v1.10.0-rc2

1.10.0-rc2 (February 2, 2021)

Features:

Core

• Added support for Nvidia HPC SDK
• Added support for latest PGI and Clang
• Added support for ROCM-3.7+ (warning generated if older version detected)

Architecture

• Added Arm SVE memcpy()
• Redesigned Arm WFE support
• Improved clear_cache performance for Arm
• Added architecture detection for Zhaoxin CPU

CI

• Added release builds on CUDA 11
• Enabled performance validation in gtest

UCP

• Added locality awareness to the transport selection logic for GPU devices
• Added put/offload/short and put/offload/zcopy protocols
• Added receive message nbx routine
• Reworked AM implementation and API, which adds support for RNDV semantics
• Added support for multi-lane connection manager over TCP
• Added support for printing AM tls with info log level
• Implement flush and destroy for UCT EPs on UCP worker
• Reduced UCP request size
• Added support for keepalive protocol
• Added support for multi-fragment protocol
• Added implementation for protocol progress for eager, bcopy, and multicopy
• Improved selection logic for protocol selection
• Added new protocols for UCP get operation
• Added bcopy protocols with support for GPU memory
• Added RNDV protocol implementation for GPU devices (CUDA, ROCm)
• Set SOCKADDR_CM_ENABLE=y by default
• Added support for fast-path short with new tag protocols
• Added a new parameter to control the CM listener's backlog
• Added support sending AM RTS over short message protocol
• Added support for shared memory multi-lane when CM is used

UCT

• Added API for keepalive_timeout value
• Added add uct_completion.status
• Allowed transports to access multiple mem_types
• Removed status arg from uct_completion_callback_t
• Restructured uct_mem_alloc/uct_md_mem_alloc to use mem_type
• Updated documentation for uct_listener_params
• Lowered the log level for certain network errors
• Added cuda_copy wakeup feature
• Added wakeup support for shared memory

UCS

• Added "inf" and "auto" values to time units
• Added on-stack constructors for array and string buffer
• Added ucs_ptr_map_t data structure
• Added bool CSWAP
• Improved logging
• Added optimization for namespace processing
• Fixes for connection matching functionality

RDMA CORE (IB, ROCE, etc.)

• Added support for auto detection of adapative routing settings
• Added an option to poll TX CQ every progress iteration
• Added local and remote addresses to the reject error message
• Added support for UAR allocation with non-cacheable memory type
• Added support for multiple flush cancel without completion
• Added async events callback support
• Added detection for ConnectX-6, ConnectX-7 and BlueField-1/2 devices
• Added support for connection matching for UD
• Added a check for AM ordering

Java (preview)

• Added support for a different javadoc executable path for different java versions
• Added UCS memory type constants
• Added support build on Java10+
• Added support for io-vector datatype.

Tests

• Added CI for CUDA 11
• Added test_ucp_sockaddr_protocols.stream_short
• Reimplemented tests using NBX API
• Added flush(cancel) test
• Added memory_wait mode to perftest
• Added support for clang 10
• Refactored RMA and atomic tests, add memtype support
• Added test for uct_md_mem_query()
• Added request interrupt support
• Added support for connection manager fallbacks
• Added new ucp request test checking for leaks from the ptr_map

Documentation

• Added glossaries

Bugfixes:

Portability

• Fixes in print functions to use format string like PRIx64, etc.

Continues Integration:

• Fixes in Github release flow
• Fixes in docker image

Packaging

• Removed deb package dependencies
• Fixes in SPEC to make the RPM relocatable

Documentation

• Fixes in documentation for ucp_am_recv_data_nbx
• Fixes in quick start example
• Fixes in installation instruction

Tests

• Fixes for failures under valgrind runtime
• Fixes in mmap tests for 0-length RMA
• Fixes in definition of LAST_WQE wait timeout
• Fixes in ROCm for mem_buffer test
• Fixes in test name printing format
• Fixes in tcp_sockcm test

UCP

• Fixes in worker cleanup flow

CUDA

• Fixes in managed memory support

RDMA CORE (IB, ROCE, etc.)

• Fixes in assert definitions
• Fixes in printing an error about invalid AM Bcopy length for UD
• Fixes for thread safety support
• Fixes to get ROCE device name according to GID
• Fixes for SL selection
• Fixes in create STRICT_ORDER key
• Fixes addressing performance degradation in UD transport due to excess async events

UGNI

• Fixing disable logic in config
• Fixing clang 11 warnings

Java

• Fixes in build dependencies
• Fixes in constructing UcpRequest object on error
• Fixes in exception handling on endpoint closure request
• Fixes for segfault in UcpErrorHandler

UCP

• Fixes in datatype support for get_zcopy RNDV
• Fixes in connection manager disconnect
• Fixes in assert definitions
• Fixes in completion flow for failed EP
• Fixes in flush error handling flow
• Fixes in latency calculations for wireup protocol
• Fixes in offload completion with inlined data
• Fixes in unpacking flow
• Fixes in error handling for various protocols

UCT

• Fixes in flush TX
• Fixes in checks for enabling GPU Direct RDMA

UCS

• Fixes for crashes on incorrect value set in config
• Fixes in ptr_array
• Fixes in maximal size for ucs_snprintf_safe()
• Fixes in compilation warning
• Fixes in ucs_aarch64_dsb(_op) definition

TCP

• Fixes in default route interface confirmation flow
• Fixes in PUT protocol
• Fixes in max connection limit and improved error reporting

UCM

• Fixing crash on prevent unload
• Fixes in libucm_rocm
• Fixes for few racing conditions