On Sat, Sep 13, 2008 at 12:52 AM, Elad Lahav <elahav@xxxxxxxxxxxx> wrote: > I hacked the kernel so that account_system_time() can differentiate between > "true" soft > IRQs and code running under local_bh_disable(). (The hack itself is quite > ugly, but does > the job). An experiment that sends UDP packets as quickly as possible, now > shows that over > 25% of the time is taken by the latter. Here's a snapshot of the /proc/stat > values, > normalised and displayed as percentage of the total time (similar to > mpstat): > > CPU %user %nice %sys %iowait %irq %soft %steal %guest %bh_dis > %idle > 0 2.01 0.00 46.73 0.00 3.02 23.62 0.00 0.00 24.62 > 0.00 > 0 2.49 0.00 44.28 0.00 2.99 21.89 0.00 0.00 28.36 > 0.00 > 0 1.52 0.00 44.95 0.00 3.03 22.22 0.00 0.00 28.28 > 0.00 > 0 1.49 0.00 48.51 0.00 2.97 21.29 0.00 0.00 25.74 > 0.00 > 0 3.50 0.00 48.50 0.00 3.50 23.50 0.00 0.00 21.00 > 0.00 > 0 2.51 0.00 44.72 0.00 3.02 23.12 0.00 0.00 26.63 > 0.00 > 0 3.98 0.00 47.76 0.00 2.49 23.38 0.00 0.00 22.39 > 0.00 > 0 2.00 0.00 45.00 0.00 2.50 24.00 0.00 0.00 26.50 > 0.00 > 0 1.51 0.00 49.25 0.00 2.51 23.62 0.00 0.00 23.12 > 0.00 > > The bh_dis field is the one I added. Clearly, this is not a negligible > amount of processor > time. Interesting data you brought, Elad. So clearly my theory is wrong. I encourage you to send your patch to LKML to get wider attention and feedback. To me, yes, I admit personally I have a feeling that not everybody follow the golden rule: disable bottom halves as fast as possible. So since everybody will likely agree that Linux kernel need to improve responsiveness and scalability, this kind of statistic improvement will be a great help. Thanks for sharing Elad.. regards, Mulyadi. -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
