<html>
  <head>
    <meta content="text/html; charset=utf-8" http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    <p>Nezih<br>
    </p>
    <br>
    <div class="moz-cite-prefix">On 05/01/2017 10:42 AM, nezih yigitbasi
      wrote:<br>
    </div>
    <blockquote
cite="mid:CALur2hUgpbNUiFL8DW2Od=xQj-4A0FFfRkF-8CaVEHnTwhEUQA@mail.gmail.com"
      type="cite">
      <div dir="ltr">Hi,
        <div>We have THP enabled for memory regions that explicitly
          request huge pages using madvise.  We don't have swapping
          enabled. We run Presto (a distributed SQL engine, <a
            moz-do-not-send="true" href="http://prestodb.io">http://prestodb.io</a>)
          on Java 1.8.0_112 and the machine has ~250GB memory and 12
          physical cores.</div>
        <div>I have some followup questions:</div>
        <div>- I was always thinking that remembered sets were updated
          in the "Update RS" phase, can you please elaborate on the
          difference between rset update vs. object copy phases in terms
          of the rset processing that is being done in each?</div>
      </div>
    </blockquote>
    When the application allocates objects, the remember set is updated
    concurrently with the application. What can not be updated during
    the concurrent stage, will be updated during the pause 'Update RS'.
    But after evacuating the objects, we also need to fix the pointers
    of the objects that are moved. <br>
    <blockquote
cite="mid:CALur2hUgpbNUiFL8DW2Od=xQj-4A0FFfRkF-8CaVEHnTwhEUQA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div>- Is high system cpu time a good indicator for high
          contention?</div>
      </div>
    </blockquote>
    High system cpu time indicates some work in the OS kernel level.<br>
    <blockquote
cite="mid:CALur2hUgpbNUiFL8DW2Od=xQj-4A0FFfRkF-8CaVEHnTwhEUQA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div><br>
        </div>
        <div>Thanks!</div>
        <div>Nezih</div>
        <div class="gmail_extra"><br>
          <div class="gmail_quote">2017-04-28 15:44 GMT-07:00 <a
              moz-do-not-send="true" href="mailto:yu.zhang@oracle.com">yu.zhang@oracle.com</a>
            <span dir="ltr"><<a moz-do-not-send="true"
                href="mailto:yu.zhang@oracle.com" target="_blank">yu.zhang@oracle.com</a>></span>:<br>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
              0.8ex;border-left:1px solid
              rgb(204,204,204);padding-left:1ex">
              <div bgcolor="#FFFFFF">
                <p>Hi, Nezih,</p>
                I agree with your estimation. There is not much
                promotion going on.<br>
                <br>
                Though I do not see high system cpu, if you are on a
                Linux system, can you verify that the OS Transparent
                Huge Page is off?<br>
                <br>
                Remember set is the data structure that keep track of
                pointers from other regions to the region we are
                operating on. If there are pointers from old regions to
                the objects in the young regions, we can not collect
                that object.<br>
                After a gc pause, the live objects are evacuated to
                either the survivor or old gen regions. Then we need to
                update the Remember set. That is what I was referring
                to.<br>
                It is part of object copy.<br>
                <br>
                Like Kirk mentioned, the parallelism of the gc threads
                seems fine. Charlie.Hunt suggested maybe the object
                graph is very deep so that the gc threads can not steal
                work and end up spinning.<br>
                <br>
                Some things we can try:<br>
                -XX:MaxGCPauseMillis=<200> Maybe increase it to
                400<br>
                Maybe try to reduce the -XX:ParallelGCThreads<br>
                <br>
                <pre style="box-sizing:border-box;font-family:sfmono-regular,consolas,"liberation mono",menlo,courier,monospace;font-size:13.6px;font-style:normal;font-variant-ligatures:normal;font-variant-caps:normal;font-weight:normal;font-stretch:normal;line-height:1.45;word-wrap:normal;padding:16px;overflow:auto;background-color:rgb(246,248,250);border-radius:3px;color:rgb(36,41,46);letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;word-spacing:0px;margin-top:0px;margin-bottom:0px"><code style="box-sizing:border-box;font-family:sfmono-regular,consolas,"liberation mono",menlo,courier,monospace;font-size:13.6px;padding:0px;margin:0px;background:transparent;border-radius:3px;word-break:normal;white-space:pre-wrap;border:0px;display:inline;overflow:visible;line-height:inherit;word-wrap:normal">-XX:+G1SummarizeRSetStats
-XX:<wbr>G1SummarizeRSetStatsPeriod=10</code></pre>
    

    It is very expensive to print those statistics, for now we know
    there is a lot of coarsening. We can make
    G1SumarizeRSetStatsPeriod=100

    

    The RSet footprint is so big,  increasing <span style="font-family:georgia;font-size:18px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;float:none;display:inline">-XX:G1RSetRegionEntries
      will get rid of coarsening but make the memory footprint bigger.
      We can delay this for now.

      

      What is your hardware/software configuration? cpu/memory/cores?
      There is no swapping, right?

