Interesting data on network interrupt - part II
Julian Elischer
julian at elischer.org
Thu Apr 6 19:10:55 UTC 2006
John Baldwin wrote:
>On Wednesday 05 April 2006 17:40, Julian Elischer wrote:
>
>
>>John Baldwin wrote:
>>
>>
>>
>>>On Tuesday 04 April 2006 20:33, Paolo Pisati wrote:
>>>
>>>
>>>
>>>
>>>>Hi,
>>>>
>>>>i updated my work on interrupt profiling with sone new
>>>>experiments.
>>>>
>>>>In total we have now:
>>>>
>>>>-FreeBSD 4 PIC (no asm part)
>>>>-FreeBSD 7 APIC
>>>>-FreeBSD 7 PIC
>>>>-FreeBSD 7 PREE APIC
>>>>-FreeBSD 7 APIC JHB
>>>>
>>>>Some quick comments:
>>>>
>>>>-PIC is much slower in masking interrupt (7k in PIC vs 3k in APIC)
>>>>-PREE let new thread save less than 500 ticks of 'queue' while
>>>>preempted threads are often resumed after a lot
>>>>-JHB patch shaved 2.5k ticks in interrupt masking op
>>>>
>>>>For graphs, data and more comments:
>>>>
>>>>http://mercurio.sm.dsi.unimi.it/~pisati/interrupt/
>>>>
>>>>
>>>>
>>>>
>>>I'll commit the patch then. :) One thing you might try to do to better
>>>measure the effects of preemption is to generate kernel work so that
>>>the bge interrupts occur while the current thread is in the kernel
>>>rather than in userland. In that case preemption should provide much
>>>lower latency for interrupt handlers, as w/o preemption, an interrupt
>>>in kernel mode won't run the ithread until either curthread blocks or
>>>returns to userland.
>>>
>>>
>>>
>>>
>>it looks a bit like the preempted threads shuld be put onto a stack of
>>threads to resume
>>so that when the pre-empter finishes, teh previosly active thread is
>>resumed.
>>Basically, a preempted thread should be put at the HEAD of it's run
>>queue, and not the tail..
>>
>>
>
>You changed the scheduler to already do that.
>
>
oh, yeah,..... at least I'm consistent..
More information about the freebsd-current
mailing list