G1 GC 单个、非常长的年轻 GC 发生且 ParallelGCThreads=1

I set ParallelGCThreads=1并使用G1 GC，所有其他JVM设置为默认设置。我跑PageRank在 Spark-1.5.1 上，有两个 EC2 节点，每个节点有 100 GB 堆。

我的堆使用情况图如下（红色区域：年轻代，黑色区域：老年代）。所有的年轻GC都很小，突然来了一个年轻GC，收集了60GB，然后年轻GC又变小了。我的 GC 日志显示没有混合 GC、没有完整 GC、1 个并发标记和数十个年轻 GC。我想知道为什么会发生巨大的年轻GC？

下面是我的 GC 日志的一部分。巨大的年轻GC是带有“Heap: 84.1G”的GC

2015-12-30T06:59:02.488+0000: 245.088: [GC pause (young) 245.089: [G1Ergonomics (CSet Construction) start choosing CSet, _pending_cards: 1727, predicted base time: 24.64 ms, remaining time: 175.36 ms, target pause time: 200.00 ms]
 245.089: [G1Ergonomics (CSet Construction) add young regions to CSet, eden: 206 regions, survivors: 3 regions, predicted young region time: 148.87 ms]
 245.089: [G1Ergonomics (CSet Construction) finish choosing CSet, eden: 206 regions, survivors: 3 regions, old: 0 regions, predicted pause time: 173.51 ms, target pause time: 200.00 ms]
2015-12-30T06:59:02.531+0000: 245.131: [SoftReference, 0 refs, 0.0000520 secs]2015-12-30T06:59:02.531+0000: 245.131: [WeakReference, 21 refs, 0.0000160 secs]2015-12-30T06:59:02.531+0000: 245.131: [FinalReference, 9759 refs, 0.0084720 secs]2015-12-30T06:59:02.539+0000: 245.140: [PhantomReference, 0 refs, 14 refs, 0.0000190 secs]2015-12-30T06:59:02.539+0000: 245.140: [JNI Weak Reference, 0.0000130 secs] 245.142: [G1Ergonomics (Heap Sizing) attempt heap expansion, reason: recent GC overhead higher than threshold after GC, recent GC overhead: 12.51 %, threshold: 10.00 %, uncommitted: 0 bytes, calculated expansion amount: 0 bytes (20.00 %)]
, 0.0534140 secs]
   [Parallel Time: 42.3 ms, GC Workers: 1]
      [GC Worker Start (ms):  245088.6]
      [Ext Root Scanning (ms):  14.4]
      [Update RS (ms):  1.9]
         [Processed Buffers:  34]
      [Scan RS (ms):  0.4]
      [Code Root Scanning (ms):  0.0]
      [Object Copy (ms):  25.5]
      [Termination (ms):  0.0]
      [GC Worker Other (ms):  0.0]
      [GC Worker Total (ms):  42.3]
      [GC Worker End (ms):  245130.9]
   [Code Root Fixup: 0.0 ms]
   [Code Root Migration: 0.0 ms]
   [Clear CT: 1.6 ms]
   [Other: 9.5 ms]
      [Choose CSet: 0.0 ms]
      [Ref Proc: 8.6 ms]
      [Ref Enq: 0.2 ms]
      [Free CSet: 0.4 ms]
   [Eden: 6592.0M(6592.0M)->0.0B(58.8G) Survivors: 96.0M->128.0M Heap: 30.6G(100.0G)->24.2G(100.0G)]
 [Times: user=0.05 sys=0.00, real=0.06 secs] 
2015-12-30T06:59:43.451+0000: 286.051: [GC pause (young) 286.054: [G1Ergonomics (CSet Construction) start choosing CSet, _pending_cards: 392599, predicted base time: 367.03 ms, remaining time: 0.00 ms, target pause time: 200.00 ms]
 286.054: [G1Ergonomics (CSet Construction) add young regions to CSet, eden: 1884 regions, survivors: 4 regions, predicted young region time: 150.18 ms]
 286.054: [G1Ergonomics (CSet Construction) finish choosing CSet, eden: 1884 regions, survivors: 4 regions, old: 0 regions, predicted pause time: 517.21 ms, target pause time: 200.00 ms]
2015-12-30T06:59:47.767+0000: 290.368: [SoftReference, 0 refs, 0.0000570 secs]2015-12-30T06:59:47.768+0000: 290.368: [WeakReference, 350 refs, 0.0000640 secs]2015-12-30T06:59:47.768+0000: 290.368: [FinalReference, 99336 refs, 0.3781120 secs]2015-12-30T06:59:48.146+0000: 290.746: [PhantomReference, 0 refs, 1 refs, 0.0000290 secs]2015-12-30T06:59:48.146+0000: 290.746: [JNI Weak Reference, 0.0000140 secs] 290.767: [G1Ergonomics (Heap Sizing) attempt heap expansion, reason: recent GC overhead higher than threshold after GC, recent GC overhead: 11.74 %, threshold: 10.00 %, uncommitted: 0 bytes, calculated expansion amount: 0 bytes (20.00 %)]
, 4.7153740 secs]
   [Parallel Time: 4313.9 ms, GC Workers: 1]
      [GC Worker Start (ms):  286053.9]
      [Ext Root Scanning (ms):  15.2]
      [Update RS (ms):  86.3]
         [Processed Buffers:  1557]
      [Scan RS (ms):  4.1]
      [Code Root Scanning (ms):  0.2]
      [Object Copy (ms):  4208.1]
      [Termination (ms):  0.0]
      [GC Worker Other (ms):  0.0]
      [GC Worker Total (ms):  4313.9]
      [GC Worker End (ms):  290367.8]
   [Code Root Fixup: 0.0 ms]
   [Code Root Migration: 0.3 ms]
   [Clear CT: 15.1 ms]
   [Other: 386.0 ms]
      [Choose CSet: 0.0 ms]
      [Ref Proc: 378.4 ms]
      [Ref Enq: 1.7 ms]
      [Free CSet: 3.3 ms]
   [Eden: 58.9G(58.8G)->0.0B(3456.0M) Survivors: 128.0M->1664.0M Heap: 84.1G(100.0G)->26.7G(100.0G)]
 [Times: user=0.78 sys=3.94, real=4.71 secs] 

尝试堆扩展，原因：GC后最近的GC开销高于阈值，最近的GC开销：11.74％，阈值：10.00％

我的猜测是这正在推动 G1 的决定。您可以通过设置来放松它-XX:GCTimeRatio=4，这将允许它占用相对于应用程序时间的 20% 的 CPU 周期来进行 GC，而不是 10%。

如果那太多了，你应该

• 允许它使用更多的 CPU 核心 - 这将更容易满足其暂停时间目标，这反过来又意味着它可以推迟收集更长时间，从而更容易满足吞吐量目标。
  是的，这确实意味着使用更多内核实际上可以使用更少的 CPU 周期。
• 放宽暂停时间目标，因此必须减少收集频率