Performance benchmark, profiling and optimization about grpc-go

[codahale]

I did some initial benchmarking of a gRPC service here, including profiling: https://gist.github.com/codahale/b3db28bfa3f7dd59d048.

It’s slower than net/http, and @kenkeiter and I think it’s probably due to IO contention. net/http uses bufio to buffer writes; grpc-go, on the other hand, uses a bare net.Conn.

[iamqizhao]

Thanks for benchmarking this. Per my reply to #108, this is because we have not done any batching IO so far, which is the focus of performance optimization in the next a couple of months.

It would be highly appreciated if you can have a pull request to wrap up what you have done to kick off the benchmark work on github.

[codahale]

It’s not much, but the code’s here: https://github.com/codahale/grpc-example.

I just spun that up on two m3.2xls.

[iamqizhao]

okay, I am going to try to push out a basic benchmark framework this week and then all the contributors can experiment various performance optimization ideas on the same ground.

[iamqizhao]

I made some improvement on client (not checked in yet) and got the significant improvement already:

http client:
2015/03/11 15:39:40 1: 5736.642904 op/sec @ p99=0.336000ms
2015/03/11 15:39:50 2: 11015.068698 op/sec @ p99=0.321000ms
2015/03/11 15:40:00 3: 17261.079718 op/sec @ p99=0.424000ms
2015/03/11 15:40:10 4: 20806.615173 op/sec @ p99=0.674000ms
2015/03/11 15:40:20 5: 917.775790 op/sec @ p99=7.762000ms

