Skip to content
Yangjun Wang edited this page Dec 22, 2015 · 5 revisions

Test(Spout -> AddTime -> Splitter -> pair -> sum -> sink) runs one one node. Whether the node is configured to run only one worker or 4 workers. The test performances are similar. Emit of spout: 1 worker(4 works/node) parallelism 1:
30s 16980
60s 56540
90s 93820
120s 130940
150s 171420

1 worker(one worker/node) parallelism 1:
30s 17520
60s 58140
90s 93340
120s 131880
150s 170940

2 workers(one worker/node) parallelism 2:
30s 24640 33280
60s 72780 82940
90s 130160 131320
120s 169100 186000
150s 222280 230740

2 workers(one worker/node) parallelism 3:
30s 32140
60s 92380
90s 151040
120s 204300
150s 260420

3 workers(one worker/node) parallelism 3:
30s 42500
60s 112340
90s 183420
120s 250060
150s 318740

4 workers(one worker/node) parallelism 4:
30s 47060 75300
60s 140540 131500
90s 232840 225380
120s 321080 316640
150s 413260 436660
180s 495200 503540

About latency, in ack case, skewed data could case throughput and latency of a certain node are both very heigh.