Wiki

Clone wiki

upcxx / ChangeLog

ChangeLog

This is the ChangeLog for public releases of UPC++.

For information on using UPC++, see: README.md
For information on installing UPC++, see: INSTALL.md

2022.03.31: Release 2022.3.0

Improvements to GPU memory kinds:

This release features a number of synergistic improvements to the UPC++ memory kinds feature that supports efficient PGAS communication involving GPU memory buffers.

  • NEW: Memory kinds support for AMD GPUs using ROCm/HIP, see INSTALL.md. New configure --enable-hip flag activates new upcxx::hip_device class. This includes native offload support for upcxx::copy() using ROCmRDMA on recent InfiniBand network hardware - see GASNet-EX documentation for details.
  • cuda_device and hip_device are derived from new abstract base class gpu_device and device_allocator<Device> is now derived from new abstract base class heap_allocator. These help enable vendor-agnostic polymorphism in use of memory kinds.
  • New optional interface to GPU memory kinds simplifies startup code, e.g.:
    auto gpu_alloc = make_gpu_allocator(2UL<<20);
    creates a 2MB device segment with a "smart" choice of GPU, and:
    auto gpu_alloc = make_gpu_allocator<hip_device>(2UL<<20, 2);
    creates a device_allocator for a segment on HIP GPU number 2.
  • Several new members have been added to device_allocator to provide convenience and support the above improvements. See the specification for details.
  • These improvements are demonstrated in example/gpu_vecadd a renamed version of the cuda_vecadd kernel example which now supports either GPU vendor.

General features/enhancements: (see specification and programmer's guide for full details)

  • Experimental support for making RPC calls to functions in executable code segments other than the core UPC++ application, such as those in dynamic libraries. For more information, see docs/ccs-rpc.md.
  • Performance improvements to atomic_domain operations using shared-memory bypass.
  • New query upcxx::local_team_position() provides job topology information
  • team and atomic_domain<T> are now DefaultConstructible and have a new is_active() query
  • team, atomic_domain<T>, cuda_device and device_allocator<Device> are now MoveAssignable

Infrastructure changes:

  • NEW initial support for the HPE Cray EX platform
    • Complete and correct, but still untuned
    • Supports Slingshot-10 and Slingshot-11 NICs via GASNet-EX's experimental support for the OFI network API (aka "libfabric").
    • Supports PrgEnv-gnu and PrgEnv-cray.
    • See INSTALL.md for instructions to enable the appropriate support for this platform.
  • Memory kinds implementation internals have been factored and restructured, simplifying the addition of new memory kinds in future releases.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #288: implementation details of future lead to type errors using when_all with inference
  • issue #494: Intermittent rpc-ctor-trace failure on copy-get-d2d(as_rpc)
  • issue #512: ADL fails with when_all
  • issue #518: configure should warn or prohibit mixed-family/mixed-version compilers
  • issue #522: Runtime crash along exception path for rpc() returning non-empty operation_cx::as_future
  • issue #523: Round-trip RPC lacking operation_cx should generate an error
  • issue #527: Raise PGI version floor to 19.3
  • issue #528: cuda_device::destroy() incorrectly perturbs CUDA Driver context stack
  • issue #534: Prune unnecessary system header includes from upcxx.hpp
  • issue #537: Numerical error in kokkos-based 3d heat conduction examples
  • spec issue 173: Add upcxx::local_team_position()
  • spec issue 188: Add cuda_device::device_n()
  • spec issue 189: Add MoveAssignable to resource object types
  • spec issue 190: device_allocator constructor has several problems

Embeds a GASNet-EX library that addresses the following notable issues (see the GASNet issue tracker for details):

  • bug4211: intermittent udp-conduit exit-time hangs on macOS
  • bug4227: Bogus maybe-uninit warning building libgasnet with GCC-11.1+
  • bug4297: incorrect nbrhd construction for some multi-homed hosts
  • bug4321: Intermittent "EBADENDPOINT" failures in single-node udp-conduit
  • bug4345: Multiply defined symbols in aries-conduit w/ recent compilers
  • bug4360: Insufficient fixed exit timeouts (ucx, ibv, ofi)
  • bug4361: (partial fix) reductions on DT_USER of unbounded length
  • bug4366: intermittent exit-time assertion failures from debug memcheck

This library release conforms to the UPC++ v1.0 Specification, Revision 2022.3.0. All currently specified features are fully implemented. See the UPC++ issue tracker for status of known bugs.

