starpu-max/ChangeLog

# StarPU --- Runtime system for heterogeneous multicore architectures.
#
# Copyright (C) 2012-2014,2016-2018                      Inria
# Copyright (C) 2009-2019                                Université de Bordeaux
# Copyright (C) 2010-2019                                CNRS
#
# StarPU is free software; you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation; either version 2.1 of the License, or (at
# your option) any later version.
#
# StarPU is distributed in the hope that it will be useful, but
# WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
#
# See the GNU Lesser General Public License in COPYING.LGPL for more details.
#

StarPU 1.4.0 (svn revision xxxx)
==============================================
New features:
  * Fault tolerance support with starpu_task_ft_failed().

StarPU 1.3.2 (git revision xxx)
==============================================

StarPU 1.3.1 (git revision 01949488b4f8e6fe26d2c200293b8aae5876b038)
==============================================

Small features:
  * Add starpu_filter_nparts_compute_chunk_size_and_offset helper.
  * Add starpu_bcsr_filter_canonical_block_child_ops.

Small changes:
  * Improve detection of NVML availability. Do not only check the
    library is available, also check the compiled code can be run.

StarPU 1.3.0 (git revision 24ca83c6dbb102e1cfc41db3bb21c49662067062)
==============================================

New features:
  * New scheduler 'heteroprio' with heterogeneous priorities
  * Support priorities for data transfers.
  * Add support for multiple linear regression performance models
    - Bump performance model file format version to 45.
  * Add MPI Master-Slave support to use the cores of remote nodes. Use the
    --enable-mpi-master-slave option to activate it.
  * Add STARPU_CUDA_THREAD_PER_DEV environment variable to support driving all
    GPUs from only one thread when almost all kernels are asynchronous.
  * Add starpu_replay tool to replay tasks.rec files with Simgrid.
  * Add experimental support of NUMA nodes. Use STARPU_USE_NUMA to activate it.
  * Add a new set of functions to make Out-of-Core based on HDF5 Library.
  * Add a new implementation of StarPU-MPI on top of NewMadeleine
  * Add optional callbacks to notify an external resource manager
    about workers going to sleep and waking up
  * Add implicit support for asynchronous partition planning. This means one
    does not need to call starpu_data_partition_submit() etc. explicitly any
    more, StarPU will make the appropriate calls as needed.
  * Add starpu_task_notify_ready_soon_register() to be notified when it is
    determined when a task will be ready an estimated amount of time from now.
  * New StarPU-MPI initialization function (starpu_mpi_init_conf())
    which allows StarPU-MPI to manage reserving a core for the MPI thread, or
    merging it with CPU driver 0.
  * Add possibility to delay the termination of a task with the
    functions starpu_task_end_dep_add() which specifies the number of
    calls to the function starpu_task_end_dep_release() needed to
    trigger the task termination, or with starpu_task_declare_end_deps_array()
    and starpu_task_declare_end_deps() to just declare termination dependencies
    between tasks.
  * Add possibility to define the sequential consistency at the task level
    for each handle used by the task.
  * Add STARPU_SPECIFIC_NODE_LOCAL, STARPU_SPECIFIC_NODE_CPU, and
    STARPU_SPECIFIC_NODE_SLOW as generic values for codelet specific memory
    nodes which can be used instead of exact node numbers.
  * Add starpu_get_next_bindid() and starpu_bind_thread_on() to allow
    binding an application-started thread on a free core. Use it in
    StarPU-MPI to automatically bind the MPI thread on an available core.
  * Add STARPU_RESERVE_NCPU environment variable and
    starpu_config::reserve_ncpus field to make StarPU use a few cores
    less.
  * Add STARPU_MAIN_THREAD_BIND environment variable to make StarPU reserve a
    core for the main thread.
  * New StarPU-RM resource management module to share processor cores and
    accelerator devices with other parallel runtime systems. Use
    --enable-starpurm option to activate it.
  * New schedulers modular-gemm, modular-pheft, modular-prandom and
    modular-prandom-prio
  * Add STARPU_MATRIX_SET_NX/NY/LD and STARPU_VECTOR_SET_NX to change a matrix
    tile or vector size without reallocating the buffer.
  * Application can change the allocation used by StarPU with
    starpu_malloc_set_hooks()
  * XML output for starpu_perfmodel_display and starpu_perfmodel_dump_xml()
    function

Small features:
  * Scheduling contexts may now be associated a user data pointer at creation
    time, that can later be recalled through starpu_sched_ctx_get_user_data().
  * New environment variables STARPU_SIMGRID_TASK_SUBMIT_COST and
    STARPU_SIMGRID_FETCHING_INPUT_COST to simulate the cost of task
    submission and data fetching in simgrid mode.
    This provides more accurate simgrid predictions, especially for the
    beginning of the execution and regarding data transfers.
  * New environment variable STARPU_SIMGRID_SCHED_COST to take into
    account the time to perform scheduling when running in SimGrid mode.
  * New configure option --enable-mpi-pedantic-isend (disabled by
    default) to acquire data in STARPU_RW (instead of STARPU_R) before
    performing MPI_Isend() call
  * New function starpu_worker_display_names() to display the names of
    all the workers of a specified type.
  * Arbiters now support concurrent read access.
  * Add a field starpu_task::where similar to starpu_codelet::where
    which allows to restrict where to execute a task. Also add
    STARPU_TASK_WHERE to be used when calling starpu_task_insert().
  * Add SubmitOrder trace field.
  * Add workerids and workerids_len task fields.
  * Add priority management to StarPU-MPI. Can be disabled with
    the STARPU_MPI_PRIORITIES environment variable.
  * Add STARPU_MAIN_THREAD_CPUID and STARPU_MPI_THREAD_CPUID environment
    variables.
  * Add disk to disk copy functions and support asynchronous full read/write
    in disk backends.
  * New starpu_task_insert() parameter STARPU_CL_ARGS_NFREE which allows
    to set codelet parameters but without freeing them.
  * New starpu_task_insert() parameter STARPU_TASK_DEPS_ARRAY which
    allows to declare task dependencies similarly to
    starpu_task_declare_deps_array()
  * Add dependency backward information in debugging mode for gdb's
    starpu-print-task
  * Add sched_data field in starpu_task structure.
  * New starpu_fxt_tool option -label-deps to label dependencies on
    the output graph
  * New environment variable STARPU_GENERATE_TRACE_OPTIONS to specify
    fxt options (to be used with STARPU_GENERATE_TRACE)
  * New function starpu_task_set() similar as starpu_task_build() but
    with a task object given as the first parameter
  * New functions
    starpu_data_partition_submit_sequential_consistency() and
    starpu_data_unpartition_submit_sequential_consistency()
  * Add a new value STARPU_TASK_SYNCHRONOUS to be used in
    starpu_task_insert() to define if the task is (or not) synchronous
  * Add memory states events in the traces.
  * Add starpu_sched_component_estimated_end_min_add() to fix termination
    estimations in modular schedulers.
  * New function starpu_data_partition_not_automatic() to disable the
    automatic partitioning of a data handle for which a asynchronous
    plan has previously been submitted
  * Add starpu_task_declare_deps()
  * New function starpu_data_unpartition_submit_sequential_consistency_cb()
    to specify a callback for the task submitting the unpartitioning
  * New tool starpu_mpi_comm_trace.py to draw heatmap of MPI
    communications
  * Support for ARM performance libraries
  * Add functionality to disable signal catching either through field
    starpu_conf::catch_signals or through the environment variable
    STARPU_CATCH_SIGNALS
  * Support for OpenMP Taskloop directive
  * Optional data interface init function (used by the vector and
    matrix interfaces)

