Update on Mina and Twisted benchmarks
2007-08-10 10:49:49 by Fabio FornoO, better, update on Mina. Following the comment on Infinite to my last post I've run the same test on Mina using Java 6, which uses epoll as default. Now the figures are pretty better, almost as good as the ones of Twisted:
| #connections | Server Load | Client Load | Latency |
| 10000 | 1-2% | 1-2% | 6-8ms |
| 15000 | 1-3% | 1-2% | 7-8ms |
There just a very sight increase of CPU (1% or less) and I've noticed some very rare spikes (8-9%) that could be due to the garbage collector.
Instead the latency and the distance of the packets remains significantly higher. With both Twisted and Mina we scheduled to send a packet each 5ms, but we receive a packet each 5-8ms using Twisted and each 11-13ms using Mina. This means that the possible throughput is between 30% and 50% higher in Twisted. Frankly I still don't explain this difference, while the the use of poll and epoll was obvious in the previous tests.
