fi_psm3(7) Libfabric Programmer's Manual
fi_psm3 - The PSM3 Fabric Provider
The psm3 provider implements a Performance Scaled Messaging capability
which supports most verbs UD and sockets devices. Additional features and
optimizations can be enabled when running over Intel’s E810 Ethernet NICs
and/or using Intel’s rendezvous kernel module (
PSM 3.x fully integrates the OFI provider and the underlying PSM3
protocols/implementation and only exports the OFI APIs.
The psm3 provider supports a subset of all the features defined in the libfabric API.
- Endpoint types
- Supports non-connection based types FI_DGRAM and FI_RDM.
- Endpoint capabilities
- Endpoints can support any combination of data transfer capabilities
FI_TAGGED, FI_MSG, FI_ATOMICS, and FI_RMA. These capabilities
can be further refined by FI_SEND, FI_RECV, FI_READ, FI_WRITE,
FI_REMOTE_READ, and FI_REMOTE_WRITE to limit the direction of
FI_MULTI_RECV is supported for non-tagged message queue only.
Scalable endpoints are supported if the underlying PSM3 library supports multiple endpoints. This condition must be satisfied both when the provider is built and when the provider is used. See the Scalable endpoints section for more information.
Other supported capabilities include FI_TRIGGER, FI_REMOTE_CQ_DATA, FI_RMA_EVENT, FI_SOURCE, and FI_SOURCE_ERR. Furthermore, FI_NAMED_RX_CTX is supported when scalable endpoints are enabled.
- FI_CONTEXT is required for the FI_TAGGED and FI_MSG capabilities. That means, any request belonging to these two categories that generates a completion must pass as the operation context a valid pointer to type struct fi_context, and the space referenced by the pointer must remain untouched until the request has completed. If none of FI_TAGGED and FI_MSG is asked for, the FI_CONTEXT mode is not required.
- The psm3 provider performs optimal with manual progress. By default, the application is expected to call fi_cq_read or fi_cntr_read function from time to time when no other libfabric function is called to ensure progress is made in a timely manner. The provider does support auto progress mode. However, the performance can be significantly impacted if the application purely depends on the provider to make auto progress.
- Scalable endpoints
- Scalable endpoints support depends on the multi-EP feature of the PSM3
library. If the PSM3 library supports this feature, the availability is
further controlled by an environment variable PSM3_MULTI_EP. The psm3
provider automatically sets this variable to 1 if it is not set. The
feature can be disabled explicitly by setting PSM3_MULTI_EP to 0.
When creating a scalable endpoint, the exact number of contexts requested should be set in the “fi_info” structure passed to the fi_scalable_ep function. This number should be set in “fi_info->ep_attr->tx_ctx_cnt” or “fi_info->ep_attr->rx_ctx_cnt” or both, whichever greater is used. The psm3 provider allocates all requested contexts upfront when the scalable endpoint is created. The same context is used for both Tx and Rx.
For optimal performance, it is advised to avoid having multiple threads accessing the same context, either directly by posting send/recv/read/write request, or indirectly by polling associated completion queues or counters.
Using the scalable endpoint as a whole in communication functions is not supported. Instead, individual tx context or rx context of the scalable endpoint should be used. Similarly, using the address of the scalable endpoint as the source address or destination address doesn’t collectively address all the tx/rx contexts. It addresses only the first tx/rx context, instead.
The psm3 provider doesn’t support all the features defined in the libfabric API. Here are some of the limitations not listed above:
- Unsupported features
- These features are unsupported: connection management, passive endpoint, and shared receive context.
The psm3 provider checks for the following environment variables:
- PSM requires that each job has a unique ID (UUID). All the processes
in the same job need to use the same UUID in order to be able to
talk to each other. The PSM reference manual advises to keep UUID
unique to each job. In practice, it generally works fine to reuse
UUID as long as (1) no two jobs with the same UUID are running at
the same time; and (2) previous jobs with the same UUID have exited
normally. If running into “resource busy” or “connection failure”
issues with unknown reason, it is advisable to manually set the UUID
to a value different from the default.
The default UUID is 00FF00FF-0000-0000-0000-00FF0F0F00FF.
It is possible to create endpoints with UUID different from the one set here. To achieve that, set ‘info->ep_attr->auth_key’ to the uuid value and ‘info->ep_attr->auth_key_size’ to its size (16 bytes) when calling fi_endpoint() or fi_scalable_ep(). It is still true that an endpoint can only communicate with endpoints with the same UUID.
