<html>
  <head>
    <meta content="text/html; charset=windows-1252"
      http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">Hi,<br>
      <br>
      Erik and Tomas asked me to separate the G1 changes which was
      pushed separately in JDK-8064473.<br>
      Erik also had some further comments offline that I have taken care
      of.<br>
      <br>
      Here is the new webrev,
      cr.openjdk.java.net/~sfriberg/8055845/webrev.06<br>
      <br>
      Thanks,<br>
      Staffan<br>
      <br>
      <br>
      On 11/06/2014 01:53 PM, Bengt Rutisson wrote:<br>
    </div>
    <blockquote cite="mid:545B6F58.2000907@oracle.com" type="cite">
      <meta content="text/html; charset=windows-1252"
        http-equiv="Content-Type">
      <br>
      <br>
      <div class="moz-cite-prefix">On 2014-11-06 14:00, Staffan Friberg
        wrote:<br>
      </div>
      <blockquote cite="mid:545B70D2.9080008@oracle.com" type="cite">
        <meta content="text/html; charset=windows-1252"
          http-equiv="Content-Type">
        <div class="moz-cite-prefix">Good catch, fixed.<br>
          <br>
                // Too large; allocate the object individually.<br>
                obj = sp->par_allocate(word_sz);<br>
                if (obj != NULL) {<br>
                 
          gc_tracer()->report_promotion_outside_plab_event(old,
          word_sz, age, false);<br>
                }<br>
          <br>
          Please let me know if anyone wants a full new webrev with
          this.<br>
        </div>
      </blockquote>
      <br>
      <br>
      Looks good. Reviewed.<br>
      <br>
      Bengt<br>
      <br>
      <blockquote cite="mid:545B70D2.9080008@oracle.com" type="cite">
        <div class="moz-cite-prefix"> <br>
          Thanks,<br>
          Staffan<br>
          <br>
          <br>
          On 11/06/2014 12:01 PM, Bengt Rutisson wrote:<br>
        </div>
        <blockquote cite="mid:545B5520.7010407@oracle.com" type="cite">
          <meta content="text/html; charset=windows-1252"
            http-equiv="Content-Type">
          <br>
          Hi Staffan,<br>
          <br>
          <div class="moz-cite-prefix">On 2014-11-06 11:12, Staffan
            Friberg wrote:<br>
          </div>
          <blockquote cite="mid:545B4986.1050700@oracle.com" type="cite">
            <meta content="text/html; charset=windows-1252"
              http-equiv="Content-Type">
            <div class="moz-cite-prefix">Hi,<br>
              <br>
              After further off list discussion it was decided to keep
              the gc_tracer in par_promote as is.<br>
              <br>
              I have uploaded a new webrev, <a moz-do-not-send="true"
                class="moz-txt-link-freetext"
                href="http://cr.openjdk.java.net/%7Esfriberg/8055845/webrev.05">http://cr.openjdk.java.net/~sfriberg/8055845/webrev.05</a><br>
              <br>
              The main change here is a rewrite of the G1 code which is
              cleaner and also reuses the read age. By sending the
              markOop down through the call we can now trust the read
              age and do not need to reread it when incrementing which
              improves the YC performance slightly as it avoids the
              rather complex bit extraction.<br>
            </div>
          </blockquote>
          <br>
          <br>
          Looks good to me. One detail in parNewGeneration.cpp:<br>
          <br>
           274     } else {<br>
           275       // Too large; allocate the object individually.<br>
           276      
          gc_tracer()->report_promotion_outside_plab_event(old,
          word_sz, age, false);<br>
           277       obj = sp->par_allocate(word_sz);<br>
           278     }<br>
          <br>
          Seems like par_allocate() return NULL. Maybe we should check
          that before reporting the event. Similarly to what you do in
          the other GCs, for example G1:<br>
          <br>
          g1Allocator.cpp<br>
          <br>
           141     obj = _g1h->par_allocate_during_gc(purpose,
          word_sz, context);<br>
           142     if (obj != NULL<br>
           143         &&
          _g1h->_gc_tracer_stw->should_report_promotion_outside_plab_event())



          {<br>
           144       bool tenured =
          _g1h->heap_region_containing_raw(obj)->is_old();<br>
           145      
          _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old,
          word_sz, <br>
           146                                                                



          age, tenured);<br>
           147     }<br>
          <br>
          Thanks,<br>
          Bengt<br>
          <br>
          <br>
          <br>
          <blockquote cite="mid:545B4986.1050700@oracle.com" type="cite">
            <div class="moz-cite-prefix"> <br>
              Cheers,<br>
              Staffan<br>
              <br>
              On 09/15/2014 02:06 PM, Bengt Rutisson wrote:<br>
            </div>
            <blockquote cite="mid:5416D64C.9020207@oracle.com"