Breaking changes:

  • UPC++ headers no longer have the undocumented side-effect of including <cassert>. Users are highly encouraged to use UPCXX_ASSERT() instead, which offers more features and automatically tracks UPCXX_CODEMODE. See docs/implementation-defined.md for more details.
  • UPC++ headers no longer have the undocumented side-effect of including some system headers. User programs should directly include system headers they need.
  • Library classes team, cuda_device, atomic_domain<T> and device_allocator<D> are now all final and may not be sub-classed.
  • atomic_domain<T> construction with an empty ops set is now prohibited.
  • The three-argument device_allocator constructor has been deprecated in favor of a new constructor that swaps argument order but provides equivalent functionality. The deprecated overload will be removed in an upcoming release.
  • An active device_allocator<Device> object must now be deactivated prior to destruction, via either Device::destroy() or device_allocator::destroy().
  • RPC calls across code segments must now use multi-segment CCS mode. Cases where this "magically" worked are now prohibited. See docs/ccs-rpc.md for more details.
  • The oldest-supported PGI compiler version is raised to 19.3 on all platforms.
  • bench/cuda_microbenchmark performance test renamed to bench/gpu_microbenchmark
  • Prior to this release, the configure script would permit values of CXX and CC which had different families or versions (as long as they were link-compatible). This was particularly easy to do on a Linux system if specifying a non-default CXX while retaining the default CC=gcc. Such mixed configurations are now prohibited. While there is a configure option to convert the enforcement to a warning, such configurations are officially unsupported.

2021.09.30: Release 2021.9.0

Improvements to on-node communication:

This release features a number of synergistic optimizations that streamline interprocess communication operations that are satisfied on-node using shared memory bypass. For details, see: doi:10.25344/S42C71

  • New as_eager_future(), as_defer_future(), as_eager_promise(), and as_defer_promise() calls for requesting eager or deferred notification of future and promise completions.
  • Existing as_future() and as_promise() calls now default to eager notification for improved performance.
  • New UPCXX_DEFER_COMPLETION macro for controlling whether as_future() and as_promise() request eager or deferred notification (see implementation-defined.md for details).
  • New overloads of fetching atomics that avoid overheads of non-empty futures and promises.
  • Performance improvements to contiguous RMA (rput, rget) using shared-memory bypass.
  • Performance improvements to upcxx::copy(), especially for cases amenable to shared-memory bypass optimizations and/or not involving device memory.
  • Performance improvements to global_ptr localization queries and operations, especially for smp-conduit.

General features/enhancements: (see specification and programmer's guide for full details)

  • upcxx::rpc and upcxx::rpc_ff calls that encounter shared heap exhaustion while allocating internal buffers will now throw an exception instead of crashing. For details, see implementation-defined.md
  • New team::create factory constructs teams with less communication than team::split when each participant can enumerate the membership of its own new team.
  • The following future operations are now permitted before UPC++ initialization: make_future(), to_future(), when_all(), assignment and copy/move constructors.
  • Added implementation-defined macros UPCXX_ASSERT and UPCXX_ASSERT_ALWAYS
  • New UPCXX_KIND_CUDA feature macro indicates the presence of CUDA support.
  • Improve error reporting on failure to open a cuda_device.
  • Add debug codemode checking for exceptions thrown out of user callbacks into library code, which is prohibited by the specification.
  • Notable GASNet performance improvements for InfiniBand (ibv) network.
  • bench/cuda_microbenchmark performance test expanded and improved

Infrastructure changes:

  • Multiple improvements to default network selection
    • In addition to selection at configure time, the default network can now be selected as late as the make install step.
    • In support of the behavior above, the make all and make install steps now report the default and available networks.
    • The build logic for Linux now makes an effort to perform more intelligent selection of a default network, rather than always defaulting to smp.
    • The default networks for macOS and Cray XC remain unchanged: smp and aries, respectively.
  • The build logic now diagnoses conditions in which CXX must be a MPI-compatible (wrapper) compiler rather than deferring discovery of the problem until make check, make test_install or even to user application link time.
  • The "NVIDIA HPC SDK" (or "nvhpc") compiler family is now supported on x86_64 and ppc64le hosts for version 20.9 and newer.
  • Intel oneAPI compilers v2021.1.2+ are now supported on x86_64 hosts.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #242: Lack of backpressure in RPC injection leads to shared memory-exhaustion crashes
  • issue #299: de-duplication of installed headers
  • issue #354: improve default network selection
  • issue #464: assert when calling global_ptr<T,kind>::local() on device pointers
  • issue #468: Harmless unused variable warnings on clang with -O -Wall
  • issue #473: Divide by zero in serialization when writing a sequence of objects that have empty UPCXX_SERIALIZED_{FIELDS,VALUES}
  • issue #477: copy(remote_cx::as_rpc) may invoke callback in the wrong context
  • issue #479: intermittent lpc-stress/opt failures on ARM64
  • issue #481: Consider use of __builtin_launder
  • issue #482: SEQ mode incorrectly requires master as current_persona for shared allocation
  • issue #487: Renaming unspecified internal UPCXX_ macros and identifiers
  • issue #488: Configure-time failure when mixing GCC + Intel
  • issue #490: Clang pedantic warnings on template destructors building the library
  • issue #495: failures with ibv-conduit recv thread enabled
  • issue #496: Configure mishandling quotes in compiler and flags settings
  • issue #500: Invalid teams created by split() are not handled according to spec
  • issue #502: Discontiguous job layouts now require configure --enable-discontig-ranks
  • spec issue 176: Change RPC injection to throw an exception on memory exhaustion

Embeds a GASNet-EX library that addresses the following notable issues (see the GASNet issue tracker for details):

  • bug4148: ibv/GDR completion issues with multiple communication paths
  • bug4150: ibv/GDR premature local completion of Puts from device memory
  • bug4209: ibv: improve ALC with respect to bounce buffer use
  • bug4230: ssh-spawer de-duplication logic is flawed
  • bug4263: remove AD_MY_NBRHD check on smp-conduit
  • bug4265: Collective scratch management is not thread safe
  • bug4266: gex_Coll_ReduceToAllNB is not thread safe
  • bug4292: ucx and aries can leak events from AM Long
  • bug4330: ibv conduit incorrectly implements HIDDEN_AM_CONCURRENCY_LEVEL

This library release conforms to the UPC++ v1.0 Specification, Revision 2021.9.0. All currently specified features are fully implemented. See the UPC++ issue tracker for status of known bugs.

