fi_psm(7) Libfabric Programmer's Manual
The PSM Fabric Provider
The psm provider runs over the PSM 1.x interface that is currently supported by the Intel TrueScale Fabric. PSM provides tag-matching message queue functions that are optimized for MPI implementations. PSM also has limited Active Message support, which is not officially published but is quite stable and well documented in the source code (part of the OFED release). The psm provider makes use of both the tag-matching message queue functions and the Active Message functions to support a variety of libfabric data transfer APIs, including tagged message queue, message queue, RMA, and atomic operations.
The psm provider can work with the psm2-compat library, which exposes a PSM 1.x interface over the Intel Omni-Path Fabric.
The psm provider doesn’t support all the features defined in the libfabric API. Here are some of the limitations:
- Endpoint types
- Only support 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
operations. The limitation is that no two endpoints can have overlapping
receive or RMA target capabilities in any of the above categories. For
example it is fine to have two endpoints with FI_TAGGED | FI_SEND,
one endpoint with FI_TAGGED | FI_RECV, one endpoint with FI_MSG,
one endpoint with FI_RMA | FI_ATOMICS. But it is not allowed to
have two endpoints with FI_TAGGED, or two endpoints with FI_RMA.
FI_MULTI_RECV is supported for non-tagged message queue only.
Other supported capabilities include FI_TRIGGER.
- 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 psm provider requires manual progress. 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.
- Unsupported features
- These features are unsupported: connection management, scalable endpoint, passive endpoint, shared receive context, send/inject with immediate data.
The psm 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 0FFF0FFF-0000-0000-0000-0FFF0FFF0FFF.
- The psm 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 fabtest)
to be written purely with libfabric, without using any out-of-band
communication mechanism. For such applications, the server would run first,
and 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 server in the dest_addr field.
The name server won’t work properly if there are more than one processes from the same job (i.e. with the same UUID) running on the same node and acting as servers. For such scenario it is recommended to have each process getting local transport address with fi_getname and exchanging the addresses with out-of-band mechanism.
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 psm provider use PSM tag-matching message queue functions to achieve higher bandwidth for large size RMA. For this purpose, a bit is reserved from the tag space 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.
- The psm provider implements the non-tagged message queue over the PSM
tag-matching message queue. One tag bit is reserved for this purpose.
Alternatively, the non-tagged message queue can be implemented over
Active Message. This experimental feature has slightly larger latency.
This option is off by default. To turn it on set the variable to 1.
- 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 1.
- Timeout (seconds) for gracefully closing PSM endpoints. A forced closing
will be issued if timeout expires.
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 affininty is set, or 1000 if not. See FI_PSM_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.