              type="cite">
              <meta content="text/html; charset=windows-1252"
                http-equiv="Content-Type">
              <br>
              Hi Staffan,<br>
              <br>
              psPromotionManager.inline.hpp<br>
              <br>
              I think the PSPromotionManager::copy_to_survivor_space()
              might send multiple events. If the allocation to the young
              gen fails we will fall through to do an old gen
              allocation. But we send the events before we realize that
              the allocation has failed, i.e. new_obj is NULL.<br>
              <br>
              I talked to Erik a bit about how to handle the gc_tracer
              in par_promote. He'll get back to you with some thoughts
              on that.<br>
              <br>
              Thanks,<br>
              Bengt<br>
              <br>
              <br>
              <div class="moz-cite-prefix">On 2014-09-06 00:20, Staffan
                Friberg wrote:<br>
              </div>
              <blockquote cite="mid:540A3730.1060603@oracle.com"
                type="cite">
                <meta content="text/html; charset=windows-1252"
                  http-equiv="Content-Type">
                <div class="moz-cite-prefix">Hi,<br>
                  <br>
                  I have uploaded a new webrev here,
                  cr.openjdk.java.net/~sfriberg/8055845/webrev.03<br>
                  <br>
                  It contains several changes<br>
                  <br>
                      - Split event into two events
                  (PromoteObjectInNewPLAB, PromoteObjectOutsidePLAB)<br>
                      - Moved events to
                  "vm/gc/detailed/PromoteObject..."<br>
                      - Supporting ParNew+CMS and ParNew+SerialOld
                  tenuring<br>
                           - Not sure if the way I do it with passing
                  the ParNewTracer is the best solution, please let me
                  know if you have an idea how to improve it<br>
                      - Simplified the G1 code to avoid sending age and
                  having a single call site<br>
                      - Fixed so that the generated event has size
                  information in bytes rather than words<br>
                  <br>
                  Thanks for offline comments and suggestions from
                  Dmitry and Thomas.<br>
                  <br>
                  Cheers,<br>
                  Staffan<br>
                  <br>
                  On 08/29/2014 03:32 PM, Staffan Friberg wrote:<br>
                </div>
                <blockquote cite="mid:5400FF6F.5040203@oracle.com"
                  type="cite">
                  <meta content="text/html; charset=windows-1252"
                    http-equiv="Content-Type">
                  <div class="moz-cite-prefix">Hi Erik,<br>
                    <br>
                    On 08/28/2014 11:34 PM, Erik Helin wrote:<br>
                  </div>
                  <blockquote cite="mid:54001EFF.3030802@oracle.com"
                    type="cite">(it seems like we lost <a
                      moz-do-not-send="true"
                      class="moz-txt-link-abbreviated"
                      href="mailto:hotspot-gc-dev@openjdk.java.net">hotspot-gc-dev@openjdk.java.net</a>
                    somewhere in this thread, I'm adding it back) <br>
                    <br>
                    On 2014-08-28 23:15, Staffan Friberg wrote: <br>
                    <blockquote type="cite">Hi Erik, <br>
                      <br>
                      Thanks for the comments. <br>
                      <blockquote type="cite">- Aren't the events for
                        promotion to the tenured generation (SerialOld)
                        <br>
                          and the CMS generation missing? <br>
                      </blockquote>
                      The reason for leaving out these two, Serial
                      completely and CMS <br>
                      promotion, was due to that neither as far as I
                      understand make use of <br>
                      PLABs. <br>
                    </blockquote>
                    <br>
                    I might be wrong here, but looking at the function
                    TenuredGeneration::par_promote (in
                    tenuredGeneration.cpp) it looks to me like SerialOld
                    is using PLABs when ParNew is promoting objects from
                    young to old. <br>
                    <br>
                    As for CMS, looking at
                    ConcurrentMarkSweepGeneration::par_promote (in
                    concurrentMarkSweepGeneration.cpp) it seems like
                    each CMSParGCThreadState has a CFLS_LAB
                    (CompactibleFreeListSpace Local Allocation Buffer)
                    that is a thread-local allocation buffer. See
                    compactibleFreeListSpace.{hpp,cpp} for more details.