- The psm3 provider has a simple built-in name server that can be used
to resolve an IP address or host name into a transport address needed
by the fi_av_insert call. The main purpose of this name server is to
allow simple client-server type applications (such as those in fabtests)
to be written purely with libfabric, without using any out-of-band
communication mechanism. For such applications, the server would run first
to allow endpoints be created and registered with the name server, and
then the client would call fi_getinfo with the node parameter set to
the IP address or host name of the server. The resulting fi_info
structure would have the transport address of the endpoint created by the
server in the dest_addr field. Optionally the service parameter can
be used in addition to node. Notice that the service number is
interpreted by the provider and is not a TCP/IP port number.
The name server is on by default. It can be turned off by setting the variable to 0. This may save a small amount of resource since a separate thread is created when the name server is on.
The provider detects OpenMPI and MPICH runs and changes the default setting to off.
- The RMA functions are implemented on top of the PSM Active Message functions.
The Active Message functions have limit on the size of data can be transferred
in a single message. Large transfers can be divided into small chunks and
be pipe-lined. However, the bandwidth is sub-optimal by doing this way.
The psm3 provider use PSM tag-matching message queue functions to achieve higher bandwidth for large size RMA. It takes advantage of the extra tag bits available in PSM3 to separate the RMA traffic from the regular tagged message queue.
The option is on by default. To turn it off set the variable to 0.
- Time (seconds) to sleep before closing PSM endpoints. This is a workaround
for a bug in some versions of PSM library.
The default setting is 0.
- Timeout (seconds) for gracefully closing PSM endpoints. A forced closing
will be issued if timeout expires.
The default setting is 5.
- Timeout (seconds) for establishing connection between two PSM endpoints.
The default setting is 5.
- When auto progress is enabled (asked via the hints to fi_getinfo),
a progress thread is created to make progress calls from time to time.
This option set the interval (microseconds) between progress calls.
The default setting is 1 if affinity is set, or 1000 if not. See FI_PSM3_PROG_AFFINITY.
- When set, specify the set of CPU cores to set the progress thread
affinity to. The format is
<start>[:<end>[:<stride>]][,<start>[:<end>[:<stride>]]]*, where each triplet
<start>:<end>:<stride>defines a block of core_ids. Both
<end>can be either the
core_id(when >=0) or
core_id - num_cores(when <0).
By default affinity is not set.
- Maximum message size allowed for fi_inject and fi_tinject calls. This is
an experimental feature to allow some applications to override default
inject size limitation. When the inject size is larger than the default
value, some inject calls might block.
The default setting is 64.
- When set, dictate the level of locking being used by the provider. Level
2 means all locks are enabled. Level 1 disables some locks and is suitable
for runs that limit the access to each PSM3 context to a single thread.
Level 0 disables all locks and thus is only suitable for single threaded
To use level 0 or level 1, wait object and auto progress mode cannot be used because they introduce internal threads that may break the conditions needed for these levels.
The default setting is 2.
- There are two strategies on when to establish connections between the PSM3
endpoints that OFI endpoints are built on top of. In eager connection mode,
connections are established when addresses are inserted into the address
vector. In lazy connection mode, connections are established when addresses
are used the first time in communication. Eager connection mode has slightly
lower critical path overhead but lazy connection mode scales better.
This option controls how the two connection modes are used. When set to 1, lazy connection mode is always used. When set to 0, eager connection mode is used when required conditions are all met and lazy connection mode is used otherwise. The conditions for eager connection mode are: (1) multiple endpoint (and scalable endpoint) support is disabled by explicitly setting PSM3_MULTI_EP=0; and (2) the address vector type is FI_AV_MAP.
The default setting is 0.
- The provider has a mechanism to automatically send disconnection notifications
to all connected peers before the local endpoint is closed. As the response,
the peers call psm3_ep_disconnect to clean up the connection state at their
side. This allows the same PSM3 epid be used by different dynamically started
processes (clients) to communicate with the same peer (server). This mechanism,
however, introduce extra overhead to the finalization phase. For applications
that never reuse epids within the same session such overhead is unnecessary.
This option controls whether the automatic disconnection notification mechanism should be enabled. For client-server application mentioned above, the client side should set this option to 1, but the server should set it to 0.
The default setting is 0.
- Select how the 96-bit PSM3 tag bits are organized. Currently three choices are
available: tag60 means 32-4-60 partitioning for CQ data, internal protocol
flags, and application tag. tag64 means 4-28-64 partitioning for internal
protocol flags, CQ data, and application tag. auto means to choose either
tag60 or tag64 based on the hints passed to fi_getinfo – tag60 is used
if remote CQ data support is requested explicitly, either by passing non-zero value
via hints->domain_attr->cq_data_size or by including FI_REMOTE_CQ_DATA in
hints->caps, otherwise tag64 is used. If tag64 is the result of automatic
selection, fi_getinfo also returns a second instance of the provider with
The default setting is auto.
Notice that if the provider is compiled with macro PSMX3_TAG_LAYOUT defined to 1 (means tag60) or 2 (means tag64), the choice is fixed at compile time and this runtime option will be disabled.