Breaking changes:

  • Existing as_future() and as_promise() calls now default to eager notification for improved performance. Deferred notification can be requested on a per-call basis by changing as_future()/as_promise() calls to as_defer_future()/as_defer_promise(), or on a translation-unit basis by defining the UPCXX_DEFER_COMPLETION macro to 1 prior to including upcxx/upcxx.hpp.
  • Many unspecified macros and identifiers that are defined by the public headers have been renamed from a UPCXX_ prefix to UPCXXI_. This naming change reflects the fact these undocumented tokens are INTERNAL to the implementation and carry no guarantee of stability or functionality. Users are strongly advised to avoid direct reference to any such interfaces.

2021.03.31: Release 2021.3.0

General features/enhancements: (see specification and programmer's guide for full details)

  • The optimizations and features supporting CUDA GPUs initially previewed in the 2020.11.0 Memory Kinds Prototype have been hardened and incorporated into this release.
  • On platforms with NVIDIA-branded CUDA devices and Mellanox-branded InfiniBand network adapters (such as OLCF Summit), upcxx::copy() uses GPUDirect RDMA (GDR) hardware support to offload RMA operations involving GPU memory.
  • See INSTALL.md for instructions to enable UPC++ CUDA support and for a list of detailed requirements and known issues.
  • New shared_segment_{size,used} queries return snapshots of the host shared segment size and utilization.
  • Updates to GASNet's support for InfiniBand networks (ibv-conduit):
    • Significantly improved performance of both RPC and RMA operations under certain conditions
    • Measurable reduction in startup time for medium-scale and large-scale jobs with wide SMP nodes.
    • Heterogeneous multirail configurations no longer reduce the size of upcxx::local_team().

Improvements to RPC and Serialization:

  • The RPC implementation has been tuned and now incurs one less payload copy on ibv and aries networks on moderately sized RPCs. Additionally, internal protocol cross-over points have been adjusted on all networks. These changes may result in noticeable performance improvement for RPCs with a total size (including serialized arguments) under about 64kb (exact limit varies with network).
  • The default aries-conduit max AM Medium size has been doubled to ~8kb to improve performance of the RPC eager protocol. See aries-conduit README for details on the available configure/envvar knobs to control this quantity.
  • Arguments to rpc, rpc_ff and remote_cx::as_rpc are now serialized synchronously before return from the communication-injection call, regardless of asynchronous source_cx completions (which are now deprecated for rpc and rpc_ff).
  • Streamlined some overheads associated with remote_cx::as_rpc and RPC replies.

Infrastructure changes:

  • The install script, deprecated since 2020.3.0, has been removed.
  • make check (and similar) now accept comma-delimited NETWORKS settings, in addition to space-delimited.
  • The PGI C++ compiler (through version 20.4) remains fully supported. The re-branded variant of this host compiler (i.e. pgc++ or nvc++ released as NVIDIA HPC SDK 20.7 and later) is not currently supported, due to critical defects.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #25: Remove non-public symbols from top-level upcxx:: namespace
  • issue #241: Intermittent validation failures in test/copy.cpp
  • issue #245: persona-example deadlocks when --with-mpsc-queue=biglock
  • issue #276: Use C++ protection features to enforce abstraction boundaries
  • issue #382: Expose shared heap usage at runtime
  • issue #408: Cannot register multiple completions against a non-copyable results type
  • issue #421: upcxx::copy() breaks with PGI optimizer
  • issue #422: Improve configure behavior for GASNet archives lacking Bootstrap
  • issue #423: Prohibit communication using non-master personas in SEQ mode
  • issue #427: Crash after write_sequence() where serialized element size is not a multiple of alignment
  • issue #428: Regression in rpc(team,rank,..,view) overload resolution
  • issue #429: upcxx library exposes dlmalloc symbols
  • issue #430: cannot disable the default network
  • issue #432: Some upcxx::copy() cases do not discharge() properly
  • issue #440: Invalid GASNet call while deserializing a global ptr
  • issue #447: REGRESSION: bulk upcxx::rput with l-value completions
  • issue #450: upcxx::lpc callback return of rvalue reference not decayed as specified
  • issue #455: Performance bug in rput(remote_cx::as_rpc(...)) with "bare" remote_cx
  • issue #459: Move unspecified identifiers into a new upcxx::experimental namespace
  • issue #460: Implementation relies on std::result_of, which is deprecated in C++17 and removed in C++20

Embeds a GASNet-EX library that addresses the following notable issues (see the GASNet issue tracker for details):

  • bug4194: ibv: unnecessarily slow startup
  • bug4208: ibv: unfortunate multi-rail interactions with PSHM and XRC

This library release conforms to the UPC++ v1.0 Specification, Revision 2021.3.0. All currently specified features are fully implemented. See the UPC++ issue tracker for status of known bugs.

