Skip to content
Permalink
Branch: master
Commits on Jun 12, 2019
  1. Feature #1436: Trace CcdCallFnAfter() causality (#2296)

    rbuch committed Jun 12, 2019
    Creates dummy message sends to trace invocations of CcdCallFnAfter back to the
    original entry method.
Commits on Jun 10, 2019
  1. ZC RO Bcast: QD Support

    nitbhat authored and evan-charmworks committed Jun 3, 2019
    Change-Id: Ie6019ed6cd7b960241d69eb2a9a9d2f5fcf07157
  2. Fix typo that removed GNI large msg registration (#2304)

    jcphill committed Jun 10, 2019
Commits on Jun 7, 2019
  1. bugfix #1473: fix ancient ordering bug in benchmark

    ericjbohm authored and evan-charmworks committed Jun 7, 2019
    Test has always incorrectly ordered the local call
    to operationFinished in the timed allocation case
    to be before the remote message send. Whereas it should
    always be after, as it is in the untimed allocation case.
    Bug revealed by Phil's introduction of size testing in 2017.
    Falsely attributed to verbs, as this bug is invariant
    with network layer.
Commits on Jun 6, 2019
  1. Fix extern declarations of numZerocopyROops

    nitbhat authored and evan-charmworks committed Jun 6, 2019
    This fixes the multicore-win-x86_64 and mpi-win-x86_64-smp autobuild
    failures on 06/05/2019 by keeping the windows linker happy.
  2. Remove divide by 0 errors in perfReport

    chin123 authored and evan-charmworks committed May 20, 2019
    This removes all the `inf`s in the generated report.
  3. Remove unused broadcast preproc directive

    epmikida authored and evan-charmworks committed May 31, 2019
    This hasn't been touched for the most part since 2012, no longer
    works, and should be subsumed by current broadcast work.
Commits on Jun 5, 2019
  1. bugfix #2268: undefined CthScheduled

    ebohm authored and evan-charmworks committed May 29, 2019
    When the CthScheduled function is inlined in the C file it cannot
    be accessed via external linkage.  Removing the pointless (outside
    of a header in C++) inline solves the problem.
  2. ZC RO Bcast: Ensure that numZerocopyROops gets the correct value on n…

    nitbhat committed Jun 4, 2019
    …on-0 threads
    
    This is required so that the _initDone is called correctly on non-0 threads.
  3. ZC RO Bcast: Wait for Rget completion on secondary worker threads

    nitbhat committed Jun 3, 2019
    The primary worker thread (rank 0) performs the Rget, the secondary
    worker threads (rank != 0) receive the RODataMsg. However, previously
    it would directly call _initDone. With this fix, it waits for the Rget
    to be completed on the primary worker thread.
  4. Add UCX machine layer (#2249)

    brminich authored and nitbhat committed Jun 5, 2019
    This commit adds a new UCX machine layer, which is LRTS based.
    It is a potential replacement for the Verbs machine layer and other
    networking layers depending on the supported UCX networking backends.
    
    In this commit, the ZC API implementation has been disabled because
    of some changes required in the LRTS API to support network de-registration
    in the UCX layer.
  5. AMPI: Fail compilation when unable to get binary path (#2282)

    matthiasdiener committed Jun 5, 2019
    In this case, any AMPI application would abort anyway on startup, so better to fail at AMPI compilation already.
Commits on Jun 4, 2019
  1. ZC API: set isRegistered after registration/de-registration

    nitbhat committed Jun 4, 2019
    This commit also adds the code for de-registration of source
    buffers for Bcast operations implemented using reverse operations
    when the buffer information object's mode is CK_BUFFER_UNREG.
  2. ZC Bcast Post API: Fix verbs-smp crash on intermediate nodes (#2289)

    nitbhat committed Jun 4, 2019
    This fixes the verbs-smp autobuild failures seen in 
    tests/charm++/zerocopy/zc_post_modify_size.
Commits on Jun 3, 2019
  1. Doc: Fix typo in quickstart and clarify AMPI (#2292)

    matthiasdiener committed Jun 3, 2019
  2. ZC RO Bcast: Send message to CmiNodeFirst(t.parent) instead of t.pare…

    nitbhat committed Jun 3, 2019
    …nt (#2290)
    
    t.parent represents the logical node, which is the parent of my node.
    However, CmiSyncSendAndFree expects a PE for its first argument and for that
    reason, the ROChildCompletionMsg should be sent to the first PE of the parent.
Commits on May 30, 2019
  1. PAPI: dynamically check for availability of highest cache level even…

    chin123 authored and rbuch committed May 30, 2019
    …ts to add (#2274)
  2. Small README fixes (#2283)

    matthiasdiener committed May 30, 2019
    * Small README fixes
    
    * Whitespace cleanups
Commits on May 29, 2019
  1. AMPI: use Ssend protocol for Node-local sends to avoid intermediate copy

    stwhite91 authored and evan-charmworks committed May 29, 2017
    In the payload of the Ssend message we serialize an SsendInfo object
    that contains the sender's information, which can be used by the recv'er
    if in the same logical node (process) to directly copy the real message
    payload directly from the src buffer.
    
