test

This is for an acceptor which streams market data.

How are you measuring the relative latency between NIO and blocking IO? Are you setting TCP_NODELAY? Are you wanting to process 25000 msg/sec (500 users x 500 msg/sec) with a single QFJ process? To anything close to that performance, I'd recommend studying the QFJ acceptor implementation and MINA. There are tuning options that might be useful for your application that are currently not directly supported by QFJ (but could be added). Which acceptor implementation are you using (single thread or thread-per-session)?

1) I am running couple of worker clients that initiate 50 fix client sessions per worker and check the delay between current time the fix sending time.
2) I have the nagle algorithm disabled (TcpNoDelay=true)
3) I tried with 1-10 QFJ processors
4) I am using SingleThread as I don't get many messages to the server. It's a streaming server. So, one thread is sufficinet for me.

Could you pls. advise some of the tuning options to achieve better results?

Thanks for the quick response.

Hi Steve,
Also, the throughput on blocked I/O is 5% less compared to non-blocking i/o. But, latency is more important than the throughput for us. The reason for going to non-blocking to avoid delays due to context switching when running 500+ users and our app already uses 2 threads for other purposes.

Hi rohit,

I saw your questions in the MINA list. If I understand them correct, we are doing what they suggested to minimize latency.