Breaking changes:

  • When compiling for the default "seq" threading mode, inter-process communication may only be initiated by the primordial thread, and now additionally requires use of the master persona. For details, see docs/implementation-defined.md.
  • Array types are now prohibited as the element-type template argument to upcxx::new_ and upcxx::new_array.
  • The following unspecified identifiers, previously in the upcxx namespace, have all been moved to the new upcxx::experimental namespace. These interfaces all remain unspecified and experimental, and they are subject to change without notice in future revisions:
    • broadcast_nontrivial, reduce_one_nontrivial, reduce_all_nontrivial
    • The non-fast op_* reduction constants (e.g. op_add)
    • os_env
    • say
    • destroy_heap and restore_heap
  • The unspecified/obsolete UPCXX_REFLECTED() macro and upcxx::wait() function have been removed.
  • Many other unspecified internal functions and members have been renamed. Applications should avoid depending on unspecified functions or members, which are subject to change without notice. This in particular includes anything in the upcxx sub-namespaces (e.g. upcxx::detail and upcxx::backend).

2020.10.30: Memory Kinds Prototype 2020.11.0

This is a prototype release of UPC++ demonstrating the new GPUDirect RDMA (GDR) native implementation of memory kinds for NVIDIA-branded CUDA devices with Mellanox-branded InfiniBand network adapters.

As a prototype, it has not been validated as widely as normal stable releases, and may include features and behaviors that are subject to change without notice. This prototype is recommended for any users who want to exercise the memory kinds feature with CUDA-enabled GPUs. All other UPC++ users are recommended to use the latest stable release.

See INSTALL.md for instructions to enable UPC++ CUDA support and for a list of caveats and known issues with the current GDR-accelerated implementation.

Recent changes to the memory kinds feature:

  • Relax the restriction that a given CUDA device ID may only be opened once per process using cuda_device.
  • Add a device_allocator::is_active() query, and fix several subtle defects with inactive devices/allocators.
  • Resource exhaustion failures that occur while allocating a device segment now throw upcxx::bad_segment_alloc, a new subclass of std::bad_alloc.
  • Debug-mode global_ptr checking for device pointers has been strengthened when using GDR-accelerated memory kinds.

Requirements changes:

  • The PGI/NVIDIA C++ compiler is not supported in this prototype release, due to a known problem with the optimizer. Users are advised to use a supported version of the Intel, GNU or LLVM/Clang C++ compiler instead. See INSTALL.md for details on supported compilers.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #221: upcxx::copy() mishandling of private memory arguments
  • issue #421: Regression with upcxx::copy(remote_cx::as_rpc)

This prototype library release conforms to the UPC++ v1.0 Specification, Revision 2020.11.0-draft. All currently specified features are fully implemented. See the UPC++ issue tracker for status of known bugs.

Breaking changes:

  • device_allocator construction is now a collective operation with user-level progress.
  • device_allocator::device_id() is now restricted to global_ptr arguments with affinity to the calling process.

2020.10.30: Release 2020.10.0

General features/enhancements: (see specification and programmer's guide for full details)

  • Added support for global_ptr<const T> with implicit conversion from global_ptr<T> to global_ptr<const T>. Communication operations now take a global_ptr<const T> where appropriate.
  • local_team() creation during upcxx::init() in multi-node runs is now more scalable in time and memory, providing noticeable improvements at large scale.
  • team::destroy() now frees GASNet-level team resources that were previously leaked.
  • when_all(...) now additionally accepts non-future arguments and implicitly promotes them to trivially ready futures (as with to_future) before performing future concatenation. The main consequence is when_all() can now replace most uses of make_future and to_future for constructing trivially ready futures, and continue to serve as the future combinator.
  • Static checking has been strengthened in a number of places, to help provide or improve compile-time diagnostics for incorrect or otherwise problematic use cases.
  • Added many precondition sanity checks in debug mode, to help users find bugs in their code when compiling with -codemode=debug (aka, upcxx -g).
  • Shared heap exhaustion in upcxx::new_(array) now throws upcxx::bad_shared_alloc (a type derived from std::bad_alloc) which provides additional diagnostics about the failure.