Changes:
  * Vastly improve simgrid simulation time.
  * Switch default scheduler to lws.
  * Add "to" parameter to pull_task and can_push methods of
    components.
  * Deprecate starpu_data_interface_ops::handle_to_pointer interface
    operation in favor of new starpu_data_interface_ops::to_pointer
    operation.
  * Sort data access requests by priority.
  * Cluster support is disabled by default, unless the configure
    option --enable-cluster is specified
  * For unpack operations, move the memory deallocation from
    starpu_data_unpack() to the interface function
    starpu_data_interface_ops::unpack_data(). Pack and unpack
    functions of predefined interfaces
    use public API starpu_malloc_on_node_flags() and
    starpu_free_on_node_flags() to allocate and de-allocate memory

Small changes:
  * Use asynchronous transfers for task data fetches with were not prefetched.
  * Allow to call starpu_sched_ctx_set_policy_data on the main
    scheduler context
  * Fonction starpu_is_initialized() is moved to the public API.
  * Fix code to allow to submit tasks to empty contexts
  * STARPU_COMM_STATS also displays the bandwidth
  * Update data interfaces implementations to only use public API

StarPU 1.2.9 (git revision xxx)
==============================================
  * Add STARPU_SIMGRID_TRANSFER_COST environment variable to easily disable
    data transfer costs.

StarPU 1.2.8 (git revision f66374c9ad39aefb7cf5dfc31f9ab3d756bcdc3c)
==============================================

Small features:
  * Minor fixes

StarPU 1.2.7 (git revision 07cb7533c22958a76351bec002955f0e2818c530)
==============================================

Small features:
  * Add STARPU_HWLOC_INPUT environment variable to save initialization time.
  * Add starpu_data_set/get_ooc_flag.

StarPU 1.2.6 (git revision 23049adea01837479f309a75c002dacd16eb34ad)
==============================================

Small changes:
  * Fix crash for lws scheduler
  * Avoid making hwloc load PCI topology when CUDA is not enabled

StarPU 1.2.5 (git revision 22f32916916d158e3420033aa160854d1dd341bd)
==============================================

Small features:
  * Add a new value STARPU_TASK_COLOR to be used in
    starpu_task_insert() to pick up the color of a task in dag.dot
  * Add starpu_data_pointer_is_inside().

Changes:
  * Do not export -lcuda -lcudart -lOpenCL in *starpu*.pc.

StarPU 1.2.4 (git revision 255cf98175ef462749780f30bfed21452b74b594)
==============================================

Small features:
   * Catch of signals SIGINT and SIGSEGV to dump fxt trace files.
   * New configure option --disable-icc to disable the compilation of
     specific ICC examples
   * Add starpu_codelet_pack_arg_init, starpu_codelet_pack_arg,
     starpu_codelet_pack_arg_fini for more fine-grain packing capabilities.
   * Add starpu_task_insert_data_make_room,
     starpu_task_insert_data_process_arg,
     starpu_task_insert_data_process_array_arg,
     starpu_task_insert_data_process_mode_array_arg
   * Do not show internal tasks in fxt dag by default. Allow to hide
     acquisitions too.
   * Add a way to choose the dag.dot colors.


StarPU 1.2.3 (svn revision 22444)
==============================================

New features:
  * Add per-node MPI data.

Small features:
  * When debug is enabled, starpu data accessors first check the
    validity of the data interface type
  * Print disk bus performances when STARPU_BUS_STATS is set
  * Add starpu_vector_filter_list_long filter.
  * Data interfaces now define a name through the struct starpu_data_interface_ops
  * StarPU-MPI :
    - allow predefined data interface not to define a mpi datatype and
      to be exchanged through pack/unpack operations
    - New function starpu_mpi_comm_get_attr() which allows to return
      the value of the attribute STARPU_MPI_TAG_UB, i.e the upper
      bound for tag value.
    - New configure option enable-mpi-verbose to manage the display of
      extra MPI debug messages.
  * Add STARPU_WATCHDOG_DELAY environment variable.
  * Add a 'waiting' worker status
  * Allow new value 'extra' for configure option --enable-verbose

Small changes:
  * Add data_unregister event in traces
  * StarPU-MPI
    - push detached requests at the back of the testing list, so they
      are tested last since they will most probably finish latest
  * Automatically initialize handles on data acquisition when
    reduction methods are provided, and make sure a handle is
    initialized before trying to read it.

StarPU 1.2.2 (svn revision 21308)
==============================================

New features:
  * Add starpu_data_acquire_try and starpu_data_acquire_on_node_try.
  * Add NVCC_CC environment variable.
  * Add -no-flops and -no-events options to starpu_fxt_tool to make
    traces lighter
  * Add starpu_cusparse_init/shutdown/get_local_handle for proper CUDA
    overlapping with cusparse.
  * Allow precise debugging by setting STARPU_TASK_BREAK_ON_PUSH,
    STARPU_TASK_BREAK_ON_SCHED, STARPU_TASK_BREAK_ON_POP, and
    STARPU_TASK_BREAK_ON_EXEC environment variables, with the job_id
    of a task. StarPU will raise SIGTRAP when the task is being
    scheduled, pushed, or popped by the scheduler.

