Release Notes v2.3.0
Summary
This version brings bug fixes and updates to our v2.0.0 release.
New features
- [HG/NA]
- Add
HG_Init_opt2()
/HG_Core_init_opt2()
/NA_Initialize_opt2()
to safely pass updated init info while maintaining ABI compatibility between versions - Add
HG_Get_na_protocol_info()
/HG_Free_na_protocol_info()
and addhg_info
utility for basic listing of protocols
- Add
- [HG]
- Add support for multi-recv operations (OFI plugin only)
- Currently disable multi-recv when auto SM is on
- Posted recv operations are in that case decoupled from pool of RPC handles
- Add
release_input_early
init info flag to attempt to release input buffers early once input is decoded - Add
HG_Release_input_buf()
to manually release input buffer. - Add also
no_multi_recv
init info option to force disabling multi-recv
- Make use of subsys logs (
cls
,ctx
,addr
,rpc
,poll
) to control log output - Add init info struct versioning
- Add
HG_Context_unpost()
/HG_Core_context_unpost()
for optional 2-step context shutdown
- Add support for multi-recv operations (OFI plugin only)
- [HG bulk]
- Update to new logging system through
bulk
subsys log.
- Update to new logging system through
- [HG proc]
- Update to new logging system through
proc
subsys log.
- Update to new logging system through
- [HG Test]
- Refactor tests to separate perf tests from unit tests
- Add NA/HG test common library
- Add
hg_rate
/hg_bw_write
andhg_bw_read
perf tests- Perf test now supports multi-client / multi-server workloads
- Add
BUILD_TESTING_UNIT
andBUILD_TESTING_PERF
CMake options
- [NA]
- Add support for multi-recv operations
- Add
NA_Msg_multi_recv_unexpected()
andna_cb_info_multi_recv_unexpected
cb info - Add
flags
parameter toNA_Op_create()
andNA_Msg_buf_alloc()
- Add
NA_Has_opt_feature()
to query multi recv capability
- Add
- Remove
int
return type from NA callbacks and returnvoid
- Remove unused
timeout
parameter fromNA_Trigger()
NA_Addr_free()
/NA_Mem_handle_free()
andNA_Op_destroy()
now returnvoid
na_mem_handle_t
andna_addr_t
types no longer include pointer type- Add support for dynamically loaded plugins
- Add
NA_PLUGIN_PATH
env variable to optionally control plugin loading path (default isNA_INSTALL_PLUGIN_DIR
) - Add
NA_INSTALL_PLUGIN_DIR
variable to control plugin install path (default is lib install path) - Add
NA_USE_DYNAMIC_PLUGINS
CMake option (OFF by default)
- Add
- Add ability to query protocol info from plugins
- Add
NA_Get_protocol_info()
/NA_Free_protocol_info()
API routines - Add
na_protocol_info
struct to na_types
- Add
- Bump NA library version to 4.0.0
- Add support for multi-recv operations
- [NA OFI]
- Add support for multi-recv operations and use
FI_MSG
- Allocate multi-recv buffers using hugepages when available
- Switch to using
fi_senddata()
with immediate data for unexpected msgsNA_OFI_UNEXPECTED_TAG_MSG
can be set to switch back to former behavior that uses tagged messages instead
- Remove support for deprecated
psm
provider - Control CQ interrupt signaling with
FI_AFFINITY
(only used if thread is bound to a single CPU ID) - Enable
cxi
provider to useFI_WAIT_FD
- Add
NA_OFI_OP_RETRY_TIMEOUT
andNA_OFI_OP_RETRY_PERIOD
- Once
NA_OFI_OP_RETRY_TIMEOUT
milliseconds elapse, retry is stopped and operation is aborted (default is 120000ms) - When
NA_OFI_OP_RETRY_PERIOD
is set, operations are retried only everyNA_OFI_OP_RETRY_PERIOD
milliseconds (default is 0)
- Once
- Add support for
tcp
with and withoutofi_rxm
tcp
defaults totcp;ofi_rxm
for libfabric < 1.18
- Enable plugin to be built as a dynamic plugin
- Add support for
get_protocol_info
to query list of protocols - Add support for libfabric log redirection
- Requires libfabric >= 1.16.0, disabled if FI_LOG_LEVEL is set
- Add
libfabric
log subsys (off by default) - Bump FI_VERSION to 1.13 when log redirection is supported
- Add support for multi-recv operations and use
- [NA UCX]
- Attempt to disable UCX backtrace if
UCX_HANDLE_ERRORS
is not set - Add support for
UCP_EP_PARAM_FIELD_LOCAL_SOCK_ADDR
- With UCX >= 1.13 local src address information can now be specified on client to use specific interface and port
- Set
CM_REUSEADDR
by default to enable reuse of existing listener addr after a listener exits abnormally - Attempt to reconnect EP if disconnected
- This concerns cases where a peer would have reappeared after a previous disconnection
- Add support for
get_protocol_info
to query list of protocols - Enable plugin to be built as a dynamic plugin
- Attempt to disable UCX backtrace if