Improvements to RPC and Serialization:

  • Added support for serialization of reference types where the referent is Serializable.
  • RPC's which return future values experience one less heap allocation and virtual dispatch in the runtime's critical path.
  • RPC's which return future values no longer copy the underlying data prior to serialization.
  • dist_object<T>::fetch() no longer copies the remote object prior to serialization.
  • Added deserializing_iterator<T>::deserialize_into to avoid copying large objects when iterating over a view of non-TriviallySerializable elements.
  • Objects passed as lvalues to RPC will now be serialized directly from the provided object, reducing copy overhead and enabling passing of non-copyable (but movable) types.
  • Non-copyable (but movable) types can now be returned from RPC by reference

Build system changes:

  • Improved compatibility with heap analysis tools like Valgrind (at some potential performance cost) using new configure option --enable-valgrind
  • configure --enable-X is now equivalent to --with-X, similarly for --disable/--without
  • Tests run by make check and friends now enforce a 5m time limit by default

Requirements changes:

  • The minimum required gcc environment module version for PrgEnv-gnu and PrgEnv-intel on a Cray XC has risen from 6.4.0 to 7.1.0.
  • The minimum required Intel compiler version for PrgEnv-intel on a Cray XC has risen from 17.0.2 to 18.0.1.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #151: Validate requested completions against the events supported by an operation
  • issue #262: Implement view buffer lifetime extension for remote_cx::as_rpc
  • issue #288: (partial fix) future-producing calls like upcxx::make_future now return the exact type future<T>. Sole remaining exception is when_all.
  • issue #313: implement future::{result,wait}_reference
  • issue #336: Add static_assert to prohibit massive types as top-level arguments to RPC
  • issue #344: poor handling for make install prefix=relative-path
  • issue #345: configure with single-dash arguments
  • issue #346: configure --cross=cray* ignores --with-cc/--with-cxx
  • issue #355: upcxx::view<T> broken with asymmetric deserialization of T
  • issue #361: upcxx::rpc broken when mixing arguments of T&& and dist_object&
  • issue #364: Stray "-e" output on macOS and possibly elsewhere
  • issue #375: Improve error message for C array types by-value arguments to RPC
  • issue #376: warnings from GCC 10.1 in reduce.hpp for boolean reductions
  • issue #384: finalize can complete without invoking progress, leading to obscure leaks
  • issue #386: upcxx_memberof_general prohibits member designators that end with an array access
  • issue #388: deserialized_value() overflows buffer for massive static types
  • issue #389: future::result*() should assert readiness
  • issue #391: View of containers of UPCXX_SERIALIZED_FIELDS crashes in deserialization
  • issue #392: Prevent silent use of by-value communication APIs for huge types
  • issue #393: Improve type check error for l-value reference args to RPC callbacks
  • issue #400: UPCXX_SERIALIZED_VALUES() misoptimized by GCC{7,8,9} with -O2+
  • issue #402: Cannot use promise<T>::fulfill_result() when T is not MoveConstructible
  • issue #405: regression: upcxx::copy(T*,global_ptr<T>,size_t) fails to compile
  • issue #407: RPC breaks if an argument asymmetrically deserializes to a type that itself has asymmetric deserialization
  • issue #412: entry barriers deadlock when invoked inside user-level progress callbacks
  • issue #413: LPC callback that returns a reference produces a future containing a dangling reference
  • issue #419: Ensure correct/portable behavior of upcxx::initialized() in static destructors
  • spec issue 104: Provide a universal variadic factory for future
  • spec issue 158: prohibit reference types in global_ptr and upcxx_memberof_general
  • spec issue 160: Deadlocks arising from synchronous collective calls with internal progress
  • spec issue 167: dist_object<T>::fetch does not correctly handle T with asymmetric serialization
  • spec issue 169: Deprecate collective calls inside progress callbacks
  • spec issue 170: Implement upcxx::in_progress() query

This library release conforms to the UPC++ v1.0 Specification, Revision 2020.10.0. All currently specified features are fully implemented. See the UPC++ issue tracker for status of known bugs.

Breaking changes:

  • Build-time UPCXX_CODEMODE/-codemode value "O3" has been renamed to "opt". For backwards compat, the former is still accepted.
  • upcxx_memberof(_general)(gp, mem) now produce a global_ptr<T> when mem names an array whose element type is T.
  • atomic_domain construction now has user-level progress
  • Initiating collective operations with progress level user from inside the restricted context (within a callback running inside progress) is now prohibited, and diagnosed with a runtime error. Most such calls previously led to silent deadlock.
  • Initiating collective operations with a progress level of internal or none from within the restricted context (within a callback running inside progress) is now a deprecated behavior, and diagnosed with a runtime warning. For details, see spec issue 169.

2020.07.17: Bug-fix release 2020.3.2

New features/enhancements:

  • Shared heap exhaustion in upcxx::new_(array) now throws upcxx::bad_shared_alloc (a type derived from std::bad_alloc) which provides additional diagnostics about the failure.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #343: Guarantee equality for default-constructed invalid upcxx::team_id
  • issue #353: configure: automatically cross-compile on Cray XC
  • issue #356: SERIALIZED_{FIELDS|VALUES} incorrectly require public constructors
  • issue #369: completion_cx::as_future() typically leaks
  • issue #371: team_ids are not "universal" as documented
  • issue #373: No python in $PATH in recent Linux distros
  • issue #380: Compile regression on bulk upcxx::rput with source+operation completions

Breaking changes:

  • Configure-time envvar CROSS has been renamed to UPCXX_CROSS. For backwards compat, the former is still accepted when the latter is unset.
  • Construction of upcxx::team_id is no longer Trivial (was never guaranteed to be). It remains DefaultConstructible, TriviallyCopyable, StandardLayoutType, EqualityComparable

2020.03.12: Release 2020.3.0

Infrastructure and requirements changes:

  • The install and run-tests scripts have been replaced with a configure script (with GNU autoconf-compatible options) and corresponding make all, make check and make install steps. See INSTALL.md for step-by-step instructions.
  • The minimum required GNU Make has risen from 3.79 to 3.80 (released in 2002).
  • There is no longer a requirement for Python2.7. While upcxx-run still requires a Python interpreter, both Python3 and Python2 (>= 2.7.5) are acceptable.

