JDK-8229517: Buffered Logging?

[Liu, Xin]

Hi, Thomas, 

I will work on the tip of jdk. Certainly, It will support unified log(UL).
In our current webrev,  we moved all file IOs out of GC safepoints. It helps for p99/p100 GC pause time. We started with optimizing GC pause time in long tail, that's why we hacked GCLog first.    
Ideally, we should add async API to UL. Users of UL who are very sensitive to delay can choose asynchronous API.  
 
About parameters, I have the same concerns. To be honest, I don't have answer to make everybody happy. What I did is guarantee that no gclog will miss when logs are very verbose and bogus delay is 5m for every File write. 
I don't think it's good idea to use dynamic allocation for  log buffers.  If kernel had very long pause, JVM would run out of memory.  I don't like to introduce a new problem when I try to solve one. 
Here is  the script I used to collect data. 
https://github.com/navyxliu/JavaGCworkload/blob/master/t.sh#L7
Those parameters are tuning.  We can define a bottom line. 
https://cr.openjdk.java.net/~xliu/8229517/webrev03/webrev/src/share/vm/runtime/globals.hpp.udiff.html

I used a mutex without safepoint check.  It should guarantee that the synchronization between read/writer won't wait for safepoints.  I am quite new for hotspot synchronizations. Could you help me review this part? 
https://cr.openjdk.java.net/~xliu/8229517/webrev03/webrev/src/share/vm/utilities/ostream.cpp.udiff.html
Two buffers switch their roles in flush_log().  We also explored lock-free queue using  boost. It works but it's non-trivial for me to implement one in hotspot.  I would like to hold until we identify async buffer is bottleneck. 

thanks,
--lx


From: Thomas Stüfe <thomas.stuefe at gmail.com>
Date: Wednesday, January 29, 2020 at 3:13 AM
To: "Liu, Xin" <xxinliu at amazon.com>
Cc: hotspot-runtime-dev <hotspot-runtime-dev at openjdk.java.net>, John Coomes <jcoomes at twitter.com>, "Mathiske, Bernd" <mathiske at amazon.com>, Ramki Ramakrishna <rramakrishna at twitter.com>
Subject: Re: JDK-8229517: Buffered Logging?

Hi Xin, 

I was not sure whether to respond in JBS or here, so I do it here.

Like others I think that if we really do this it would make more sense to add this as a general feature of UL, not of the GC logging in particular. We have the problem (logging may block or take time) in other areas too. I have thought about a similar solution in the past (offloading UL writes to a dedicated thread) but usually when I thought about the details I got unsure. But maybe I do not see the whole picture, and you can convince me :)

Some questions (assuming we go the more general direction of buffering UL output):

- How do you determine which buffer size is large enough? Lets say UL gets really chatty, how do you deal with buffer overflow? Do you stop the writers until the buffer is drained or do you discard log output? Both options seems similar bad.

- We need synchronization now between reader and writers. Slow IO can stall the reader, which may block the writers (if you stop writers on a full buffer). I am not sure, but this may be worse than before, where we had the writer threads effectively write in parallel to the same fd.

Thanks & Cheers, Thomas


On Sun, Jan 26, 2020 at 12:39 AM Liu, Xin <mailto:xxinliu at amazon.com> wrote:
Dear hotspot developers, 

    I think runtime might be a better forum to continue this discussion.  
    The JBS: https://bugs.openjdk.java.net/browse/JDK-8229517

    Now our latest patch is here: https://cr.openjdk.java.net/~xliu/8229517/webrev02/webrev/
    I know it’s based on jdk8u. I really appreciate if someone reviews it and gives me advice.  If the community is still interested, we can keep developing using better approach or on the tip of jdk. 

    I’d like to present our experiments here. I try to post data on the JDK-8229517, but It doesn't look good in comment area.

    Experiments
    We can’t reproduce the problem in [1] because harddisks are not accessible for us. To simulate IO blocking in JVM, we introduce a bogus delay. It spins DelayInGCLogging milliseconds for every fwrite. 
      notproduct(intx, DelayInGCLogging, 0,
      "Delay milliseconds to simulate IO blocking for GC log testing")

    It spins in gcLogFileStream::write_blocking. https://cr.openjdk.java.net/~xliu/8229517/webrev02/webrev/src/share/vm/utilities/ostream.cpp.cdiff.html

    We observe GC pause time using -XX:+PrintGCApplicationStoppedTime for JavaWorkload with 2g heap. The unit is second.  The program is on github. https://github.com/zhenyun/JavaGCworkload.
    $java -Xmx2g -Xms2g -Xloggc:gc.log -XX:+UseG1GC -XX:+PrintGCApplicationStoppedTime -XX:+PrintGCDetails -XX:+UnlockExperimentalVMOptions Test
    sync means jdk8u-dev. async means jdk8u-dev with nonblocking gclog change. delayXXX denotes an artifactual delay given by -XX:DelayInGCLogging=XXX. 