Small features:
  * New function starpu_worker_get_job_id(struct starpu_task *task)
    which returns the job identifier for a given task
  * Show package/numa topology in starpu_machine_display
  * MPI: Add mpi communications in dag.dot
  * Add STARPU_PERF_MODEL_HOMOGENEOUS_CPU environment variable to
    allow having one perfmodel per CPU core
  * Add starpu_perfmodel_arch_comb_fetch function.
  * Add starpu_mpi_get_data_on_all_nodes_detached function.

Small changes:
  * Output generated through STARPU_MPI_COMM has been modified to
    allow easier automated checking
  * MPI: Fix reactivity of the beginning of the application, when a
    lot of ready requests have to be processed at the same time, we
    want to poll the pending requests from time to time.
  * MPI: Fix gantt chart for starpu_mpi_irecv: it should use the
    termination time of the request, not the submission time.
  * MPI: Modify output generated through STARPU_MPI_COMM to allow
    easier automated checking
  * MPI: enable more tests in simgrid mode
  * Use assumed-size instead of assumed-shape arrays for native
    fortran API, for better backward compatibility.
  * Fix odd ordering of CPU workers on CPUs due to GPUs stealing some
    cores

StarPU 1.2.1 (svn revision 20299)
==============================================

New features:
  * Add starpu_fxt_trace_user_event_string.
  * Add starpu_tasks_rec_complete tool to add estimation times in tasks.rec
    files.
  * Add STARPU_FXT_TRACE environment variable.
  * Add starpu_data_set_user_data and starpu_data_get_user_data.
  * Add STARPU_MPI_FAKE_SIZE and STARPU_MPI_FAKE_RANK to allow simulating
    execution of just one MPI node.
  * Add STARPU_PERF_MODEL_HOMOGENEOUS_CUDA/OPENCL/MIC/SCC to share performance
    models between devices, making calibration much faster.
  * Add modular-heft-prio scheduler.
  * Add starpu_cublas_get_local_handle helper.
  * Add starpu_data_set_name, starpu_data_set_coordinates_array, and
    starpu_data_set_coordinates to describe data, and starpu_iteration_push and
    starpu_iteration_pop to describe tasks, for better offline traces analysis.
  * New function starpu_bus_print_filenames() to display filenames
    storing bandwidth/affinity/latency information, available through
    tools/starpu_machine_display -i
  * Add support for Ayudame version 2.x debugging library.
  * Add starpu_sched_ctx_get_workers_list_raw, much less costly than
    starpu_sched_ctx_get_workers_list
  * Add starpu_task_get_name and use it to warn about dmda etc. using
    a dumb policy when calibration is not finished
  * MPI: Add functions to test for cached values

Changes:
  * Fix performance regression of lws for small tasks.
  * Improve native Fortran support for StarPU

Small changes:
  * Fix type of data home node to allow users to pass -1 to define
    temporary data
  * Fix compatibility with simgrid 3.14

StarPU 1.2.0 (svn revision 18521)
==============================================

New features:
  * MIC Xeon Phi support
  * SCC support
  * New function starpu_sched_ctx_exec_parallel_code to execute a
    parallel code on the workers of the given scheduler context
  * MPI:
        - New internal communication system : a unique tag called
	  is now used for all communications, and a system
	  of hashmaps on each node which stores pending receives has been
	  implemented. Every message is now coupled with an envelope, sent
	  before the corresponding data, which allows the receiver to
	  allocate data correctly, and to submit the matching receive of
	  the envelope.
        - New function
   	  starpu_mpi_irecv_detached_sequential_consistency which
	  allows to enable or disable the sequential consistency for
	  the given data handle (sequential consistency will be
	  enabled or disabled based on the value of the function
	  parameter and the value of the sequential consistency
	  defined for the given data)
        - New functions starpu_mpi_task_build() and
  	  starpu_mpi_task_post_build()
        - New flag STARPU_NODE_SELECTION_POLICY to specify a policy for
          selecting a node to execute the codelet when several nodes
	  own data in W mode.
	- New selection node policies can be un/registered with the
	  functions starpu_mpi_node_selection_register_policy() and
	  starpu_mpi_node_selection_unregister_policy()
	- New environment variable STARPU_MPI_COMM which enables
	  basic tracing of communications.
	- New function starpu_mpi_init_comm() which allows to specify
	  a MPI communicator.
  * New STARPU_COMMUTE flag which can be passed along STARPU_W or STARPU_RW to
    let starpu commute write accesses.
  * Out-of-core support, through registration of disk areas as additional memory
    nodes. It can be enabled programmatically or through the STARPU_DISK_SWAP*
    environment variables.
  * Reclaiming is now periodically done before memory becomes full. This can
    be controlled through the STARPU_*_AVAILABLE_MEM environment variables.
  * New hierarchical schedulers which allow the user to easily build
    its own scheduler, by coding itself each "box" it wants, or by
    combining existing boxes in StarPU to build it. Hierarchical
    schedulers have very interesting scalability properties.
  * Add STARPU_CUDA_ASYNC and STARPU_OPENCL_ASYNC flags to allow asynchronous
    CUDA and OpenCL kernel execution.
  * Add STARPU_CUDA_PIPELINE and STARPU_OPENCL_PIPELINE to specify how
    many asynchronous tasks are submitted in advance on CUDA and
    OpenCL devices. Setting the value to 0 forces a synchronous
    execution of all tasks.
  * Add CUDA concurrent kernel execution support through
    the STARPU_NWORKER_PER_CUDA environment variable.
  * Add CUDA and OpenCL kernel submission pipelining, to overlap costs and allow
    concurrent kernel execution on Fermi cards.
  * New locality work stealing scheduler (lws).
  * Add STARPU_VARIABLE_NBUFFERS to be set in cl.nbuffers, and nbuffers and
    modes field to the task structure, which permit to define codelets taking a
    variable number of data.
  * Add support for implementing OpenMP runtimes on top of StarPU
  * New performance model format to better represent parallel tasks.
    Used to provide estimations for the execution times of the
    parallel tasks on scheduling contexts or combined workers.
  * starpu_data_idle_prefetch_on_node and
    starpu_idle_prefetch_task_input_on_node allow to queue prefetches to be done
    only when the bus is idle.
  * Make starpu_data_prefetch_on_node not forcibly flush data out, introduce
    starpu_data_fetch_on_node for that.
  * Add data access arbiters, to improve parallelism of concurrent data
    accesses, notably with STARPU_COMMUTE.
  * Anticipative writeback, to flush dirty data asynchronously before the
    GPU device is full. Disabled by default. Use STARPU_MINIMUM_CLEAN_BUFFERS
    and STARPU_TARGET_CLEAN_BUFFERS to enable it.
  * Add starpu_data_wont_use to advise that a piece of data will not be used
    in the close future.
  * Enable anticipative writeback by default.
  * New scheduler 'dmdasd' that considers priority when deciding on
    which worker to schedule
  * Add the capability to define specific MPI datatypes for
    StarPU user-defined interfaces.
  * Add tasks.rec trace output to make scheduling analysis easier.
  * Add Fortran 90 module and example using it
  * New StarPU-MPI gdb debug functions
  * Generate animated html trace of modular schedulers.
  * Add asynchronous partition planning. It only supports coherency through
    the home node of data for now.
  * Add STARPU_MALLOC_SIMULATION_FOLDED flag to save memory when simulating.
  * Include application threads in the trace.
  * Add starpu_task_get_task_scheduled_succs to get successors of a task.
  * Add graph inspection facility for schedulers.
  * New STARPU_LOCALITY flag to mark data which should be taken into account
    by schedulers for improving locality.
  * Experimental support for data locality in ws and lws.
  * Add a preliminary framework for native Fortran support for StarPU