                    <br>
                    <br>
                    Given this, I think we should add events for Serial
                    and CMS as well. <br>
                    <br>
                  </blockquote>
                  <br>
                  Ok I see what you mean with CMS, basically the
                  equivalent to getting a PLAB would be when we refill
                  the CFLS_LAB in the alloc function. It still does the
                  allocation to a small memory (~ size of object) area
                  from the freelist, but at least we did extra work to
                  refill the local CFLS_LAB. Need to do some analysis to
                  see how often we refill so the overhead doesn't get
                  too high.<br>
                  The only issue I run into is how I can in a nice way
                  get access to the ParNewTracer from ParNewGeneration
                  to call the report function. Let's sync up next week
                  and see how it can be solved.<br>
                  <br>
                  The tenured GC requires something similar to be
                  supported, however since ParNewGC is deprecated for
                  usage without CMS in JDK 8 we might want to skip that
                  combination.<br>
                  <br>
                  <br>
                  <blockquote cite="mid:54001EFF.3030802@oracle.com"
                    type="cite"> <br>
                    On 2014-08-28 23:15, Staffan Friberg wrote: <br>
                    <blockquote type="cite">
                      <blockquote type="cite">- Would it make sense to
                        differentiate, either by separate events or by <br>
                          a field in a event, between promotions to
                        to-space and to the old <br>
                          generation? <br>
                        - The are two events for TLAB allocations, <br>
                          java/object_alloc_in_new_TLAB and
                        java/object_alloc_outside_TLAB. <br>
                          What do you think about using two events for
                        PLAB allocations as well: <br>
                          - java/object_alloc_in_new_PLAB <br>
                          - java/object_alloc_outside_PLAB <br>
                      </blockquote>
                      I think this is a matter of taste and probably how
                      similar we want the <br>
                      event to be to the existing allocation event. I
                      personally prefer a <br>
                      single event but if the GC team and serviceability
                      team feel it would be <br>
                      better to have two I can certainly rewrite. The
                      reason for me preferring <br>
                      a single event is just ease of analysis, you can
                      easily filter a list of <br>
                      events on a field, it is harder to merge two
                      different events with <br>
                      different fields and work with them in an straight
                      forward fashion. <br>
                      <br>
                      Any general preference for having a single or
                      multiple events? <br>
                    </blockquote>
                    <br>
                    I would prefer to have two events in this case and
                    try to follow the existing allocation events as much
                    as possible (both in naming and in style). Keeping
                    the tenured field (I missed it the first time I read
                    the patch) is good. <br>
                    <br>
                  </blockquote>
                  Yes, tenured would be independent of having two
                  events, only PLAB size and directAllocation would be
                  affected when having two event types.<br>
                  <br>
                  <b>Erik Gahlin</b>, any preference from your end?<br>
                  <br>
                  <br>
                  <br>
                  <blockquote cite="mid:54001EFF.3030802@oracle.com"
                    type="cite">On 2014-08-28 23:15, Staffan Friberg
                    wrote: <br>
                    <blockquote type="cite">
                      <blockquote type="cite">- In PSPromotionManager,
                        instead of utilizing the C++ friendship with <br>
                          PSScavenge, should we add a getter function
                        for the gc_tracer? <br>
                      </blockquote>
                      Created a getter function. <br>
                    </blockquote>
                    <br>
                    Thanks! If you make report_promotion_sample const,
                    then the getter can return a const
                    ParallelScavengeTracer*, right? <br>
                    <br>
                  </blockquote>
                  Done<br>
                  <br>
                  //Staffan<br>
                </blockquote>
                <br>
              </blockquote>
              <br>
            </blockquote>
            <br>
          </blockquote>
          <br>
        </blockquote>
        <br>
      </blockquote>
      <br>
    </blockquote>
    <br>
  </body>
</html>