New features/enhancements: (see specification and programmer's guide for full details)

  • Added support for non-trivial serialization of user-defined types. See the new chapter of the programmer's guide for an introduction, and the specification for all the details.
  • Implement upcxx_memberof(_general)(), enabling RMA access to fields of remote objects
  • upcxx::promise now behaves as a CopyAssignable handle to a reference-counted hidden object, meaning users no longer have to worry about promise lifetime issues.
  • atomic_domain<T> is now implemented for all 32- and 64-bit integer types, in addition to the fixed-width integer types, (u)int{32,64}_t.
  • atomic_domain<T> implementation has been streamlined to reduce overhead, most notably for targets with affinity to local_team using CPU-based atomics.
  • Improve GASNet polling heuristic to reduce head-of-line blocking
  • Installed CMake package file is now more robust, and supports versioning.
  • global_ptr<T> now enforces many additional correctness checks in debug mode
  • Significantly improve compile latency associated with upcxx compiler wrapper script
  • Improve handling of upcxx-run -v verbose options

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #83: strengthen global_ptr correctness assertions
  • issue #142: Clarify semantic restrictions on UPCXX_THREADMODE=seq
  • issue #173: upcxx-run for executables in CWD and error messages
  • issue #247: unused-local-typedef warning for some compilers
  • issue #266: Redundant completion handler type breaks compile
  • issue #267: Same completion value can't be received multiple times
  • issue #271: Use of -pthread in example/prog-guide/Makefile breaks PGI
  • issue #277: Ensure completion_cx::as_promise(p) works even when p is destroyed prior to fulfillment
  • issue #280: Off-by-one error in promise constructor leads to breakage with any explicit argument
  • issue #282: Improve installed CMake package files
  • issue #287: Bogus install warnings on Cray XC regarding CC/CXX
  • issue #289: link failure for PGI with -std=c++17
  • issue #294: redundant-move warning in completion.hpp from g++-9.2
  • issue #304: Bad behavior for misspelled CC or CXX
  • issue #323: Incorrect behavior for global_ptr<T>(nullptr).is_local() in multi-node jobs
  • issue #333: Multiply defined symbol detail::device_allocator_core<cuda_device>::min_alignment w/ std=c++2a
  • spec issue 148: Missing const qualifiers on team and other API objects
  • spec issue 155: value argument type to value collectives is changed to a simple by-value T

This library release mostly conforms to the UPC++ v1.0 Specification, Revision 2020.3.0. The following features from that specification are not yet implemented:

  • view buffer lifetime extension for remote_cx::as_rpc (issue #262)
  • boost_serialization adapter class is not yet implemented (issue #330)
  • future<T&&>::{result,wait}_reference currently return the wrong type. E.g. future<int&&>::result_reference() returns int const& instead of int&& (issue #313)

Breaking changes:

  • atomic_domain has been specified as non-DefaultConstructible since UPC++ spec draft 8 (2018.9) This release removes the default constructor implementation from atomic_domain (issue #316).
  • The trait template classes upcxx::is_definitely_trivially_serializable and upcxx::is_definitely_serializable have been renamed such that the "definitely" word has been dropped, e.g.: upcxx::is_trivially_serializable.
  • future::{result,wait}_moved, deprecated since 2019.9.0, have been removed. New future::{result,wait}_reference calls provide related functionality.

2019.09.14: Release 2019.9.0

New features/enhancements: (see specification and programmer's guide for full details)

  • upcxx has several new convenience options (see upcxx -help)
  • upcxx::rput(..., remote_cx::as_rpc(...)) has received an improved implementation for remote peers where the dependent RPC is injected immediately following the put. This pipelining reduces latency and sensitivity to initiator attentiveness, improving performance in most cases (for the exception, see issue #261).
  • Accounting measures have been added to track the shared-heap utilization of the UPC++ runtime (specifically rendezvous buffers) so that in the case of shared-heap exhaustion an informative assertion will fire. Also, fewer rendezvous buffers are now required by the runtime, thus alleviating some pressure on the shared heap.
  • Environment variable UPCXX_OVERSUBSCRIBED has been added to control whether the runtime should yield to OS (sched_yield()) within calls to upcxx::progress()). See docs/oversubscription.md.
  • Release tarball downloads now embed a copy of GASNet-EX that is used by default during install. Git clones of the repo will still default to downloading GASNet-EX during install. The GASNET envvar can still be set at install time to change the default behavior.
  • A CMake module for UPC++ is now installed. See 'Using UPC++ with CMake' in README.md
  • atomic_domain<float> and atomic_domain<double> are now implemented
  • Interoperability support for Berkeley UPC's -pthreads mode, see docs/upc-hybrid.md
  • New define UPCXX_SPEC_VERSION documents the implemented revision of the UPC++ specification