Small features:
  * Tasks can now have a name (via the field const char *name of
    struct starpu_task)
  * New functions starpu_data_acquire_cb_sequential_consistency() and
    starpu_data_acquire_on_node_cb_sequential_consistency() which allows
    to enable or disable sequential consistency
  * New configure option --enable-fxt-lock which enables additional
    trace events focused on locks behaviour during the execution
  * Functions starpu_insert_task and starpu_mpi_insert_task are
    renamed in starpu_task_insert and starpu_mpi_task_insert. Old
    names are kept to avoid breaking old codes.
  * New configure option --enable-calibration-heuristic which allows
    the user to set the maximum authorized deviation of the
    history-based calibrator.
  * Allow application to provide the task footprint itself.
  * New function starpu_sched_ctx_display_workers() to display worker
    information belonging to a given scheduler context
  * The option --enable-verbose can be called with
    --enable-verbose=extra to increase the verbosity
  * Add codelet size, footprint and tag id in the paje trace.
  * Add STARPU_TAG_ONLY, to specify a tag for traces without making StarPU
    manage the tag.
  * On Linux x86, spinlocks now block after a hundred tries. This avoids
    typical 10ms pauses when the application thread tries to submit tasks.
  * New function char *starpu_worker_get_type_as_string(enum starpu_worker_archtype type)
  * Improve static scheduling by adding support for specifying the task
    execution order.
  * Add starpu_worker_can_execute_task_impl and
    starpu_worker_can_execute_task_first_impl to optimize getting the
    working implementations
  * Add STARPU_MALLOC_NORECLAIM flag to allocate without running a reclaim if
    the node is out of memory.
  * New flag STARPU_DATA_MODE_ARRAY for the function family
    starpu_task_insert to allow to define a array of data handles
    along with their access modes.
  * New configure option --enable-new-check to enable new testcases
    which are known to fail
  * Add starpu_memory_allocate and _deallocate to let the application declare
    its own allocation to the reclaiming engine.
  * Add STARPU_SIMGRID_CUDA_MALLOC_COST and STARPU_SIMGRID_CUDA_QUEUE_COST to
    disable CUDA costs simulation in simgrid mode.
  * Add starpu_task_get_task_succs to get the list of children of a given
    task.
  * Add starpu_malloc_on_node_flags, starpu_free_on_node_flags, and
    starpu_malloc_on_node_set_default_flags to control the allocation flags
    used for allocations done by starpu.
  * Ranges can be provided in STARPU_WORKERS_CPUID
  * Add starpu_fxt_autostart_profiling to be able to avoid autostart.
  * Add arch_cost_function perfmodel function field.
  * Add STARPU_TASK_BREAK_ON_SCHED, STARPU_TASK_BREAK_ON_PUSH, and
  STARPU_TASK_BREAK_ON_POP environment variables to debug schedulers.
  * Add starpu_sched_display tool.
  * Add starpu_memory_pin and starpu_memory_unpin to pin memory allocated
    another way than starpu_malloc.
  * Add STARPU_NOWHERE to create synchronization tasks with data.
  * Document how to switch between differents views of the same data.
  * Add STARPU_NAME to specify a task name from a starpu_task_insert call.
  * Add configure option to disable fortran --disable-fortran
  * Add configure option to give path for smpirun executable --with-smpirun
  * Add configure option to disable the build of tests --disable-build-tests
  * Add starpu-all-tasks debugging support
  * New function
    void starpu_opencl_load_program_source_malloc(const char *source_file_name, char **located_file_name, char **located_dir_name, char **opencl_program_source)
    which allocates the pointers located_file_name, located_dir_name
    and opencl_program_source.
  * Add submit_hook and do_schedule scheduler methods.
  * Add starpu_sleep.
  * Add starpu_task_list_ismember.
  * Add _starpu_fifo_pop_this_task.
  * Add STARPU_MAX_MEMORY_USE environment variable.
  * Add starpu_worker_get_id_check().
  * New function starpu_mpi_wait_for_all(MPI_Comm comm) that allows to
    wait until all StarPU tasks and communications for the given
    communicator are completed.
  * New function starpu_codelet_unpack_args_and_copyleft() which
    allows to copy in a new buffer values which have not been unpacked by
    the current call
  * Add STARPU_CODELET_SIMGRID_EXECUTE flag.
  * Add STARPU_CODELET_SIMGRID_EXECUTE_AND_INJECT flag.
  * Add STARPU_CL_ARGS flag to starpu_task_insert() and
    starpu_mpi_task_insert() functions call

Changes:
  * Data interfaces (variable, vector, matrix and block) now define
    pack und unpack functions
  * StarPU-MPI: Fix for being able to receive data which have not yet
    been registered by the application (i.e it did not call
    starpu_data_set_tag(), data are received as a raw memory)
  * StarPU-MPI: Fix for being able to receive data with the same tag
    from several nodes (see mpi/tests/gather.c)
  * Remove the long-deprecated cost_model fields and task->buffers field.
  * Fix complexity of implicit task/data dependency, from quadratic to linear.