- [NA Test]
- Update NA test perf to use multi-recv feature
- Update perf test to use hugepages
- Add support for multi-targets and add lookup test
- Install perf tests if
BUILD_TESTING_PERF
isON
- [HG util]
- Change return type of
hg_time_less()
tobool
- Add
HG_LOG_WRITE_FUNC()
macro to pass func/line info- Add also
module
/no_return
parameters to hg_log_write()
- Add also
- Add support for hugepage allocations
- Use
isb
forcpu_spinwait
onaarch64
- Add
mercury_dl
to support dynamically loaded modules - Bump HG util version to 4.0.0
- Change return type of
Bug fixes
- [HG]
- Ensure init info version is compatible with previous versions of the struct
- Clean up and refactoring fixes
- Fix race condition in
hg_core_forward
with debug enabled - Simplify RPC map and fix hashing for RPC IDs larger than 32-bit integer
- Refactor context pools and cleanup
- Fix potential leak on ack buffer
- Ensure list of created RPC handles is empty before closing context
- Bump default number of pre-allocated requests from 256 to 512 to make use of 2M hugepages by default
- Add extra error checking to prevent class mismatch
- Fix potential race when sending one-way RPCs to ourself
- [HG Bulk]
- Add extra error checking to prevent class mismatch
- [HG Test]
- Refactor
test_rpc
to correctly handle timeout return values - Fix overflow of number of target / classes
- Number of targets was limited to
UINT8_MAX
- Number of targets was limited to
- Refactor
- [NA OFI]
- Fix handling of extra caps to not always follow advertised caps
- Ensure also that extra caps passed are honored by provider
- Force
sockets
provider to use shared domains- This prevents a performance regression when multiple classes are
being used (
FI_THREAD_DOMAIN
is therefore disabled for this provider)
- This prevents a performance regression when multiple classes are
being used (
- Refactor unexpected and expected sends, retry of OFI operations, handling of RMA operations
- Always include
FI_DIRECTED_RECV
in primary caps - Disable use of
FI_SOURCE
for most providers to reduce lookup overhead- Separate code paths for providers that do not support
FI_SOURCE
- Remove insert of FI addr into secondary table if
FI_SOURCE
is not used
- Separate code paths for providers that do not support
- Remove
NA_OFI_SOURCE_MSG
flag that was matchingFI_SOURCE_ERR
- Fix potential refcount race when sharing domains
- Check domain's optimal MR count if non-zero
- Fix potential double free of src_addr info
- Refactor auth key parsing code to build without extension headers
- Merge latest changes required for
opx
provider enablement- Pass
FI_COMPLETION
to RMA ops as flag is currently not ignored (prov/opx
tmp fix)
- Pass
- Add runtime version check
- Ensure that runtime version is greater than min version
- Fix handling of extra caps to not always follow advertised caps
- [NA SM]
- Fix handling of 0-size messages when no receive has been posted
- Fix issue where an expected msg that is no longer posted arrives
- In that particular case just drop the incoming msg
- Add perf warning message for unexpected messages without recv posted
- [NA UCX]
- Fix handling of UCS return types to match NA types
- Enforce src_addr port used for connections to be 0
- This fixes a port conflict between listener and connection ports
- Fix handling of unexpected messages without pre-posted recv
- [NA BMI]
- Clean up and fix some coverity warnings
- [NA MPI]
- Clean up and fix some coverity warnings
- [NA Test]
- Fix NA latency test to ensure recvs are always pre-posted
- Do not use MPI_Init_thread() if not needed
- Fix missing return check of na_test_mpi_init()
- [HG util]
- Clean up logging and set log root to
hg_all
hg_all
subsys can now be set to turn on logging in all subsystems
- Set log subsys to
hg_all
if log level env is set - Fixes to support WIN32 builds
- Clean up logging and set log root to
- [CMake]
- Fix internal/external dependencies that were not correctly set
- Fix pkg-config entries wrongly set as public/private
- Ensure
VERSION
/SOVERSION
is not set onMODULE
libraries - Allow for in-source builds (RPM support)
- Add
DL
lib dependency - Fix object target linking on CMake < 3.12
- Ensure we build with PIC and PIE when available
- [Examples]
- Allow examples to build without Boost support
Known Issues
- [NA OFI]
- [tcp/verbs;ofi_rxm] Using more than 256 peers requires
FI_UNIVERSE_SIZE
to be set.
- [tcp/verbs;ofi_rxm] Using more than 256 peers requires
Last update:
June 15, 2023