Support has been added for the following compilers/platforms (for details, see 'System Requirements' in INSTALL.md):

  • PGI v19.1+ on Linux/x86_64
  • PGI v18.10+ on Linux/ppc64le
  • clang v5.0+ on Linux/ppc64le
  • PrgEnv/cray with CCE v9.0+ on the Cray XC
  • ALCF's PrgEnv/llvm v4.0+ on the Cray XC
  • NEW platform: Linux/aarch64 (aka "arm64" or "armv8")
    • gcc v6.4.0+
    • clang 4.0.0+

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #140: upcxx::discharge() does not discharge remote_cx::as_rpc()
  • issue #168: upcxx::progress_required always returns 0
  • issue #170: team_id::when_here() is unimplemented
  • issue #181: Library linkage failures when user compiles with a different -std=c++ level
  • issue #184: bench/put_flood crashes on opt/Linux
  • issue #203: strict aliasing violations in device_allocator
  • issue #204: No support for nvcc --compiler-bindir=...
  • issue #210: cuda_device::default_alignment() not implemented
  • issue #223: operator<<(std::ostream, global_ptr<T>) does not match spec
  • issue #224: missing const qualifier on dist_object<T>.fetch()
  • issue #228: incorrect behavior for upcxx -g -O
  • issue #229: Teach upcxx wrapper to compile C language files
  • issue #234: Generalized operation completion for barrier_async and broadcast
  • issue #243: Honor $UPCXX_PYTHON during install
  • issue #260: GASNET_CONFIGURE_ARGS can break UPC++ build
  • issue #264: upcxx-meta CXX and CC are not full-path expanded
  • issue #268: Completion handlers can't accept && references
  • spec issue 141: resolve empty transfer ambiguities (count=0 RMA)
  • spec issue 142: add persona::active_with_caller()

This library release mostly conforms to the UPC++ v1.0 Specification, Revision 2019.9.0. The following features from that specification are not yet implemented:

  • view buffer lifetime extension for remote_cx::as_rpc (issue #262)
  • User-defined Class Serialization interface (coming soon!)

Breaking changes:

  • Applications are recommended to replace calls to std::getenv with upcxx::getenv_console, to maximize portability to loosely coupled distributed systems.
  • envvar UPCXX_GASNET_CONDUIT has been renamed to UPCXX_NETWORK. For backwards compat, the former is still accepted when the latter is unset.
  • upcxx::allocate() and device_allocator<Device>::allocate() have changed signature. The alignment parameter has moved from being a final defaulted template argument to being a final defaulted function argument.
  • future<T...>::result_moved() and future<T...>::wait_moved() are deprecated, and will be removed in a future release.

2019.05.27: Bug-fix release 2019.3.2

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #209: Broken install defaulting of CC/CXX on macOS

Embeds a GASNet-EX library that addresses the following notable issues (see the GASNet issue tracker for details):

  • bug3943: infrequent startup hang with PSHM and over 62 PPN

2019.03.15: Release 2019.3.0

New features/enhancements: (see specification and programmer's guide for full details)

  • Prototype Memory Kinds support for CUDA-based NVIDIA GPUs, see INSTALL.md. Note the CUDA support in this UPC++ release is a proof-of-concept reference implementation which has not been tuned for performance. In particular, the current implementation of upcxx::copy does not utilize hardware offload and is expected to underperform relative to solutions using RDMA, GPUDirect and similar technologies. Performance will improve in an upcoming release.
  • Support for interoperability with Berkeley UPC, see upc-hybrid.md
  • There is now an offline installer package for UPC++, for systems lacking connectivity
  • Barrier synchronization performance has been improved
  • Installer now defaults to more build parallelism, improving efficiency (see UPCXX_MAKE)

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #100: Fix shared heap setting propagation on loosely-coupled clusters
  • issue #118: Enforce GASNet-EX version interlock at compile time
  • issue #177: Completion broken for non-fetching binary AMOs
  • issue #183: bench/{put_flood,nebr_exchange} were failing to compile
  • issue #185: Fix argument order for dist_object constructor to match spec
  • issue #187: Improve Python detection logic for the install script
  • issue #190: Teach upcxx-run to honor UPCXX_PYTHON
  • issue #202: Make global_ptr::operator bool conversion explicit
  • issue #205: incorrect metadata handling in ~persona_scope()

This library release mostly conforms to the UPC++ v1.0 Draft 10 Specification. The following features from that specification are not yet implemented:

  • barrier_async() and broadcast() only support default future-based completion (issue #234)
  • atomic_domain<float> and atomic_domain<double> are not yet implemented (issue #235)
  • team_id::when_here() is unimplemented (issue #170)
  • User-defined Class Serialization interface

Breaking changes:

