fi_efa(7) Libfabric Programmer's Manual
fi_efa - The Amazon Elastic Fabric Adapter (EFA) Provider
The EFA provider supports the Elastic Fabric Adapter (EFA) device on Amazon EC2. EFA provides reliable and unreliable datagram send/receive with direct hardware access from userspace (OS bypass).
The following features are supported:
- Endpoint types
- The provider supports endpoint type FI_EP_DGRAM, and FI_EP_RDM on a new Scalable (unordered) Reliable Datagram protocol (SRD). SRD provides support for reliable datagrams and more complete error handling than typically seen with other Reliable Datagram (RD) implementations. The EFA provider provides segmentation, reassembly of out-of-order packets to provide send-after-send ordering guarantees to applications via its FI_EP_RDM endpoint.
- RDM Endpoint capabilities
- The following data transfer interfaces are supported via the FI_EP_RDM endpoint: FI_MSG, FI_TAGGED, and FI_RMA. FI_SEND, FI_RECV, FI_DIRECTED_RECV, FI_MULTI_RECV, and FI_SOURCE capabilities are supported. The endpoint provides send-after-send guarantees for data operations. The FI_EP_RDM endpoint does not have a maximum message size.
- DGRAM Endpoint capabilities
- The DGRAM endpoint only supports FI_MSG capability with a maximum message size of the MTU of the underlying hardware (approximately 8 KiB).
- Address vectors
- The provider supports FI_AV_TABLE and FI_AV_MAP address vector types. FI_EVENT is unsupported.
- Completion events
- The provider supports FI_CQ_FORMAT_CONTEXT, FI_CQ_FORMAT_MSG, and FI_CQ_FORMAT_DATA. FI_CQ_FORMAT_TAGGED is supported on the RDM endpoint. Wait objects are not currently supported.
- The provider requires the use of FI_MSG_PREFIX when running over the DGRAM endpoint, and requires FI_MR_LOCAL for all memory registrations on the DGRAM endpoint.
- Memory registration modes
- The RDM endpoint does not require memory registration for send and receive operations, i.e. it does not require FI_MR_LOCAL. Applications may specify FI_MR_LOCAL in the MR mode flags in order to use descriptors provided by the application. The FI_EP_DGRAM endpoint only supports FI_MR_LOCAL.
- RDM and DGRAM endpoints support FI_PROGRESS_MANUAL. EFA erroneously claims the support for FI_PROGRESS_AUTO, despite not properly supporting automatic progress. Unfortunately, some Libfabric consumers also ask for FI_PROGRESS_AUTO when they only require FI_PROGRESS_MANUAL, and fixing this bug would break those applications. This will be fixed in a future version of the EFA provider by adding proper support for FI_PROGRESS_AUTO.
- The RDM endpoint supports FI_THREAD_SAFE, the DGRAM endpoint supports FI_THREAD_DOMAIN, i.e. the provider is not thread safe when using the DGRAM endpoint.
The DGRAM endpoint does not support FI_ATOMIC interfaces. For RMA operations, completion events for RMA targets (FI_RMA_EVENT) is not supported. The DGRAM endpoint does not fully protect against resource overruns, so resource management is disabled for this endpoint (FI_RM_DISABLED).
No support for selective completions.
No support for counters for the DGRAM endpoint.
No support for inject.
When using FI_HMEM for either CUDA and Neuron buffers, the provider requires peer to peer transaction support between the EFA and the FI_HMEM device. Therefore, the FI_HMEM_P2P_DISABLED option is not supported by the EFA provider.
PROVIDER SPECIFIC ENDPOINT LEVEL OPTION
- Defines the number of RNR retry. The application can use it to reset RNR retry counter via the call to fi_setopt. Note that this option must be set before the endpoint is enabled. Otherwise, the call will fail. Also note that this option only applies to RDM endpoint.
- Maximum number of transmit operations before the provider returns -FI_EAGAIN. For only the RDM endpoint, this parameter will cause transmit operations to be queued when this value is set higher than the default and the transmit queue is full.
- Maximum number of receive operations before the provider returns -FI_EAGAIN.
- Maximum number of IOVs for a transmit operation.
- Maximum number of IOVs for a receive operation.
RUNTIME PARAMETERS SPECIFIC TO RDM ENDPOINT
These OFI runtime parameters apply only to the RDM endpoint.
- Maximum number of MTU-sized messages that can be in flight from any single endpoint as part of long message data transfer.
- Depth of transmit queue opened with the NIC. This may not be set to a value greater than what the NIC supports.
- Size of out of order reorder buffer (in messages). Messages received out of this window will result in an error.
- Size of any cq created, in number of entries.
- Enables using the mr cache and in-line registration instead of a bounce buffer for iov’s larger than max_memcpy_size. Defaults to true. When disabled, only uses a bounce buffer
- Sets the maximum number of memory registrations that can be cached at any time.
- Sets the maximum amount of memory that cached memory registrations can hold onto at any time.
- Threshold size switch between using memory copy into a pre-registered bounce buffer and memory registration on the user buffer.
- Overrides the default MTU size of the device.
- Enables the use of a separate pool of bounce-buffers to copy unexpected messages out of the pre-posted receive buffers.
- Enables the use of a separate pool of bounce-buffers to copy out-of-order RTS packets out of the pre-posted receive buffers.
- Maximum timeout (us) for backoff to a peer after a receiver not ready error.
- Time interval (us) for the base timeout to use for exponential backoff to a peer after a receiver not ready error.
- Enable SHM provider to provide the communication across all intra-node processes.
SHM transfer will be disabled in the case where
ptrace protectionis turned on. You can turn it off to enable shm transfer.
- Defines the maximum number of entries in SHM provider’s address vector.
- Defines the switch point between small/medium message and large message. The message larger than this switch point will be transferred with large message protocol. NOTE: This parameter is now deprecated.
- The maximum size for inter EFA messages to be sent by using medium message protocol. Messages which can fit in one packet will be sent as eager message. Messages whose sizes are smaller than this value will be sent using medium message protocol. Other messages will be sent using CTS based long message protocol.
- Enable fork() support. This may have a small performance impact and should only be set when required. Applications that require to register regions backed by huge pages and also require fork support are not supported.
- The maximum number of bytes that will be eagerly sent by inflight messages uses runting read message protocol (Default 307200).
- Set CU_POINTER_ATTRIBUTE_SYNC_MEMOPS for cuda ptr. (Default: 1)