      

      Thanks<span class="gmail-HOEnZb"><font color="#888888">

      Jenny

      

    </font></span></span><div><div class="gmail-h5">
    <div class="gmail-m_2983664130649143190moz-cite-prefix">On 04/27/2017 08:52 PM, nezih yigitbasi
      wrote:

    </div>
    <blockquote type="cite">
      <div dir="ltr">My first question is incorrect actually, so I am
        giving a better example to rephrase my question. At time
        "2017-04-26T03:12:24.259-0700"  there is a young GC that
        took 35.47 s, where object copy took 28983.4 ms. In that event I
        see the following log:
        <div>

        </div>
        <div>
          <div>   [Eden: 7168.0M(7168.0M)->0.0B(7168.<wbr>0M) Survivors:
            1024.0M->1024.0M Heap: 153.4G(160.0G)->149.4G(160.0G)<wbr>]</div>
          <div>

          </div>
          <div>My interpretation for this is, ~4GB of garbage was
            collected from heap in total and we see that eden usage goes
            down by ~7GB, this means ~3GB of the eden was live objects.
            Is this interpretation correct? If it is, how come copying
            over 3GB takes ~29s? In your answer you said "most of the
            object copy time is dealing with the Remember Set", can you
            please give some details about what kind of operations on
            rsets are done during the object copy phase, and can we see
            that from these logs?</div>
          <div>

          </div>
          <div>Thanks again,</div>
          <div>Nezih</div>
          <div>

          </div>
        </div>
      </div>
      <div class="gmail_extra">

        <div class="gmail_quote">2017-04-27 17:20 GMT-07:00 nezih
          yigitbasi <span dir="ltr"><<a moz-do-not-send="true" href="mailto:nezihyigitbasi@gmail.com" target="_blank">nezihyigitbasi@gmail.com</a>></span>:

          <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
            <div dir="ltr">Thanks for the suggestions. We use the
              default pause time. And here is our entire set of JVM
              args: <a moz-do-not-send="true" href="https://gist.github.com/nezihyigitbasi/04f5fdb9c32ac56097011819e20602d8" target="_blank">https://gist.github.com/<wbr>nezihyigitbasi/04f5fdb9c32ac56<wbr>097011819e20602d8</a>
              <div>

              </div>
              <div>I have some followup questions:

                <div>
                  <div>

                  </div>
                  <div>- In some case the object copy took 39406.8 ms,
                    even if the remembered set is ~30GB isn't this too
                    slow (that's <1GB/s of data)?</div>
                  <div>- Is there any way to reduce the rset overhead?</div>
                  <div>- My initial thought when I saw the high object
                    copy times was there may be some sort of contention
                    to have such a low throughput during the copy.
                    Although it may not be the case here, I just wonder
                    whether there is a way to see the amount of
                    contention from the gc logs?</div>
                  <span class="gmail-m_2983664130649143190HOEnZb"><font color="#888888">
                      <div>

                      </div>
                      <div>Nezih</div>
                    </font></span>
                  <div>
                    <div class="gmail-m_2983664130649143190h5">
                      <div>

                      </div>
                      <div>
                        <div>
                          <div class="gmail_extra">

                            <div class="gmail_quote">2017-04-27 16:58
                              GMT-07:00 Jenny Zhang <span dir="ltr"><<a moz-do-not-send="true" href="mailto:yu.zhang@oracle.com" target="_blank">yu.zhang@oracle.com</a>></span>:

                              <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Hi,
                                Hezih,

                                

                                It seems this workload is very heavy on
                                Remember Set. It has about 31G native
                                memory for RSet (old gen) and still with
                                coarsening.

                                

                                What is you pause time goal? The default
                                (200ms) might be too small for you.  Can
                                you increase that so G1 can increase the
                                young gen size? Since there is not much
                                promotion, I guess most of the object
                                copy time is dealing with the Remember
                                Set.

                                

                                There are other things you can try, like
                                increase the G1RSetReginEntries, but the
                                memory footprint will be bigger.

                                

                                So if you can, I suggest increase the
                                pause time goal first.

                                

                                Thanks<span class="gmail-m_2983664130649143190m_8770557251507851110gmail-HOEnZb"><font color="#888888">

                                    

                                    Jenny</font></span>
                                <div class="gmail-m_2983664130649143190m_8770557251507851110gmail-HOEnZb">
                                  <div class="gmail-m_2983664130649143190m_8770557251507851110gmail-h5">

                                    

                                    

                                    

                                    On 4/27/2017 9:22 AM, nezih
                                    yigitbasi wrote:

                                    <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                                      Hi,

                                      We see huge object copy times (and
                                      relatively high termination times)
                                      during young GCs in our production
                                      system running on Java
                                      1.8.0_112-b15. You can find the GC
                                      logs here: <a moz-do-not-send="true" href="https://gist.github.com/nezihyigitbasi/1f7a92da7860908a611cb1197bd8626b" rel="noreferrer" target="_blank">https://gist.github.com/nezihy<wbr>igitbasi/1f7a92da7860908a611cb<wbr>1197bd8626b</a>

                                      

                                      The young GC times start going
                                      high after the timestamp
                                      "2017-04-26T03:07:22.164-0700"<wbr>.

                                      

                                      I will appreciate if you can give
                                      some details about:

                                      - what goes into the "Object Copy"
                                      phase during young GCs and how we
                                      can reduce it.

                                      - why we see high Termination
                                      times and what we can do about it

                                      

                                      Thanks,

                                      Nezih

                                    </blockquote>
                                    

                                  </div>
                                </div>
                              </blockquote>
                            </div>
                            

                          </div>
                        </div>
                      </div>
                    </div>
                  </div>
                </div>
              </div>
            </div>
          </blockquote>
        </div>
        

      </div>
    </blockquote>
    

  </div></div></div>

</blockquote></div>
</div></div>



</blockquote>
</body></html>