  • envvar UPCXX_SEGMENT_MB has been renamed to UPCXX_SHARED_HEAP_SIZE. For backwards compat, the former is still accepted when the latter is unset.
  • The minimum-supported version of GNU g++ is now 6.4.0
    • This also applies to the stdlibc++ used by Clang or Intel compilers
  • The minimum-supported version of llvm/clang for Linux is now 4.0

2018.09.26: Release 2018.9.0

New features/enhancements: (see specification and programmer's guide for full details)

  • Subset teams and team-aware APIs are added and implemented
  • Non-Blocking Collective operations, with team support: barrier, broadcast, reduce
  • New atomic operations: mul, min, max, bit_{and,or,xor}
  • future::{wait,result}* return types are now "smarter", allowing more concise syntax
  • New upcxx compiler wrapper makes it easier to build UPC++ programs
  • upcxx-run: improved functionality and handling of -shared-heap arguments
  • New supported platforms:
    • GNU g++ compiler on macOS (e.g., as installed by Homebrew or Fink)
    • PrgEnv-intel version 17.0.2 or later on Cray XC x86-64 systems
    • Intel C++ version 17.0.2 or later on x86-64/Linux
    • GNU g++ compiler on ppc64le/Linux
  • rput_{strided,(ir)regular} now provide asynchronous source completion
  • Performance improvements to futures, promises and LPCs
  • UPC++ library now contains ident strings that can be used to query version info from a compiled executable, using the UNIX ident tool.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #49: stability and portability issues caused by C++ thread_local
  • issue #141: missing promise move assignment operator

This library release mostly conforms to the UPC++ v1.0 Draft 8 Specification. The following features from that specification are not yet implemented:

  • barrier_async() and broadcast() only support default future-based completion (issue #234)
  • atomic_domain<float> and atomic_domain<double> are not yet implemented (issue #235)
  • team_id::when_here() is unimplemented (issue #170)
  • User-defined Serialization interface

Breaking changes:

  • global_ptr<T>(T*) "up-cast" constructor has been replaced with to_global_ptr<T>(T*)
  • atomic_domain now requires a call to new collective destroy() before destructor
  • allreduce has been renamed to reduce_all

2018.05.10: Release 2018.3.2

This is a re-release of version 2018.3.0 (see below) that corrects a packaging error.

2018.03.31: Release 2018.3.0

New features/enhancements:

  • Non-Contiguous One-Sided RMA interface is now fully implemented.
  • Remote Atomics have been revamped, expanded and implemented. See the updated specification for usage details. The current implementation leverages hardware support in shared memory and NIC offload support in Cray Aries.
  • View-Based Serialization - see the specification for details
  • Implementation of local memory translation (achieved with global_ptr::local() / global_ptr(T*)). This encompasses a limited implementation of teams to support upcxx::world and upcxx::local_team so clients may query their local neighborhood of ranks.

Notable issues resolved (see the UPC++ issue tracker for details):

  • issue #119: Build system is now more robust to GASNet-EX download failures.
  • issue #125: Fix upcxx-run exit code handling.
  • Minor improvements to upcxx-run and run-tests.

This library release mostly conforms to the UPC++ v1.0 Draft 6 Specification. The following features from that specification are not yet implemented:

  • Teams: team::split, team_id, collectives over teams, passing team& arguments to rpcs, constructing dist_object over teams.
  • Vector broadcast broadcast(T *buf, size_t count, ...)
  • barrier_async
  • User-defined Serialization interface

This release is not yet performant, and may be unstable or buggy.

Please report any problems in the issue tracker.

2018.01.31: Release 2018.1.0 BETA

This is a BETA preview release of UPC++ v1.0.

New features/enhancements:

  • Generalized completion. This allows the application to be notified about the status of UPC++ operations in a handful of ways. For each event, the user is free to choose among: futures, promises, callbacks, delivery of remote procedure calls, and in some cases even blocking until the event has occurred.
  • Internal use of lock-free datastructures for lpc queues.
  • Improvements to the upcxx-run command.
  • Improvements to internal assertion checking and diagnostics.

This library release mostly conforms to the UPC++ v1.0 Draft 5 Specification. The following features from that specification are not yet implemented:

  • Teams
  • Vector broadcast broadcast(T *buf, size_t count, ...)
  • barrier_async
  • Serialization
  • Non-contiguous transfers
  • Atomics

This release is not performant, and may be unstable or buggy.

2017.09.30: Release 2017.9.0

The initial public release of UPC++ v1.0.

This library release mostly conforms to the UPC++ v1.0 Draft 4 Specification. The following features from that specification are not yet implemented:

  • Continuation-based and Promise-based completion (use future completion for now)
  • rput_then_rpc
  • Teams
  • Vector broadcast broadcast(T *buf, size_t count, ...)
  • barrier_async
  • Serialization
  • Non-contiguous transfers

This release is not performant, and may be unstable or buggy.

2017.09.01: Release v1.0-pre

This is a pre-release of v1.0. This pre-release supports most of the functionality covered in the UPC++ specification, except personas, promise-based completion, teams, serialization, and non-contiguous transfers. This pre-release is not performant, and may be unstable or buggy. Please notify us of issues by sending email to upcxx@googlegroups.com.

Updated