Le 6/9/13 9:19 AM, Bernd Fondermann a écrit : > Hi, > > interesting, thanks for sharing! > > In general, did you conduct any comparisons varying the number of > clients/CPUs used? Yes, we did a bit of that. We ran the server with one singe thread to handle all the load (so with one selector only), and we saw no difference. Things seemed to have evolved a lot since Java 4, and now, all the load seems to be on the OS shoulders, not on the Java code.
Now, my computer has only 2 cores, it would be very valuable to test that on a more powerful machine with 4 CPU and 8 cores, for instance. > I'm not too much into the details of MINA and netty, but an important > thing for me to consider is also how performance changes with the > number of cores used. Is allocation of resources taking the cores into > account? I'm afraid that the OS is he bottleneck here. I'm quite sure that we need to tune various OS parameters to get a better performance. This is something that needs to be conducted pn a dedicated environment, with a Gb ethernet or even a 10 Gb ethernet (sending 70 000 1kb messages per second pretty much saturate a 1Gb network...). Note that the message size does not make any difference here, even with 10 bytes messages we get the system CPU jumpning to 85%. -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
