I think the problem is the 2.4 scheduler. It essentially has to wade through each of the processes in the process table for just about every scheduling call. 2.6 (and possibly RH9 has a patched 2.4) has an O(1) scheduler that should alleviate you. In addition, RH9 has a completely new threading implementation that may help you out.
Anyway, I would test RH9 on that load and see what happens. Jon On Thu, 26 Jun 2003, Fred Whipple wrote: > Hi all, > > I think I may be beating-up one of my servers too badly, but I'm not > sure what the hard limits are within the 2.4 Kernel. > > I have a server running right now with about 250 processes. About 150 > of those are Java VM processes, each with between 20-100 of its own > threads. A couple have many more threads, one has ~430 threads, the > other ~220. Of course, each thread gets its own PID, and overall the > system is running at about 2600 PID's allocated (again, to about 250 > processes, 150 of which are Java VM's). Each JVM has a unique User ID, > so the problems we're experiencing are not, I would think, related to a > per-user limitation. > > We're starting to see the server fail to execute new processes (such as > killall or shutdown!) with a "segmentation fault" and the server is, at > times, dead in the water until it gets itself a hard reboot. Actually I > have a couple tricks to get around a hard reboot but they're not pretty. > The server's memory and swap are no where near exhausted, and the CPU > util. is relatively low. > > Anyhow, my questions: > > 1. Am I brushing up against some limitations of the kernel here? Or is > there something else I may be missing afoot? > > 2. If there is a limitation to the # of PID's the kernel will dole out, > is this limitation limited explicitely to threads, or is it a matter of > total # of processes? > > 3. I've noticed that the stock Red Hat 9 kernel more gracefully reports > a single process per Java VM rather than showing all threads (at least > when querying with a 'ps xauwww'). Is the above (presumed) limitation > still implicit, or does it go away once threads are no longer assigned > their own PID? The implication here is, should I go through a painful > upgrade to solve the problem, or should I find a more interesting problem? > > 4. Are there any good solutions to this issue without upgrading? > Perhaps customizing the kernel somehow? > > TIA, > > -Fred > > > -- > redhat-list mailing list > unsubscribe mailto:[EMAIL PROTECTED] > https://www.redhat.com/mailman/listinfo/redhat-list > -- redhat-list mailing list unsubscribe mailto:[EMAIL PROTECTED] https://www.redhat.com/mailman/listinfo/redhat-list
