Re: squid benchmarking results - squid-3.0 + epoll()

hai david,

    thanks for your reply.

    i have tested with squid + epoll() using null fs. it is consuming 90% CPU at peak rate ( 180 req / sec ). system cpu idleness is
in the range of 0.
    Are you have any benchmarking results to epoll() with squid there?

    how to integrate shared memory with null fs ? I have the configuration option to include null fs on squid-3.0 as
    cache_dir null /dev/null
    cache_mem 200MB

    epoll() is servicing only 20 req more compared to poll()? ( 180 req / sec epoll() - 160 req / sec poll() ). I have included
latest patch for epoll().
    I have tuned kernel parameters as,

    # polyserver, polyclient, squid
    echo 1 > /proc/sys/net/ipv4/tcp_timestamps
    echo 1 > /proc/sys/net/ipv4/tcp_window_scaling
    echo 1 > /proc/sys/net/ipv4/tcp_sack

    echo 8388608 > /proc/sys/net/core/wmem_max
    echo 8388608 > /proc/sys/net/core/rmem_max
    echo "4096 87380 4194304" > /proc/sys/net/ipv4/tcp_rmem
    echo "4096 65536 4194304" > /proc/sys/net/ipv4/tcp_wmem

    # polyserver, polyclient
    ulimit -HSn 8192
    sysctl -w net.ipv4.ip_forward=1
    sysctl -w net.ipv4.tcp_tw_recycle=1
    sysctl -w net.ipv4.ip_local_port_range="32768 60001"

    # squid server
    ulimit -HSn 32768
    sysctl -w net.ipv4.tcp_syncookies=1

thanks
muthukumar

> Hi,
>
> Based upon my experience. aufs works best when you have CPU to spare. More often than not it ends up eating up more CPU to
> scheduling of the threads, than you gain. Perhaps try ufs in comparison. For a lot of our workloads it is actually faster than
> aufs on decent disks. Or even better use ufs or the null fs in combination with /dev/shm (if you can spare the memory).
>
Received on Wed Dec 08 2004 - 22:21:20 MST