I`ve experienced a very strange relationship between
ISA sb card generated interrupts and network NIC perfomance.
The situation was the next: i`ve performed a data transfer from
a remote samba share to the local box thru a 10Mbit ISA NIC.
What has been varying was the interrupt load of the sound card.
With esd started the sound card generated 500-600 interrupts/second,
with no esd, it didnt generated interrupts.
Theoretically the sbcard interrupts should make the NIC starve
and reduce the NIC performance. But the situation i`ve experienced
was weird as hell.
With esd started the samba transfer maxes out at 400kb/s.
With esd stopped it immediately drops down to 30-35 kb/s.
This is also well reflected by the NIC interrupt rate cutting down
by 90% with esd stopped.
Info: linux-2.4.17-rmap12e, NE2k ISA NIC, SB16 Vibra ISA
regards, Samium Gromoff
On Feb 19, 2002 16:18 +0300, Samium Gromoff wrote:
> With esd started the sound card generated 500-600 interrupts/second,
> with no esd, it didnt generated interrupts.
This is a known problem with esd - it plays "the sound of silence" all
the time. This means it is generating interrupts even when you aren't
listening to anything.
Cheers, Andreas
--
Andreas Dilger
http://sourceforge.net/projects/ext2resize/
http://www-mddsp.enel.ucalgary.ca/People/adilger/
Andreas Dilger wrote:
>
> On Feb 19, 2002 16:18 +0300, Samium Gromoff wrote:
> > With esd started the sound card generated 500-600 interrupts/second,
> > with no esd, it didnt generated interrupts.
>
> This is a known problem with esd - it plays "the sound of silence" all
> the time. This means it is generating interrupts even when you aren't
> listening to anything.
>
But the odd thing is that the soundcard interrupts tentupled the
samba throughput. Which sounds like a driver-handoff-to-ksoftirqd
problem.
-
> > the time. This means it is generating interrupts even when you aren't
> > listening to anything.
>
> But the odd thing is that the soundcard interrupts tentupled the
> samba throughput. Which sounds like a driver-handoff-to-ksoftirqd
> problem.
Or an IRQ routing bug
" Alan Cox wrote:"
>
> > > the time. This means it is generating interrupts even when you aren't
> > > listening to anything.
> >
> > But the odd thing is that the soundcard interrupts tentupled the
> > samba throughput. Which sounds like a driver-handoff-to-ksoftirqd
> > problem.
>
> Or an IRQ routing bug
>
Okay guys, new data on this issue.
the kernel 2.4.17 which i used so far had this issue - ie the transfer rate drop
400k -> 20-30k without esd.
i recently switched to 2.4.19-pre3 and the behaviour reverted:
now it drops 300k -> 50k due to esd active.
i think both behaviours are broken, cause the 2.4.17 case clearly shows that
ne2k perfectly runs at full speed with _active_ esd.
ie summarizing we have:
2.4.17: esd -> 20k bumps to 400k
2.4.19p3: esd -> 300k drops to 50k
regards, Samium Gromoff