squid-3.0 withepoll() vs withoutepoll()

From: Muthukumar <kmuthu_gct@dont-contact.us>
Date: Mon, 20 Dec 2004 17:24:01 +0530

Hello Development Team,

    We are benchmarking squid performance with epoll() / without epoll() on the hardware configuration of,

    model name : Pentium III (Coppermine)
    cpu MHz : 927.753
    RAM size : 512 MB
    version : Fedora core 2 - 2.6.5

    Results with squid-3.0 ( Development ) and squid-2.5S7 (Stable ) are as,

    squid-3.0 Pre3 + epoll() + in-core memory + /dev/null fs ( with epoll() ) :
    =================================================
            req.rate = 371 req / sec
            rep.rate = 370 rep / sec

        Problem : CPU usage to 100% - CPU bound

  squid-3.0 Pre3 + poll() + in-core memory + /dev/null fs ( without epoll() ) :
  ==================================================
            req.rate = 345 req / sec
            rep.rate = 344 rep / sec

        Problem : CPU usage to 100% - CPU bound

    poll() vs epoll() :
        using poll(), squid is using 50% - 60% usage for even 50 - 150 requests / sec. epoll() is utilizing around 10-15% cpu usage.
( report based on top tool )

    squid-2.5 Stable 7 + poll() + in-core memory + /dev/null fs ( Stable version ) :
    ====================================================
        req.rate=612 req / sec
        rep.rate=611 rep / sec

        Problem : CPU usage to 100% - CPU bound

    Why squid-3.0 withoutepoll() benchmark and squid-2.5 stable 7 results are getting differed? epoll() method is making 100% cpu
usage.? Is it a right behaviour of epoll() I/O method? ( I have used recent 100% usage patch of epoll() ). NO error messages
appeared in cache.log file.

    IF you want to have more informations, let us know.

    Thanks for your help.

Regards
Visolve Dev. team.
Received on Mon Dec 20 2004 - 04:55:04 MST

This archive was generated by hypermail pre-2.1.9 : Fri Dec 31 2004 - 12:00:05 MST