Re: ksoftirqd uses 99% CPU triggered by network traffic (maybeRLT-8139 related)

[hector@marcansoft.com]

> On Tue, 27 Jul 2004, Robert Olsson wrote:
>
>>  > Yeah, when the ksoftirqd is taking all the cpu it will be like that,
>> but
>>  > when the kernel is behaving normally the starving diff is between
>> 0->1sec.
>>  Well ksoftirqd makes your kernel load just visible which is good and
>>  ksofirqd gets accounted for this when softirq's get deferred to it.
>>  It may look like goes from 0 to 100% but thats probably not the case.
>>  The problem is we can starve userland at high loads. As said we were
>>  trying some way to cure this I may have some old patch if you like to
>> try.
>
> Ok, as I said before I'm willing to test your patches.
>
> It would be nice that one could use the full capacity of his/her computer.
> This is not a big problem for everyday use for a workstation but prevents
> 2.6-series to be used in production-enviroments in the servers.
> But hey.. we need to do some work and maybe we will resolve this. =)
>
> --
> Pasi Sjöholm

Hi all, and first of all thanks for all the feedback!

Currently i'm on vacation (reading this through webmail, if you notice
anything strange please tell me) and I won't be able to test anything, but
Monday I'm back.

As I mentioned in the first e-mail, I can reproduce this under heavy
_packet_ load, UDP load actually, with 5-20 byte packets at a moderate
rate. No other CPU usage is needed to reproduce it, so it seems to be more
related to the number of packets than to the bytes/sec. Presumably more
packets means more interrupts, hence this behaviour. Also, the fact that
this packet stream is actually a debug-mesage protocol stream which
basically gets dumped to the screen (each UDP packet is one line) might
account for a slightly increased CPU load.

I'm willing to test any patches and other stuff that may be needed, I have
free time currently.

BTW this actually is a pretty big pain in the ass for me, since I need
those messages to get feedback from an application I'm developing. I can
reduce them, but then I get less feedback.

This seems to be something resource-related since the first 15-30 minutes
all goes well, but it just starts (not having touched the computer) and
then there's no going back. I can (with some effort) close down X and
telnet in, but performance is very bad (even the framebuffer screen takes
some time to refresh after a line scroll--and that's a kernel job AFAIK)

BTW when I get back I'll test it with the old 10mbps card that I use for
the external connection (internet) to see if it still happens with the
other driver. I might be able to try other cards as well.