Performance

Overview #

It’s hard to do a full benchmark for the scheduler. We need to choose a scenario in which the scheduler will be busy creating, scheduling, running and destroying the processes.

We continue to talk about the topic of the sum example in the tutorial. But now the computing rule is changed to a divide-and-conquer one: sum(low, high) = sum(low, mid) + sum(mid + 1, high) in which mid is (low + high) / 2.

In a basic version, it may look like:

int sum(int low, int high) {
  if (low == high) {
    return low;
  }
  int mid = low + ((high - low) >> 1);
  return sum(low, mid) + sum(mid + 1, high);
}

In our benchmark, we compute the sub-parts sum in two processes each. And it becomes:

proc void sum(int low, int high, int *result) {
  if (low == high) {
    *result = low;
    return;
  }
  int mid = low + ((high - low) >> 1), left, right;
  sync(sum(low, mid, &left); sum(mid + 1, high, &right));
  *result = left + right;
}

Go to sum.go for the equivalent golang function.

Environment #

  • OS: Linux ubuntu-bionic 4.15.0-88-generic x86_64
  • CPU Cores: 8
  • Memory: 8192M
  • Go: v1.14
  • GCC: v8.3.0

Benchmark #

Now compute sum(0, 10000000) for 10 rounds and print the spent time per round in seconds.

$ for i in $(seq 8); do make clean && make CC=gcc-8 CPU_CORES=${i} benchmark_sum_libcsp; done
The result is 50000005000000, ran 10 rounds, 3.047471 seconds per round.
The result is 50000005000000, ran 10 rounds, 1.579780 seconds per round.
The result is 50000005000000, ran 10 rounds, 1.082895 seconds per round.
The result is 50000005000000, ran 10 rounds, 0.847325 seconds per round.
The result is 50000005000000, ran 10 rounds, 0.787095 seconds per round.
The result is 50000005000000, ran 10 rounds, 0.744351 seconds per round.
The result is 50000005000000, ran 10 rounds, 0.717903 seconds per round.
The result is 50000005000000, ran 10 rounds, 0.726390 seconds per round.

$ for i in $(seq 8); do make clean && env GOMAXPROCS=${i} make benchmark_sum_go; done
The result is 50000005000000, ran 10 rounds, 31.145248 seconds per round.
The result is 50000005000000, ran 10 rounds, 17.112188 seconds per round.
The result is 50000005000000, ran 10 rounds, 12.460803 seconds per round.
The result is 50000005000000, ran 10 rounds, 11.172224 seconds per round.
The result is 50000005000000, ran 10 rounds, 9.097314 seconds per round.
The result is 50000005000000, ran 10 rounds, 8.355515 seconds per round.
The result is 50000005000000, ran 10 rounds, 8.135163 seconds per round.
The result is 50000005000000, ran 10 rounds, 8.029795 seconds per round.

Below is the result in table:

CPU Cores Time(s) libcsp/go Memory(%) libcsp/go
1 3.047471 / 31.145248 0.2 / 27.8
2 1.579780 / 17.112188 0.5 / 29.4
3 1.082895 / 12.460803 1.0 / 28.9
4 0.847325 / 11.172224 1.2 / 30.7
5 0.787095 / 9.097314 1.6 / 32.0
6 0.744351 / 8.355515 1.9 / 29.5
7 0.717903 / 8.135163 2.0 / 29.5
8 0.726390 / 8.029795 2.3 / 31.2

We will see that libcsp is at least 10 times faster and use less memory than golang in the benchmark.