bad lmbench results on 2.6.x kernel with SMP - Linux

This is a discussion on bad lmbench results on 2.6.x kernel with SMP - Linux ; Hello group! When we migrate from 2.4.20 kernel to the various 2.6.x kernel versions we have noticed severe degradation in the lmbench context switch benchmark on the various dual Xeon SMP blades. On the single CPU blades everything works OK. ...

+ Reply to Thread
Results 1 to 2 of 2

Thread: bad lmbench results on 2.6.x kernel with SMP

  1. bad lmbench results on 2.6.x kernel with SMP

    Hello group!

    When we migrate from 2.4.20 kernel to the various 2.6.x kernel versions we
    have noticed severe degradation in the lmbench context switch benchmark on
    the various dual Xeon SMP blades. On the single CPU blades everything works
    OK. Also other benchmarks in the lmbench shows comparable results between
    2.4.20 and 2.6.x kernels.

    We found that on 2.6.x disabling the SMP support in kernel actually
    increases the lmbench context switch result. On 2.4.20 disabling the SMP
    support decreased performances (logically). Also using CPU affinity to lock
    processes to cores helps. But still we are far away from results we achieved
    with 2.4.20 kernel.

    As our application uses a lot of message passing and processes context
    switches we are not able to utilize SMP CPU blades or multicore CPU under
    2.6 kernels as we expect.

    Did anyone encounter similar problem? Any suggestion what to do?

    Regards,

    Sani




  2. Re: bad lmbench results on 2.6.x kernel with SMP

    Sani Rus wrote:

    > Hello group!
    >
    > When we migrate from 2.4.20 kernel to the various 2.6.x kernel versions we
    > have noticed severe degradation in the lmbench context switch benchmark on
    > the various dual Xeon SMP blades. On the single CPU blades everything
    > works OK. Also other benchmarks in the lmbench shows comparable results
    > between 2.4.20 and 2.6.x kernels.
    >
    > We found that on 2.6.x disabling the SMP support in kernel actually
    > increases the lmbench context switch result. On 2.4.20 disabling the SMP
    > support decreased performances (logically). Also using CPU affinity to
    > lock processes to cores helps. But still we are far away from results we
    > achieved with 2.4.20 kernel.
    >
    > As our application uses a lot of message passing and processes context
    > switches we are not able to utilize SMP CPU blades or multicore CPU under
    > 2.6 kernels as we expect.
    >
    > Did anyone encounter similar problem? Any suggestion what to do?
    >
    > Regards,
    >
    > Sani


    Unless your application requires features of a 2.6 kernel, it would appear
    for your application (if it requires that particular feature) to remain on
    a 2.4 kernel.
    It might be be worth while getting in touch with your distro's support or
    perhaps the kernel mailing list, it could be that there is setting to
    optimise this in a 2.6 kernel.

    Bruce S.
    --
    Replace the by by blueyonder

+ Reply to Thread