Small changes:
  * Rename function starpu_trace_user_event() as
    starpu_fxt_trace_user_event()
  * "power" is renamed into "energy" wherever it applies, notably energy
    consumption performance models
  * Update starpu_task_build() to set starpu_task::cl_arg_free to 1 if
    some arguments of type ::STARPU_VALUE are given.
  * Simplify performance model loading API
  * Better semantic for environment variables STARPU_NMIC and
    STARPU_NMICDEVS, the number of devices and the number of cores.
    STARPU_NMIC will be the number of devices, and STARPU_NMICCORES
    will be the number of cores per device.

StarPU 1.1.5 (svn revision xxx)
==============================================
The scheduling context release

  * Add starpu_memory_pin and starpu_memory_unpin to pin memory allocated
    another way than starpu_malloc.
  * Add starpu_task_wait_for_n_submitted() and
    STARPU_LIMIT_MAX_NSUBMITTED_TASKS/STARPU_LIMIT_MIN_NSUBMITTED_TASKS to
    easily control the number of submitted tasks by making task submission
    block.

StarPU 1.1.4 (svn revision 14856)
==============================================
The scheduling context release

New features:
  * Fix and actually enable the cache allocation.
  * Enable allocation cache in main RAM when STARPU_LIMIT_CPU_MEM is set by
    the user.
  * New MPI functions starpu_mpi_issend and starpu_mpi_issend_detached
    to send data using a synchronous and non-blocking mode (internally
    uses MPI_Issend)
  * New data access mode flag STARPU_SSEND to be set when calling
    starpu_mpi_insert_task to specify the data has to be sent using a
    synchronous and non-blocking mode
  * New environment variable STARPU_PERF_MODEL_DIR which can be set to
    specify a directory where to store performance model files in.
    When unset, the files are stored in $STARPU_HOME/.starpu/sampling
  * MPI:
      - New function starpu_mpi_data_register_comm to register a data
        with another communicator than MPI_COMM_WORLD
      - New functions starpu_mpi_data_set_rank() and starpu_mpi_data_set_tag()
        which call starpu_mpi_data_register_comm()

Small features:
  * Add starpu_memory_wait_available() to wait for a given size to become
    available on a given node.
  * New environment variable STARPU_RAND_SEED to set the seed used for random
    numbers.
  * New function starpu_mpi_cache_set() to enable or disable the
    communication cache at runtime
  * Add starpu_paje_sort which sorts Pajé traces.

Changes:
  * Fix complexity of implicit task/data dependency, from quadratic to linear.

StarPU 1.1.3 (svn revision 13450)
==============================================
The scheduling context release

New features:
  * One can register an existing on-GPU buffer to be used by a handle.
  * Add the starpu_paje_summary statistics tool.
  * Enable gpu-gpu transfers for matrices.
  * Let interfaces declare which transfers they allow with the can_copy
    methode.

Small changes:
  * Lock performance model files while writing and reading them to avoid
    issues on parallel launches, MPI runs notably.
  * Lots of build fixes for icc on Windows.

StarPU 1.1.2 (svn revision 13011)
==============================================
The scheduling context release

New features:
  * The reduction init codelet is automatically used to initialize temporary
    buffers.
  * Traces now include a "scheduling" state, to show the overhead of the
    scheduler.
  * Add STARPU_CALIBRATE_MINIMUM environment variable to specify the minimum
    number of calibration measurements.
  * Add STARPU_TRACE_BUFFER_SIZE environment variable to specify the size of
    the trace buffer.

StarPU 1.1.1 (svn revision 12638)
==============================================
The scheduling context release

New features:
  * MPI:
        - New variable STARPU_MPI_CACHE_STATS to print statistics on
   	  cache holding received data.
        - New function starpu_mpi_data_register() which sets the rank
  	  and tag of a data, and also allows to automatically clear
	  the MPI communication cache when unregistering the data. It
	  should be called instead of both calling
	  starpu_data_set_tag() and starpu_data_set_rank()
  * Use streams for all CUDA transfers, even initiated by CPUs.
  * Add paje traces statistics tools.
  * Use streams for GPUA->GPUB and GPUB->GPUA transfers.

Small features:
  * New STARPU_EXECUTE_ON_WORKER flag to specify the worker on which
    to execute the task.
  * New STARPU_DISABLE_PINNING environment variable to disable host memory
    pinning.
  * New STARPU_DISABLE_KERNELS environment variable to disable actual kernel
    execution.
  * New starpu_memory_get_total function to get the size of a memory node.
  * New starpu_parallel_task_barrier_init_n function to let a scheduler decide
    a set of workers without going through combined workers.

Changes:
  * Fix simgrid execution.
  * Rename starpu_get_nready_tasks_of_sched_ctx to starpu_sched_ctx_get_nready_tasks
  * Rename starpu_get_nready_flops_of_sched_ctx to starpu_sched_ctx_get_nready_flops
  * New functions starpu_pause() and starpu_resume()
  * New codelet specific_nodes field to specify explicit target nodes for data.
  * StarPU-MPI: Fix overzealous allocation of memory.
  * Interfaces: Allow interface implementation to change pointers at will, in
    unpack notably.

Small changes:
  * Use big fat abortions when one tries to make a task or callback
    sleep, instead of just returning EDEADLCK which few people will test
  * By default, StarPU FFT examples are not compiled and checked, the
    configure option --enable-starpufft-examples needs to be specified
    to change this behaviour.

