Features:
- Added stream-based communication API to UCP
- Added support for GPU platforms: Nvidia CUDA and AMD ROCM software stacks
- Added API for client/server based connection establishment
- Added support for TCP transport
- Support for InfiniBand tag-matching offload for DC and accelerated transports
- Multi-rail support for eager and rendezvous protocols
- Added support for tag-matching communications with CUDA buffers
- Added ucp_rkey_ptr() to obtain pointer for shared memory region
- Avoid progress overhead on unused transports
- Improved scalability of software tag-matching by using a hash table
- Added transparent huge-pages allocator
- Added non-blocking flush and disconnect for UCP
- Support fixed-address memory allocation via ucp_mem_map()
- Added ucp_tag_send_nbr() API to avoid send request allocation
- Support global addressing in all IB transports
- Add support for external epoll fd and edge-triggered events
- Added registration cache for knem
- Initial support for Java bindings
Bugfixes:
- Multiple bugfixes (full list on githib)
Bugfixes since RC1:
- Fix flow control for DC transport
- Fix compilation issue with mlx5 on ARM
- Disable GDR-copy when ODP is used
- Fixes for gcc8 compilation
- Fix missing initialization of rndv_send_nbr thresholds
- Fix mlx5 srq cleanup
- Fix ep info print when there is no wireup lane
- Optimize ugni locking
Tested configurations:
- InfiniBand: MLNX_OFED 4.2, inbox OFED drivers.
- CUDA: gdrcopy 1.2, cuda 9.1.85
- XPMEM: 2.6.2
- KNEM: 1.1.2
Known issues:
#2047 - UCP: ucp_do_am_bcopy_multi drops data on UCS_ERROR_NO_RESOURCE
#2047 - failure in ud/uct_flush_test.am_zcopy_flush_ep_nb/1
#1977 - failure in shm/test_ucp_rma.blocking_small/0
#1926 - Timeout in mpi_test_suite with HW TM