[8] RFR: Shenandoah: specialize String Table scans for better pause performance

[Zhengyu Gu]

  935   assert(UseShenandoahGC, "Only for Shenandoah");
  936   const int chunk_size = limit / (ParallelGCThreads * 10);
  937

Maybe need to guard against minimum chunk_size? even through, it is 
unlikely, but possible chunk_size == 0.

Otherwise, looks good.

-Zhengyu


On 5/19/20 5:23 AM, Aleksey Shipilev wrote:
> This is sh/jdk8 specific fix. String Table scans dominate some workloads, notably Cassandra. But
> they are also visible on simple hello world scenarios. The underlying reason is that parallel
> StringTable scans use a very small chunk size, so on decently-sized tables we hit the contended
> atomic very often. Larger chunk size is needed to improve performance.
> 
> Since that chunk size is hard-coded and entrenched in the 8u upstream code, I chose to copy-paste
> the relevant method and specialize it for Shenandoah. 11u and later code handles it better with
> OopStorage and concurrent hash tables, so this specialization is only needed for 8u.
> 
> 8u webrev:
>    https://cr.openjdk.java.net/~shade/shenandoah/8u-stringtable-scan/webrev.01/
> 
> This yields the impressive improvements on Cassandra perf tests, see example GC cycles:
> 
> === Baseline:
> 
> Pause Init Mark (N)           3646 us
>    Accumulate Stats              18 us
>    Make Parsable                  4 us
>    Update Region States          86 us
>    Scan Roots                  3478 us
>      S: <total>               26681 us
>      S: Thread Roots           1521 us, workers (us): 468, 392, 111, 122, 113, 106, 111,  98,
>      S: Universe Roots            3 us, workers (us):   3, ---, ---, ---, ---, ---, ---, ---,
>      S: JNI Handles Roots         3 us, workers (us):   3, ---, ---, ---, ---, ---, ---, ---,
>      S: JNI Weak Roots           25 us, workers (us): ---, ---, ---,  25, ---, ---, ---, ---,
>      S: String Table Roots    24644 us, workers (us): 2857, 2935, 3141, 3122, 3148, 3148, 3148, 3145,
>      S: Synchronizer Roots        3 us, workers (us):   2,   0,   0,   0,   0,   0,   0,   0,
>      S: Flat Profiler Roots       2 us, workers (us): ---, ---,   2, ---, ---, ---, ---, ---,
>      S: Management Roots          2 us, workers (us):   2, ---, ---, ---, ---, ---, ---, ---,
>      S: System Dict Roots        10 us, workers (us): ---, ---,  10, ---, ---, ---, ---, ---,
>      S: CLDG Roots              458 us, workers (us):  12,  13,  75,  72,  74,  75,  65,  72,
>      S: JVMTI Roots              10 us, workers (us): ---,  10, ---, ---, ---, ---, ---, ---,
>    Resize TLABs                  11 us
> 
> === Patched:
> Pause Init Mark (N)            999 us
>    Accumulate Stats              17 us
>    Make Parsable                  4 us
>    Update Region States          92 us
>    Scan Roots                   824 us
>      S: <total>                5494 us
>      S: Thread Roots           1584 us, workers (us): 121, 613, 114, 131, 301, 110, 102,  93,
>      S: Universe Roots            2 us, workers (us):   2, ---, ---, ---, ---, ---, ---, ---,
>      S: JNI Handles Roots         4 us, workers (us): ---,   4, ---, ---, ---, ---, ---, ---,
>      S: JNI Weak Roots           20 us, workers (us): ---, ---, ---, ---, ---,  20, ---, ---,
>      S: String Table Roots     3374 us, workers (us): 506,  81, 479, 487, 372, 482, 473, 494,
>      S: Synchronizer Roots        4 us, workers (us):   0,   0,   3,   0,   0,   0,   0,   0,
>      S: Flat Profiler Roots       5 us, workers (us): ---,   5, ---, ---, ---, ---, ---, ---,
>      S: Management Roots          1 us, workers (us):   1, ---, ---, ---, ---, ---, ---, ---,
>      S: System Dict Roots        10 us, workers (us): ---, ---,  10, ---, ---, ---, ---, ---,
>      S: CLDG Roots              478 us, workers (us):  79,  12,  77,  69,  13,  75,  75,  78,
>      S: JVMTI Roots              10 us, workers (us):  10, ---, ---, ---, ---, ---, ---, ---,
>    Resize TLABs                  10 us
> 
> 
> Testing: hotspot_gc_shenandoah, Cassandra benches, eyeballing gc logs
>