StarPU 1.1.0 (svn revision 11960)
==============================================
The scheduling context release

New features:
  * OpenGL interoperability support.
  * Capability to store compiled OpenCL kernels on the file system
  * Capability to load compiled OpenCL kernels
  * Performance models measurements can now be provided explicitly by
    applications.
  * Capability to emit communication statistics when running MPI code
  * Add starpu_data_unregister_submit, starpu_data_acquire_on_node and
    starpu_data_invalidate_submit
  * New functionnality to wrapper starpu_insert_task to pass a array of
	data_handles via the parameter STARPU_DATA_ARRAY
  * Enable GPU-GPU direct transfers.
  * GCC plug-in
	- Add `registered' attribute
	- A new pass was added that warns about the use of possibly
	  unregistered memory buffers.
  * SOCL
        - Manual mapping of commands on specific devices is now
	  possible
        - SOCL does not require StarPU CPU tasks anymore. CPU workers
          are automatically disabled to enhance performance of OpenCL
          CPU devices
  * New interface: COO matrix.
  * Data interfaces: The pack operation of user-defined data interface
    defines a new parameter count which should be set to the size of
    the buffer created by the packing of the data.
  * MPI:
        - Communication statistics for MPI can only be enabled at
	  execution time by defining the environment variable
	  STARPU_COMM_STATS
        - Communication cache mechanism is enabled by default, and can
	  only be disabled at execution time by setting the
	  environment variable STARPU_MPI_CACHE to 0.
        - Initialisation functions starpu_mpi_initialize_extended()
  	  and starpu_mpi_initialize() have been made deprecated. One
	  should now use starpu_mpi_init(int *, char ***, int). The
	  last parameter indicates if MPI should be initialised.
        - Collective detached operations have new parameters, a
	  callback function and a argument. This is to be consistent
	  with the detached point-to-point communications.
        - When exchanging user-defined data interfaces, the size of
	  the data is the size returned by the pack operation, i.e
	  data with dynamic size can now be exchanged with StarPU-MPI.
  * Add experimental simgrid support, to simulate execution with various
    number of CPUs, GPUs, amount of memory, etc.
  * Add support for OpenCL simulators (which provide simulated execution time)
  * Add support for Temanejo, a task graph debugger
  * Theoretical bound lp output now includes data transfer time.
  * Update OpenCL driver to only enable CPU devices (the environment
        variable STARPU_OPENCL_ONLY_ON_CPUS must be set to a positive
	value when executing an application)
  * Add Scheduling contexts to separate computation resources
    	- Scheduling policies take into account the set of resources corresponding
	to the context it belongs to
	- Add support to dynamically change scheduling contexts
	(Create and Delete a context, Add Workers to a context, Remove workers from a context)
	- Add support to indicate to which contexts the tasks are submitted
  * Add the Hypervisor to manage the Scheduling Contexts automatically
    	- The Contexts can be registered to the Hypervisor
	- Only the registered contexts are managed by the Hypervisor
	- The Hypervisor can detect the initial distribution of resources of
	a context and constructs it consequently (the cost of execution is required)
    	- Several policies can adapt dynamically the distribution of resources
	in contexts if the initial one was not appropriate
	- Add a platform to implement new policies of redistribution
	of resources
  * Implement a memory manager which checks the global amount of
    memory available on devices, and checks there is enough memory
    before doing an allocation on the device.
  * Discard environment variable STARPU_LIMIT_GPU_MEM and define
    instead STARPU_LIMIT_CUDA_MEM and STARPU_LIMIT_OPENCL_MEM
  * Introduce new variables STARPU_LIMIT_CUDA_devid_MEM and
    STARPU_LIMIT_OPENCL_devid_MEM to limit memory per specific device
  * Introduce new variable STARPU_LIMIT_CPU_MEM to limit memory for
    the CPU devices
  * New function starpu_malloc_flags to define a memory allocation with
    constraints based on the following values:
    - STARPU_MALLOC_PINNED specifies memory should be pinned
    - STARPU_MALLOC_COUNT specifies the memory allocation should be in
      the limits defined by the environment variables STARPU_LIMIT_xxx
      (see above). When no memory is left, starpu_malloc_flag tries
      to reclaim memory from StarPU and returns -ENOMEM on failure.
  * starpu_malloc calls starpu_malloc_flags with a value of flag set
    to STARPU_MALLOC_PINNED
  * Define new function starpu_free_flags similarly to starpu_malloc_flags
  * Define new public API starpu_pthread which is similar to the
    pthread API. It is provided with 2 implementations: a pthread one
    and a Simgrid one. Applications using StarPU and wishing to use
    the Simgrid StarPU features should use it.
  * Allow to have a dynamically allocated number of buffers per task,
    and so overwrite the value defined --enable-maxbuffers=XXX
  * Performance models files are now stored in a directory whose name
    include the version of the performance model format. The version
    number is also written in the file itself.
    When updating the format, the internal variable
    _STARPU_PERFMODEL_VERSION should be updated. It is then possible
    to switch easily between differents versions of StarPU having
    different performance model formats.
  * Tasks can now define a optional prologue callback which is executed
    on the host when the task becomes ready for execution, before getting
    scheduled.
  * Small CUDA allocations (<= 4MiB) are now batched to avoid the huge
    cudaMalloc overhead.
  * Prefetching is now done for all schedulers when it can be done whatever
    the scheduling decision.
  * Add a watchdog which permits to easily trigger a crash when StarPU gets
    stuck.
  * Document how to migrate data over MPI.
  * New function starpu_wakeup_worker() to be used by schedulers to
    wake up a single worker (instead of all workers) when submitting a
    single task.
  * The functions starpu_sched_set/get_min/max_priority set/get the
    priorities of the current scheduling context, i.e the one which
    was set by a call to starpu_sched_ctx_set_context() or the initial
    context if the function has not been called yet.
  * Fix for properly dealing with NAN on windows systems

Small features:
  * Add starpu_worker_get_by_type and starpu_worker_get_by_devid
  * Add starpu_fxt_stop_profiling/starpu_fxt_start_profiling which permits to
    pause trace recording.
  * Add trace_buffer_size configuration field to permit to specify the tracing
    buffer size.
  * Add starpu_codelet_profile and starpu_codelet_histo_profile, tools which draw
    the profile of a codelet.
  * File STARPU-REVISION --- containing the SVN revision number from which
    StarPU was compiled --- is installed in the share/doc/starpu directory
  * starpu_perfmodel_plot can now directly draw GFlops curves.
  * New configure option --enable-mpi-progression-hook to enable the
    activity polling method for StarPU-MPI.
  * Permit to disable sequential consistency for a given task.
  * New macro STARPU_RELEASE_VERSION
  * New function starpu_get_version() to return as 3 integers the
    release version of StarPU.
  * Enable by default data allocation cache
  * New function starpu_perfmodel_directory() to print directory
    storing performance models. Available through the new option -d of
    the tool starpu_perfmodel_display
  * New batch files to execute StarPU applications under Microsoft
    Visual Studio (They are installed in path_to_starpu/bin/msvc)/
  * Add cl_arg_free, callback_arg_free, prologue_callback_arg_free fields to
    enable automatic free(cl_arg); free(callback_arg);
    free(prologue_callback_arg) on task destroy.
  * New function starpu_task_build
  * New configure options --with-simgrid-dir
    --with-simgrid-include-dir and --with-simgrid-lib-dir to specify
    the location of the SimGrid library

Changes:
  * Rename all filter functions to follow the pattern
    starpu_DATATYPE_filter_FILTERTYPE. The script
    tools/dev/rename_filter.sh is provided to update your existing
    applications to use new filters function names.
  * Renaming of diverse functions and datatypes. The script
    tools/dev/rename.sh is provided to update your existing
    applications to use the new names. It is also possible to compile
    with the pkg-config package starpu-1.0 to keep using the old
    names. It is however recommended to update your code and to use
    the package starpu-1.1.