                                 median     p95         p99           p100        losses
    -------------------------------------------------------------------
    async                   0.01357   0.36187  0.48691   0.61792
    sync                     0.01339   0.3618    0.48646   0.63539
    -------------------------------------------------------------------
    delay5ms-async 0.01516  0.40564  0.486        0.61053  148156
    delay5ms-sync   0.20444  0.4428    0.56733    0.69522
    -------------------------------------------------------------------
    delay10ms-async0.0123   0.47451  0.48623    0.61083  195460
    delay10ms-sync  0.38732 0.52424  0.64783    0.78087
    -------------------------------------------------------------------

    async and sync perform equally in happy case. JavaWorkload didn’t lose any GC log even though the GC activities are intensive and the outputs are verbose. If we introduce 5 ms delay for each file-writing, delay5ms-sync starts to show longer stop time for p99 and p100. eg. p99 increases from 0.486s to 0.567s, which is 16,7% longer. delay5ms-async remains the same p99 and p100 at expense of losing 148k logging events. Because the outputs GC logs are contiguous, they are still readable and helpful for problem analysis. After we increase delay to 10ms, delay10ms-async still have the same p99 and p100 stop time while delay10ms-sync has 22.9% longer stop time. 

    [1]Eliminating Large JVM GC Pauses Caused by Background IO Traffic
    https://engineering.linkedin.com/blog/2016/02/eliminating-large-jvm-gc-pauses-caused-by-background-io-traffic

    Thanks,
    --lx





    From: Ramki Ramakrishna <mailto:rramakrishna at twitter.com>
    Date: Tuesday, August 13, 2019 at 4:16 PM
    To: "mailto:hotspot-dev at openjdk.java.net" <mailto:hotspot-dev at openjdk.java.net>
    Cc: "Liu, Xin" <mailto:xxinliu at amazon.com>, "Mathiske, Bernd" <mailto:mathiske at amazon.com>, "Hohensee, Paul" <mailto:hohensee at amazon.com>, John Coomes <mailto:jcoomes at twitter.com>, Tony Printezis <mailto:tprintezis at twitter.com>
    Subject: Re: Buffered Logging?


    I filed https://bugs.openjdk.java.net/browse/JDK-8229517.  Please correct/complete if I mangled or neglected to fill any specific required fields.

    Patches forthcoming in the next few days.
    -- ramki

    On Tue, Aug 13, 2019 at 2:53 PM Ramki Ramakrishna <mailto:mailto:rramakrishna at twitter.com> wrote:

    I'll go ahead and open a JBS ticket to keep the ball rolling on this one, and so we have a place to hang some patches and discussions off of. Scream if I shouldn't.

    -- ramki

    On Mon, Aug 12, 2019 at 6:19 PM Ramki Ramakrishna <mailto:mailto:rramakrishna at twitter.com> wrote:

    (Resent... Apologies to anyone who might see duplicates. There was an issue with my membership in the list which, I am guessing, has now been resolved.)

    ---------- Forwarded message ----------
    From: Ramki Ramakrishna <mailto:mailto:rramakrishna at twitter.com>
    To: mailto:mailto:hotspot-dev at openjdk.java.net
    Cc: "Liu, Xin" <mailto:mailto:xxinliu at amazon.com>, "Mathiske, Bernd" <mailto:mailto:mathiske at amazon.com>, "Hohensee, Paul" <mailto:mailto:hohensee at amazon.com>, John Coomes <mailto:mailto:jcoomes at twitter.com>, Tony Printezis <mailto:mailto:tprintezis at twitter.com>
    Bcc: 
    Date: Mon, 12 Aug 2019 15:48:17 -0700
    Subject: Buffered Logging?

    In the JDK 7 time-frame, some folks noticed much longer than expected pauses that seemed to coincide with GC logging in the midst of a GC safepoint. In that setup, the GC logs were going to a disk file (these were often useful for post-mortem analyses) rather than to a RAM-based tmpfs which had been the original design center assumption. The vicissitudes of the dirty page flushing policy in Linux when 
    IO load on the machine (not necessarily the JVM process doing the logging)
    could affect the length and duration of these inline logging stalls.

    A buffered logging scheme was then implemented by us (and independently by
    others) which we have used successfully to date to avoid these pauses in high i/o
    multi-tenant environments.

    We have recently ported the OpenJDK 8u based scheme past the Unified Logging related changes and have been using it on OpenJDK 11u.

    In various fora, some of us have heard people express an interest in such an implementation.

    Would there be appetite for this in openjdk (there are some differences between our scheme in OpenJDK 8u and their subsequent post-Unified-Logging port to OpenJDK 11)?

    If so, we'd like to open a JBS ticket, attach some patches rebased to current OpenJDK tip, and start preparing a webrev for discussion with a view to possible upstreaming if it converges past review/discussion/debate.

    PS: In recent discussion at the JVMLS, mention was made of something that might be similar in support of JFR streaming, so our rebase to tip and familiarization with any JFR-associated buffered logging might well make this easier or, possibly, moot.

    Thoughts, questions, comments, feedback?
    -- ramki

    -- 
    JVM Team, Platform Engineering, Twitter (San Francisco)



    -- 
    JVM Team, Infrastructure Engineering, Twitter (San Francisco)



    -- 
    JVM Team, Infrastructure Engineering, Twitter (San Francisco)