Add throughput distribution and coarse datapoints for StressWorkerBench

[twalluxio]

For random reads, bytes read per file is not a constant any more. In spite of existing duration distribution, need a throughput distribution for better understanding of reading performance.

Also, when duration too long, grpc will receive huge size of output data. Should aggregate data points to transfer more datapoints with limited output size.

Group datapoints by threads and time slices:
Example:

nodeResults: {
  worker-0: {
    dataPoints: [
      data: [
        { // worker 0, thread 0, slice 0
          count: 1,
          iobytes: 33554432,
        }, { // worker 0, thread 0, other slices
          …
        }
      ], [ // worker 0, other workers
        …
      ]
    ]
    throughputPercentiles: […]
  },
  worker-1: { // other workers
    …
  }
}

Slice time with --slice-size flag, e.g. --slice-size 1s.

[jiacheliu3]

thanks for the work! I left some comments :)

[jiacheliu3]

added a few more comments. I think it's good to merge after our comments are resolved. Thanks for the work!

[jiacheliu3]

Suggested change

	new ArrayList<>(), new ArrayList<>());
	new ArrayList<>(), new ArrayList<>());

are you going to use these lists and add to them? It's better if you just init empty lists internally.

public WorkerBenchCoarseDataPoint(long, long) {
  xxx = new ArrayList<>();
  xxx = new ArrayList<>();
}

Kinda weird if you create a list, and give it to someone, you and that someone both hold the list and can add to it.

[twalluxio]

Yes I agree. There are two places creating a WorkerBenchCoarseDataPoint. One in the bench (without data and throughput arrays), and another in the deserializer (with data and throughput arrays). I will write two different constructors for them.

[jiacheliu3]

ok let's change it to ns-based in another PR. The problem with ms is the error on the clock. If an operation is longer than 1ms then you may still observe the same ms value in some extreme cases. I don't really care what the smallest unit is, I just don't want DividedByZeroException.

[jiacheliu3]

LGTM, thanks

[jiacheliu3]

alluxio-bot, merge this please