[ofiwg] 5/13 OFIWG notes

[Ingerson, Alexia]

05/13/2025
Participants:
Alexia Ingerson (Intel)
Jianxin Xiong (Intel)
Charles Shereda (Cornelis)
Juee Desai (Intel)
Ken Raffenetti (ANL)
Peinan Zhang (Intel)
Sai Sunku (AWS)
Steve Welch (HPE)

Summary:
Reminder to add any changes for 2.2.0 release - GA 6/15
Discussion regarding updates to PR #10965 adding fi_rpc API. Some small API updates were discussed as well as added proof of concept implementation in rxm and an rpc test to validate the feature.

Notes:
2.2.0 release reminder

  *   6/1 RC1
  *   6/2 RC1
  *   6/15 GA
fi_rpc: API update (PR #10965)

  *   Review of fi_rpc (summary below, see previous notes for in depth details)
     *   Sender sends rpc request, receiver sends response
     *   New ops, new format, new capability
  *   Updates:
     *   Fi_cq_rpc_entry - added timeout as union to data but that limits rpc - can't be used with CQ data. May be reverted.
  *   Added implemenation in rxm provider
     *   fi_rpc:
        *   rxm_get_conn
        *   rpc_tag=rxm_next_rpc_tag()->atomic inc
        *   util_srx_generic_trecv_no_lock: buf = resp_buf, tag = rpc_tag, flags |= FI_RPC
        *   rxm_send_common: buf = req_buf, tag = rpc_tag, data = timeout, flags |= FI_RPC, op = ofi_op_msg
        *   cancel recv if send fails
     *   fi_rpc_resp:
        *   rxm_get_conn
        *   rxm_send_common: tag = rcp_id, flags | =FI_RPC, op = ofi_op_tagged
     *   fi_rpc_discard:
        *   no op
     *   completions:
        *   FI_RPC flags is kept, tag = rpc_id, data = timeout
     *   Stale response
        *   Rxm_handle_recv_comp/get_tag - ENOENT flag FI_RPC, drop stale response
     *   Limitations:
        *   Currently tagged flow is not separate from RPC flow, not advised to mix usage of tag and FI_RPC. In the future can add separate queues
        *   Timeout not handled
  *   New test fi_rdm_rpc:
     *   Client sends 10 RPC requests using fi_rpc(), wait for RPC reponses and check for FI_RPC and FI_RECV flags
     *   Server posts 10 receives, wait for completions and check for FI_RPC and FI_RECV flags. Send RPC responses in random order using fi_rpc_resp

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ofiwg/attachments/20250514/ebea99b1/attachment.htm>