  * Fix the block filter functions.
  * Fix StarPU-MPI on Darwin.
  * The FxT code can now be used on systems other than Linux.
  * Keep only one hashtable implementation common/uthash.h
  * The cache of starpu_mpi_insert_task is fixed and thus now enabled by
    default.
  * Improve starpu_machine_display output.
  * Standardize objects name in the performance model API
  * SOCL
    - Virtual SOCL device has been removed
    - Automatic scheduling still available with command queues not
      assigned to any device
    - Remove modified OpenCL headers. ICD is now the only supported
      way to use SOCL.
    - SOCL test suite is only run when environment variable
      SOCL_OCL_LIB_OPENCL is defined. It should contain the location
      of the libOpenCL.so file of the OCL ICD implementation.
  * Fix main memory leak on multiple unregister/re-register.
  * Improve hwloc detection by configure
  * Cell:
    - It is no longer possible to enable the cell support via the
      gordon driver
    - Data interfaces no longer define functions to copy to and from
      SPU devices
    - Codelet no longer define pointer for Gordon implementations
    - Gordon workers are no longer enabled
    - Gordon performance models are no longer enabled
  * Fix data transfer arrows in paje traces
  * The "heft" scheduler no longer exists. Users should now pick "dmda"
    instead.
  * StarPU can now use poti to generate paje traces.
  * Rename scheduling policy "parallel greedy" to "parallel eager"
  * starpu_scheduler.h is no longer automatically included by
    starpu.h, it has to be manually included when needed
  * New batch files to run StarPU applications with Microsoft Visual C
  * Add examples/release/Makefile to test StarPU examples against an
    installed version of StarPU. That can also be used to test
    examples using a previous API.
  * Tutorial is installed in ${docdir}/tutorial
  * Schedulers eager_central_policy, dm and dmda no longer erroneously respect
    priorities. dmdas has to be used to respect priorities.
  * StarPU-MPI: Fix potential bug for user-defined datatypes. As MPI
    can reorder messages, we need to make sure the sending of the size
    of the data has been completed.
  * Documentation is now generated through doxygen.
  * Modification of perfmodels output format for future improvements.
  * Fix for properly dealing with NAN on windows systems
  * Function starpu_sched_ctx_create() now takes a variable argument
    list to define the scheduler to be used, and the minimum and
    maximum priority values
  * The functions starpu_sched_set/get_min/max_priority set/get the
    priorities of the current scheduling context, i.e the one which
    was set by a call to starpu_sched_ctx_set_context() or the initial
    context if the function was not called yet.
  * MPI: Fix of the livelock issue discovered while executing applications
    on a CPU+GPU cluster of machines by adding a maximum trylock
    threshold before a blocking lock.

Small changes:
  * STARPU_NCPU should now be used instead of STARPU_NCPUS. STARPU_NCPUS is
	still available for compatibility reasons.
  * include/starpu.h includes all include/starpu_*.h files, applications
	therefore only need to have #include <starpu.h>
  * Active task wait is now included in blocked time.
  * Fix GCC plugin linking issues starting with GCC 4.7.
  * Fix forcing calibration of never-calibrated archs.
  * CUDA applications are no longer compiled with the "-arch sm_13"
    option. It is specifically added to applications which need it.
  * Explicitly name the non-sleeping-non-running time "Overhead", and use
    another color in vite traces.
  * Use C99 variadic macro support, not GNU.
  * Fix performance regression: dmda queues were inadvertently made
    LIFOs in r9611.

StarPU 1.0.3 (svn revision 7379)
==============================================

Changes:
  * Several bug fixes in the build system
  * Bug fixes in source code for non-Linux systems
  * Fix generating FXT traces bigger than 64MiB.
  * Improve ENODEV error detections in StarPU FFT

StarPU 1.0.2 (svn revision 7210)
==============================================

Changes:
  * Add starpu_block_shadow_filter_func_vector and an example.
  * Add tag dependency in trace-generated DAG.
  * Fix CPU binding for optimized CPU-GPU transfers.
  * Fix parallel tasks CPU binding and combined worker generation.
  * Fix generating FXT traces bigger than 64MiB.

StarPU 1.0.1 (svn revision 6659)
==============================================

Changes:
  * hwloc support. Warn users when hwloc is not found on the system and
	produce error when not explicitely disabled.
  * Several bug fixes
  * GCC plug-in
	- Add `#pragma starpu release'
	- Fix bug when using `acquire' pragma with function parameters
	- Slightly improve test suite coverage
	- Relax the GCC version check
  * Update SOCL to use new API
  * Documentation improvement.

StarPU 1.0.0 (svn revision 6306)
==============================================
The extensions-again release

New features:
  * Add SOCL, an OpenCL interface on top of StarPU.
  * Add a gcc plugin to extend the C interface with pragmas which allows to
	easily define codelets and issue tasks.
  * Add reduction mode to starpu_mpi_insert_task.
  * A new multi-format interface permits to use different binary formats
	on CPUs & GPUs, the conversion functions being provided by the
	application and called by StarPU as needed (and as less as
	possible).
  * Deprecate cost_model, and introduce cost_function, which is provided
	with the whole task structure, the target arch and implementation
	number.
  * Permit the application to provide its own size base for performance
	models.
  * Applications can provide several implementations of a codelet for the
	same architecture.
  * Add a StarPU-Top feedback and steering interface.
  * Permit to specify MPI tags for more efficient starpu_mpi_insert_task

