Skip to content

Latest commit

 

History

History
208 lines (180 loc) · 10.4 KB

CHANGELOG.md

File metadata and controls

208 lines (180 loc) · 10.4 KB
Raw

This file contains the main features as well as overviews of specific changes to this project (since v1.1.0).

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

v4.0.2411

Added

  • PaRSEC API 4.0.
  • Add DTD CUDA support including NEW tiles in DTD.
  • Add RoCM/HIP device support.
  • Add IrisXE/Level0 device support (experimental).
  • Enable users to manage their own data copies without PaRSEC interfering. Data copies are marked as being owned by PaRSEC or not and managed by PaRSEC or not. A data copy owned by PaRSEC can be reclaimed by PaRSEC when its reference count reaches 0, a data copy managed by PaRSEC can be copied / moved onto a different device, while a data copy not managed by PaRSEC will never be moved by the runtime.
  • Add an info system, and introduce two info hooks. See parsec/class/info.h for details. The info system allows the user to register info objects with different levels of structures and dynamic objects in the PaRSEC runtime.
  • PTG supports user-defined routines to move data between GPU and CPU, and user-defined sizes for buffers allocated on the GPU.
  • PTG supports reshaping data propagated between local tasks and the speficiation of two types on acccesses to data colletions.
  • PINS log SCHEDULE_BEGIN and SCHEDULE_END events to better track tasks lifecycle.
  • Detect and report oversubscribed binding of core resources.
  • PaRSEC Thread binding can be disabled (bind_threads 0 MCA parameter).
  • Load balancing between GPUs can be tuned (device_load_balance_skew MCA parameter).
  • Load balancing exclusivity between CPU/GPUs can be disabled (device_load_balance_allow_cpu MCA parameter).
  • Data sent in messages can be of variable size.
  • New API parsec_context_query can be used to obtain information on the system, like the number of devices, ranks, etc.
  • New active-message communication API gives low-level access to the PaRSEC communication system to DSLs.

Changed

  • Single letter command line options have been replaced with --mca parameters. --help is now --parsec-help.
  • Renamed symbols related to data distribution to properly prefix them with the parsec_ prefix. The old symbols have been deprecated.
  • DTD interface change: the global array parsec_dtd_arena_datatypes is replaced with functions to create, destroy, and get arena datatypes for DTD, and these objects now live inside the parsec context.
  • PARSEC_SUCCESS changed to 0 (from -1), all values for PARSEC_ERR_XYZ changed.
  • PaRSEC now requires CMake 3.21.
  • PaRSEC profiling tools now require Python 3.x
  • PaRSEC profiling system does not require for local dicitonaries to be identical between ranks anymore.
  • time_estimate functions can be used to control task load balancing (replaces weight PTG property).

Deprecated

  • data distribution w/o the parsec_ prefix. Further documentation (including a sed script) can be found in contrib/renaming.

Removed

  • PaRSEC API 3.0
  • RECURSIVE Device support (this is temporary and will be restored in a future version).
  • Removed obsolete dbp2paje tool; h5totrace is the replacement tool to use. This removes the optional dependency on GTG.
  • Removed all command line options not prefixed by --mca, except for --parsec-help and --parsec-version.
  • Using more than PARSEC_GPU_MAX_WORKSPACE workspaces per device will now cause an error (instead of computing incorrect values).
  • PTG property weight (replaced by time_estimate).

Fixed

  • DTD Termination detection would occasionally assert.
  • Multiple bugs with GPU data ownership causing crashes and incorrect results when executing with more than 1 GPU.
  • Device-to-device memory copies would not work in some scenarios.
  • Suboptimal ordering of members in broadcast tree could cause performance reduction.
  • Cray MPI and MPICH would crash in MPI_Cancel and when using NULL datatypes.
  • Do not report incorrect flops/s capabilities (device_show_capabilities MCA parameter).
  • On some systems PaRSEC would allocate more GPU memory than is available on the device.
  • Performance with large number of GPU tasks with the same priority would be poor due to overhead of sorting by priority.

Known Bugs

  • PaRSEC Thread binding ignores externally provided binding (e.g., a cpuset enforced by srun); see issue ICLDisco/dplasma#9.
  • Enabling the RECURSIVE device will cause crashes (it is disabled by default in this release); see issues #548, #541.
  • Running out of GPU memory when using the NEW keyword in PTG may cause deadlocks; see issue #527.