improved grpc client:
2015/03/11 15:28:20 1: 3474.533592 op/sec @ p99=0.453000ms
2015/03/11 15:28:30 2: 8111.900305 op/sec @ p99=0.657000ms
2015/03/11 15:28:40 3: 14023.490286 op/sec @ p99=0.663000ms
2015/03/11 15:28:50 4: 21179.714876 op/sec @ p99=0.617000ms
2015/03/11 15:29:00 5: 27564.232005 op/sec @ p99=0.612000ms
2015/03/11 15:29:10 6: 33875.344691 op/sec @ p99=0.603000ms
2015/03/11 15:29:20 7: 39543.547520 op/sec @ p99=0.616000ms
2015/03/11 15:29:30 8: 43567.770711 op/sec @ p99=0.668000ms
2015/03/11 15:29:40 9: 46393.655755 op/sec @ p99=0.711000ms
2015/03/11 15:29:50 10: 47250.902113 op/sec @ p99=0.796000ms
2015/03/11 15:30:00 11: 47733.268011 op/sec @ p99=0.961000ms
2015/03/11 15:30:10 12: 47531.488503 op/sec @ p99=1.130000ms
2015/03/11 15:30:20 13: 50293.363756 op/sec @ p99=0.938000ms
2015/03/11 15:30:30 14: 50614.134504 op/sec @ p99=1.006000ms
2015/03/11 15:30:40 15: 50902.922158 op/sec @ p99=1.061000ms
2015/03/11 15:30:50 16: 51342.060561 op/sec @ p99=1.103000ms
2015/03/11 15:31:00 17: 51172.659114 op/sec @ p99=1.200000ms
2015/03/11 15:31:10 18: 51340.048872 op/sec @ p99=1.230000ms
2015/03/11 15:31:20 19: 51575.161160 op/sec @ p99=1.277000ms
2015/03/11 15:31:30 20: 51478.429739 op/sec @ p99=1.376000ms
2015/03/11 15:31:40 21: 49980.940379 op/sec @ p99=1.738000ms
2015/03/11 15:31:50 22: 51071.717198 op/sec @ p99=1.681000ms
2015/03/11 15:32:00 23: 52140.282287 op/sec @ p99=1.553000ms
2015/03/11 15:32:10 24: 52124.439668 op/sec @ p99=1.593000ms
2015/03/11 15:32:20 25: 52222.698417 op/sec @ p99=1.646000ms
2015/03/11 15:32:30 26: 52365.708092 op/sec @ p99=1.795000ms
2015/03/11 15:32:40 27: 52629.151019 op/sec @ p99=1.771000ms
2015/03/11 15:32:50 28: 52815.300753 op/sec @ p99=1.857000ms
2015/03/11 15:33:00 29: 53304.403154 op/sec @ p99=1.839000ms
2015/03/11 15:33:10 30: 53081.078234 op/sec @ p99=1.986000ms
2015/03/11 15:33:20 31: 53571.344040 op/sec @ p99=1.945000ms
2015/03/11 15:33:30 32: 53348.888919 op/sec @ p99=2.008000ms
2015/03/11 15:33:40 33: 53429.124904 op/sec @ p99=2.147000ms
2015/03/11 15:33:50 34: 53687.364050 op/sec @ p99=2.196000ms
2015/03/11 15:34:00 35: 54029.536526 op/sec @ p99=2.184000ms
2015/03/11 15:34:10 36: 53837.964570 op/sec @ p99=2.264000ms
2015/03/11 15:34:20 37: 53558.114539 op/sec @ p99=2.358000ms
2015/03/11 15:34:31 38: 54276.048533 op/sec @ p99=2.395000ms
2015/03/11 15:34:41 39: 54714.463152 op/sec @ p99=2.424000ms
2015/03/11 15:34:51 40: 54343.651778 op/sec @ p99=2.501000ms
2015/03/11 15:35:01 41: 54334.642340 op/sec @ p99=2.557000ms
2015/03/11 15:35:11 42: 52279.954783 op/sec @ p99=3.224000ms
2015/03/11 15:35:21 43: 54623.572538 op/sec @ p99=2.759000ms
2015/03/11 15:35:31 44: 55033.028392 op/sec @ p99=2.729000ms
2015/03/11 15:35:41 45: 55097.426098 op/sec @ p99=2.797000ms
2015/03/11 15:35:51 46: 54736.133552 op/sec @ p99=2.973000ms
2015/03/11 15:36:01 47: 54809.898913 op/sec @ p99=2.926000ms
2015/03/11 15:36:11 48: 55465.691515 op/sec @ p99=3.009000ms
2015/03/11 15:36:21 49: 55142.363729 op/sec @ p99=3.053000ms
2015/03/11 15:36:31 50: 54523.644725 op/sec @ p99=3.341000ms
2015/03/11 15:36:41 51: 51818.090431 op/sec @ p99=3.874000ms
2015/03/11 15:36:51 52: 55958.880606 op/sec @ p99=3.230000ms
2015/03/11 15:37:02 53: 55770.858429 op/sec @ p99=3.363000ms
2015/03/11 15:37:12 54: 55643.909388 op/sec @ p99=3.520000ms
2015/03/11 15:37:22 55: 53500.566081 op/sec @ p99=3.894000ms
2015/03/11 15:37:32 56: 55789.631428 op/sec @ p99=3.722000ms
2015/03/11 15:37:42 57: 56027.047875 op/sec @ p99=3.556000ms
2015/03/11 15:37:52 58: 56038.120343 op/sec @ p99=3.638000ms
2015/03/11 15:38:02 59: 56292.927614 op/sec @ p99=3.793000ms
2015/03/11 15:38:12 60: 55018.049146 op/sec @ p99=4.073000ms
2015/03/11 15:38:22 61: 55992.929617 op/sec @ p99=4.024000ms
2015/03/11 15:38:32 62: 55822.072635 op/sec @ p99=4.004000ms
2015/03/11 15:38:42 63: 50557.821204 op/sec @ p99=5.170000ms

I am going to
i) investigate the p99 latency increase when concurrency is 1, 2, 3;
ii) improve server side IO.

[hsaliak]

@iamqizhao can we close this issue, given that we have such a framework in place