider to support the following capabilities / modes:

Buffered receive (FI_BUFFERED_RECV)
FI_SOURCE
FI_AV_TABLE

Applications need to: * Support FI_MR_RAW MR mode bit to make use of FI_RMA capability. * Set FI_OFI_MRAIL_ADDR env variable (see RUNTIME PARAMETERS section below).

The provider supports only FI_EP_RDM.
The following data transfer interface is supported: FI_MSG, FI_TAGGED, FI_RMA.
# LIMITATIONS
Limitations of the underlying provider may show up as that of mrail provider.

mrail provider doesn't allow pass-through of any mode bits to the underlying provider.

The following are the major libfabric features that are not supported. Any other feature not listed in "Supported features" can be assumed as unsupported.

FI_ATOMIC
Scalable endpoints
Shared contexts
FABRIC_DIRECT
Multicast
Triggered operations

For messages (FI_MSG, FI_TAGGED), the provider uses different policies to send messages over one or more rails based on message size (See FI_OFI_MRIAL_CONFIG in the RUNTIME PARAMETERS section). Ordering is guaranteed through the use of sequence numbers.

For RMA, the data is striped equally across all rails.

The ofi_mrail provider checks for the following environment variables.

Comma delimited list of individual rail addresses. Each address can be an address in FI_ADDR_STR format, a host name, an IP address, or a netdev interface name.
Deprecated. Replaced by FI_OFI_MRAIL_ADDR.
Comma separated list of <max_size>:<policy> pairs, sorted in ascending order of <max_size>. Each pair indicated the rail sharing policy to be used for messages up to the size <max_size> and not covered by all previous pairs. The value of <policy> can be fixed (a fixed rail is used), round-robin (one rail per message, selected in round-robin fashion), or striping (striping across all the rails). The default configuration is 16384:fixed,ULONG_MAX:striping. The value ULONG_MAX can be input as -1.

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

OpenFabrics.

2020-04-14 Libfabric Programmer's Manual