Security

v3.0.2012

  • PaRSEC API 3.0
  • PaRSEC now requires CMake 3.16.
  • New configure system to ease the installation of PaRSEC. See INSTALL for details. This system automates installation on most DOE leadership systems.
  • Split DPLASMA and PaRSEC into separate repositories. PaRSEC moves from cmake-2.0 to cmake-3.12, using targets. Targets are exported for third-party integration
  • Add visualization tools to extract user-defined properties from the application (see: PR 229 visualization-tools)
  • Automate expression of required data transfers from host-to-device and device-to-host to satisfy depencencies (and anti-dependencies). PaRSEC tracks multiple versions of the same data as data copies with a coherency algorithm that initiates data transfers as needed. The heurisitic for the eviction policy in out-of-memory event on GPU has been optimized to allow for efficient operation in larger than GPU memory problems.
  • Add support for MPI out-of-order matching capabilities; Added capability for compute threads to send direct control messages to indicate completion of tasks to remote nodes (without delegation to the communication thread)
  • Remove communication mode EAGER from the runtime. It had a rare but hard to correct bug that would rarely deadlock, and the performance benefit was small.
  • Add a Map operator on the Block Cyclic matrix data collection that performs in-place data transformation on the collection with a user provided operator.
  • Add support in the runtime for user-defined properties evaluated at runtime and easy to export through a shared memory region (see: PR 229 visualization-tools)
  • Add a PAPI-SDE interface to the parsec library, to expose internal counters via the PAPI-Software Defined Events interface.
  • Add a backend support for OTF2 in the profiling mechanism. OTF2 is used automatically if a OTF2 installation is found.
  • Add a MCA parameter to control the number of ejected blocks from GPU memory (device_cuda_max_number_of_ejected_data). Add a MCA parameter to control wether or not the GPU engine will take some time to sort the first N tasks of the pending queue (device_cuda_sort_pending_list).
  • Reshape the users vision of PaRSEC: they only have to include a single header (parsec.h) for most usages, and link with a single library (-lparsec).
  • Update the PaRSEC DSL handling of initial tasks. We now rely on 2 pieces of information: the number of DSL tasks, and the number of tasks imposed by the system (all types of data transfer).
  • Add a purely local scheduler (ll), that uses a single LIFO per thread. Each schedule operation does 1 atomic (push in local queue), each select operation does up to t atomics (pop in local queue, then try any other thread's queue until they are all tested empty).
  • Add a --ignore-properties=... option to parsec_ptgpp
  • Change API of hash tables: allow keys of arbitrary size. The API features how to build a key from a task; how to hash a key into 1 <= N <= 64 bits; and how to compare twy keys (plus a printing function to debug).
  • Change behavior of DEBUG_HISTORY: log all information inside a buffer of fixed size (MCA parameter) per thread, do not allocate memory during logging, and use timestamp to re-order output when the user calls dump()
  • DTD interface is updated (new flag to send pointer as parameter, unpacking of paramteres is simpler etc).
  • DTD provides mca param (dtd_debug_verbose) to print information about traversal of DAG in a separate output stream from the default.

v2.0.0rc2

  • Rename all functions, files, directories from dague/DAGUE/DAGuE to parsec/PARSEC/PaRSEC.

v2.0.0rc1

  • .so support. Dynamic Library Build has been succesfully tested on Linux platforms, reducing significantly the size of the compiled dplasma/testing directory. Note that on modern architectures, all depending libraries must be compiled either as Position Independent Code (-fPIC) or as shared libraries. Hint: add --cflags="-fPIC" when running the plasma-installer.
  • The "inline_c %{ ... %}" block syntax has been simplified to either "%c{ ... %}" or "%{ ... %}". The change is backward compatible.

v1.2.0

  • The support for C is required from MPI.
  • Revert to an older LEX syntax (not (?i:xxx))
  • Don't recursively call the MPI communication engine progress function.
  • Protect the runtime headers from inclusion in C++ files.
  • Fix a memory leak allowing the remote data to be released once used.
  • Integrate the new profiling system and the python interface (added Panda support).
  • Change several default parameters for the DPLASMA tests.
  • Add Fortran support for the PaRSEC and the profiling API and add tests to validate it.
  • Backport all the profiling features from the devel branch (panda support, simpler management, better integration with python, support for HDF5...).
  • Change the colorscheme of the .dot generator
  • Correctly compute the identifier of the tasks (ticket #33).
  • Allocate the list items from the corresponding list using the requested gap.
  • Add support for older lex (without support for ?i).
  • Add support for 128 bits atomics and resolve the lingering ABA issue. When 128 bits atomics are not available fall back to an atomic lock implementation of the most basic data structures.
  • Required Cython 0.19.1 (at least)
  • Completely disconnect the ordering of parameters and locals in the JDF.
  • Many other minor bug fixes, code readability impeovement and typos.
  • DPLASMA:
    • Add the doxygen documentation generation.
    • Improved ztrmm with all the matrix reads unified.
    • Support for potri functions (trtri+lauum), and corresponding testings.
    • Fix bug in symmetric/hermitian norms when A.mt < P.