Changes:
  * Fix several memory leaks and race conditions
  * Make environment variables take precedence over the configuration
        passed to starpu_init()
  * Libtool interface versioning has been included in libraries names
        (libstarpu-1.0.so, libstarpumpi-1.0.so,
        libstarpufft-1.0.so, libsocl-1.0.so)
  * Install headers under $includedir/starpu/1.0.
  * Make where field for struct starpu_codelet optional. When unset, its
	value will be automatically set based on the availability of the
	different XXX_funcs fields of the codelet.
  * Define access modes for data handles into starpu_codelet and no longer
	in starpu_task. Hence mark (struct starpu_task).buffers as
	deprecated, and add (struct starpu_task).handles and (struct
	starpu_codelet).modes
  * Fields xxx_func of struct starpu_codelet are made deprecated. One
	should use fields xxx_funcs instead.
  * Some types were renamed for consistency. when using pkg-config libstarpu,
        starpu_deprecated_api.h is automatically included (after starpu.h) to
        keep compatibility with existing software. Other changes are mentioned
        below, compatibility is also preserved for them.
        To port code to use new names (this is not mandatory), the
        tools/dev/rename.sh script can be used, and pkg-config starpu-1.0 should
        be used.
  * The communication cost in the heft and dmda scheduling strategies now
        take into account the contention brought by the number of GPUs. This
        changes the meaning of the beta factor, whose default 1.0 value should
        now be good enough in most case.

Small features:
  * Allow users to disable asynchronous data transfers between CPUs and
	GPUs.
  * Update OpenCL driver to enable CPU devices (the environment variable
	STARPU_OPENCL_ON_CPUS must be set to a positive value when
	executing an application)
  * struct starpu_data_interface_ops --- operations on a data
        interface --- define a new function pointer allocate_new_data
        which creates a new data interface of the given type based on
        an existing handle
  * Add a field named magic to struct starpu_task which is set when
        initialising the task. starpu_task_submit will fail if the
        field does not have the right value. This will hence avoid
        submitting tasks which have not been properly initialised.
  * Add a hook function pre_exec_hook in struct starpu_sched_policy.
        The function is meant to be called in drivers. Schedulers
        can use it to be notified when a task is about being computed.
  * Add codelet execution time statistics plot.
  * Add bus speed in starpu_machine_display.
  * Add a STARPU_DATA_ACQUIRE_CB which permits to inline the code to be
	done.
  * Add gdb functions.
  * Add complex support to LU example.
  * Permit to use the same data several times in write mode in the
	parameters of the same task.

Small changes:
  * Increase default value for STARPU_MAXCPUS -- Maximum number of
        CPUs supported -- to 64.
  * Add man pages for some of the tools
  * Add C++ application example in examples/cpp/
  * Add an OpenMP fork-join example.
  * Documentation improvement.


StarPU 0.9 (svn revision 3721)
==============================================
The extensions release

  * Provide the STARPU_REDUX data access mode
  * Externalize the scheduler API.
  * Add theoretical bound computation
  * Add the void interface
  * Add power consumption optimization
  * Add parallel task support
  * Add starpu_mpi_insert_task
  * Add profiling information interface.
  * Add STARPU_LIMIT_GPU_MEM environment variable.
  * OpenCL fixes
  * MPI fixes
  * Improve optimization documentation
  * Upgrade to hwloc 1.1 interface
  * Add fortran example
  * Add mandelbrot OpenCL example
  * Add cg example
  * Add stencil MPI example
  * Initial support for CUDA4

StarPU 0.4 (svn revision 2535)
==============================================
The API strengthening release

  * Major API improvements
    - Provide the STARPU_SCRATCH data access mode
    - Rework data filter interface
    - Rework data interface structure
    - A script that automatically renames old functions to accomodate with the new
      API is available from https://scm.gforge.inria.fr/svn/starpu/scripts/renaming
      (login: anonsvn, password: anonsvn)
  * Implement dependencies between task directly (eg. without tags)
  * Implicit data-driven task dependencies simplifies the design of
    data-parallel algorithms
  * Add dynamic profiling capabilities
    - Provide per-task feedback
    - Provide per-worker feedback
    - Provide feedback about memory transfers
  * Provide a library to help accelerating MPI applications
  * Improve data transfers overhead prediction
    - Transparently benchmark buses to generate performance models
    - Bind accelerator-controlling threads with respect to NUMA locality
  * Improve StarPU's portability
    - Add OpenCL support
    - Add support for Windows

StarPU 0.2.901 aka 0.3-rc1 (svn revision 1236)
==============================================
The asynchronous heterogeneous multi-accelerator release

  * Many API changes and code cleanups
    - Implement starpu_worker_get_id
    - Implement starpu_worker_get_name
    - Implement starpu_worker_get_type
    - Implement starpu_worker_get_count
    - Implement starpu_display_codelet_stats
    - Implement starpu_data_prefetch_on_node
    - Expose the starpu_data_set_wt_mask function
  * Support nvidia (heterogeneous) multi-GPU
  * Add the data request mechanism
    - All data transfers use data requests now
    - Implement asynchronous data transfers
    - Implement prefetch mechanism
    - Chain data requests to support GPU->RAM->GPU transfers
  * Make it possible to bypass the scheduler and to assign a task to a specific
    worker
  * Support restartable tasks to reinstanciate dependencies task graphs
  * Improve performance prediction
    - Model data transfer overhead
    - One model is created for each accelerator
  * Support for CUDA's driver API is deprecated
  * The STARPU_WORKERS_CUDAID and STARPU_WORKERS_CPUID env. variables make it possible to
    specify where to bind the workers
  * Use the hwloc library to detect the actual number of cores

StarPU 0.2.0 (svn revision 1013)
==============================================
The Stabilizing-the-Basics release

  * Various API cleanups
  * Mac OS X is supported now
  * Add dynamic code loading facilities onto Cell's SPUs
  * Improve performance analysis/feedback tools
  * Application can interact with StarPU tasks
    - The application may access/modify data managed by the DSM
    - The application may wait for the termination of a (set of) task(s)
  * An initial documentation is added
  * More examples are supplied


StarPU 0.1.0 (svn revision 794)
==============================================
First release.

Status:
 * Only supports Linux platforms yet
 * Supported architectures
   - multicore CPUs
   - NVIDIA GPUs (with CUDA 2.x)
   - experimental Cell/BE support

Changes:
 * Scheduling facilities
   - run-time selection of the scheduling policy
   - basic auto-tuning facilities
 * Software-based DSM
   - transparent data coherency management
   - High-level expressive interface


# Local Variables:
# mode: text
# coding: utf-8
# ispell-local-dictionary: "american"
# End: