On Thu, 2002-09-26 at 19:49, Ariel Biener wrote: > > Would this count as huge ? > > 17: 186608169 186641281 IO-APIC-level eth1 > > 7:47pm up 1 day, 28 min, 5 users, load average: 0.23, 0.82, 1.25 > > If I count it right, this is: > > 186608169/86428 int/sec per cpu, that is 2159 int/second.
I'm not familier with your setup and enviorment and I dont want to say things I'm not 100% sure about, but it does sound a lot to me. Your machine seems to have done more network interrupts in 1 day then a reasnoably loaded web server I have access to did in the last 50... Again, I *don't* know if this is the cause of your problems, but taking into account that the interrupt rate and thus the CPU resources it consumes will increase exactly when the system is loaded and need CPU time most for VM managment, swapping and of course - your application it *might* be something to consider. As mentioned already teh NAPI patches can provide a solution and under normal circumstances I would advice to try them out. I don't know how feasible this is on the specific machine you are using. Maybe you have a testing facility of some sort to avoid doing this on a production unit (always a good idea IMHO :-) One other thing you have to consider is that on SMP systems having both CPUs servicing the network card means you'll end up with a LOT of packet reordering. This can increase load further as the TCP/IP stack will have to put them in order whcih is quite a costly thing - on the Internet packet reordering is quite rare. The usuall solution is to put the NICs on CPU affinity - to tie their handling witgh a specific CPU. This can be done via the /proc interface, don't remember exactly where. This two may or may not help you. Cheers, Gilad. -- Gilad Ben-Yossef <[EMAIL PROTECTED]> http://benyossef.com "Geeks rock bands cool name #8192: RAID against the machine" ================================================================= To unsubscribe, send mail to [EMAIL PROTECTED] with the word "unsubscribe" in the message body, e.g., run the command echo unsubscribe | mail [EMAIL PROTECTED]