    Also change the existing AMPI_LOCAL_THRESHOLD to AMPI_PE_LOCAL_THRESHOLD
    to distinguish it from the new AMPI_NODE_LOCAL_THRESHOLD variable.
    
    This means that all pt2pt messages with size greater than or equal to
    the AMPI_NODE_LOCAL_THRESHOLD are now 'zero-copy' if in the same node.
    
    Change-Id: Iebe9972b630fe17f4cc5c2a7ae0aa9f4b54b295b
  2. AMPI: use Ssend protocol for PE-local sends to avoid intermediate copy

    stwhite91 authored and evan-charmworks committed May 26, 2017
    PE-local sends are done inline if the message is in-order and expected.
    If not in-order and expected, then block the sender on the matching recv.
    
    Build AMPI with -DAMPI_LOCAL_IMPL=0 to disable all local/inline sends.
    
    Build AMPI with -DAMPI_LOCAL_THRESHOLD_DEFAULT=N and/or set the
    environment variable AMPI_LOCAL_THRESHOLD=N, where N is the number
    of bytes below which AMPI will not block the sender.
    
    Change-Id: I244e400fa8e8750c2a999109572ca12124f4d672
  3. ZC Bcast Post API: Deregister buffers after completion of memcpy and …

    nitbhat authored and evan-charmworks committed May 23, 2019
    …CMA transfers
    
    Previously, buffers were de-registered after completion of RDMA transfers.
    However, when buffers are declared with CK_BUFFER_REG (and PREREG for verbs
    and ucx), it is required to de-register these buffers after the completion
    of the memcpy and CMA transfers to avoid leaking of pinned memory.
Commits on May 28, 2019
  1. AMPI: implement (I)Ssend's using a synchronous rendezvous protocol

    stwhite91 authored and evan-charmworks committed May 28, 2019
    Instead of sending the message payload eagerly and then getting an ack
    when the recv'er actually matches it, now we use a rendezvous protocol,
    only sending the payload after it has been matched.
    
    The initial sync message (to ampi::genericSync) has no payload and its
    refnum is the index of the sender's request. When the sync message is
    matched by the recv'er, we delete that placeholder msg and send a
    message back to the sender (to ampi::ssendAck) to get the payload back
    in a normal AmpiMsg (via ampi::generic or ampi::genericRdma).
    
    Change-Id: I6a8db194f63321a2247f707042d624d39fa3bef8
  2. ZC P2P API: Deregister buffers after completion of memcpy and CMA tra…

    nitbhat committed May 23, 2019
    …nsfers
    
    Previously, buffers were de-registered after completion of RDMA transfers.
    However, when buffers are declared with CK_BUFFER_REG (and PREREG for verbs
    and ucx), it is required to de-register these buffers after the completion
    of the memcpy and CMA transfers to avoid leaking of pinned memory.
Commits on May 24, 2019
  1. ZC Bcast Post API: Fix multicore bug when bcast root is not rank 0

    nitbhat committed May 24, 2019
  2. ZC Post API Test: Rename identifiers for successful compilation on Wi…

    nitbhat committed May 24, 2019
    …ndows
    
    Previously, the group was named 'grp1', which was a global namespace
    macro on Windows. The group is renamed from 'grp1' to 'grp'. For
    consistency, the chare array was renamed from 'arr1' to 'arr' and
    the nodegroup was renamed from 'nodegrp1' to 'nodegrp'.
  3. Documentation #2077: RAJA interoperability (#2269)

    minitu committed May 24, 2019
    * Documentation #2077: RAJA interoperability
    
    * Fix misspelling in doc/charm++/manual.rst
Commits on May 23, 2019
  1. AMPI: Clean up pointer refreshing after potential migration

    evan-charmworks committed Apr 15, 2019
    - Replace calls to thread->suspend() and thread->schedule() with new
    ampi::block() and ampi::yield() that refresh the parent pointer
    - Use static member functions to enforce non-access of `this` after
    migration
    - Replace missed `this` accesses
    - Add __attribute__ ((warn_unused_result)) where helpful
    
    Change-Id: Ifacc90178f0cfa81f677b364922bb2cfd2247652
Older
You can’t perform that action at this time.