@@ -29,7 +29,33 @@ v1.13.0, Fri Jul 2, 2021
2929
3030## EFA
3131
32- - TODO
32+ - Minor code optimizations and bug fixed
33+ - Add support for fi_inject for RDM messages
34+ - Improve handling of RNR NACKs from NIC
35+ - Improve handling of zero copy receive case, especially when sender does not
36+ post receive buffer
37+ - Numerous RMA read bug fixes
38+ - Add unexpected receive queue for each peer
39+ - Fixed issue releasing rx entries
40+ - Decrease the initial size of the out-of-order packeet pool allocation size
41+ to reduce the common-case memory footprint
42+ - Handle FI_ADDR_NOTAVAIL in rxr_ep_get_peer
43+ - Identify and handle QP reuse
44+ - Use the memory monitor specified by the user
45+ - Replace provider code with common code in select places
46+ - Update efa_av_lookup to return correct address
47+ - Update rdm endpoint directly poll cq from ibv_cq
48+ - Avoid possible duplicate completions
49+ - Add reference counting for peer tracking
50+ - Fix EFA usage of util AV causing incorrect refcounting
51+ - Do not allow endpoints to share address vectors
52+ - Improve fork support; users can set the FI_EFA_FORK_SAFE environment variable
53+ for applications which call fork()
54+ - Adjust the timing of clearing deferred memory registration list
55+ - Do not use eager protocol for cuda message and local peer
56+ - Fixes for shm support
57+ - Enable MR cache for CUDA
58+ - Disable shm when application requests FI_HMEM
3359
3460## PSM3
3561
0 commit comments