"Jim Helm" <[EMAIL PROTECTED]> wrote: >Everything I've read as an SA (for Solaris at least - though I would >expect the other *nices to be similar) was to never set a user space >(non O/S) process to less than -15. Other than that, it's another of >those YMMV, measure before and after, and if it helps great. Trying to >second guess process schedulers is a tricky business though, and you >really need to intimately know how your system behaves before trying it. > >> -----Original Message----- >> >> Alexey Zvyagin has suggested a use of Unix process priorities >> to improve >> the performance of the web services during the peak hours: >> >> Alex writes: >> ---------------------------------------------------------
[snip] >> The CPU priorities help to handle an increased traffic on the >> overloaded >> server. >> --------------------------------------------------------- I think the key here is the fact that the system is overloaded/overcommitted. Too many processes are competeing for CPU. Putting my SA hat on, I would say the processes need to be split across multiple pieces of hardware or a new machine needs to replace the current system. That's the only real long-term solution for a system in this state. The priorities as described help set a relative importance between the processes--the front-end is more important than the database with the larger server processes in the middle. This is not unreasonable. But a serious solution to the problem of an overloaded system is to put in more system. The priorities might be helpful for the few minutes/hours/days needed to get the new hardware on the floor. Imho, a web server should be designed for the expected peak normal usage plus a fudge factor thrown in for safety and multiplied by a small integer greater than one for growth (I'm optimistic). Conclusion of my thoughts: putting in a blurb about priorities being able to set relative importance of processes is fine, but don't cast it as a solution (just as swap space is not a solution to constrained memory). It's a bit of a band-aid that can help until the problem can be fixed. -- James Smith <[EMAIL PROTECTED]>, 979-862-3725 Texas A&M CIS Operating Systems Group, Unix