Hello,

I've just recently aquired a Dual Intel Xeon 3 Ghz server and started noticing that under heavy load a process eats up a lot of resources: ksoftirqd_CPU0.
As far as I've read this means the cpu is busy treating interrupt requests ( from the network activity I have).

At the same time I see other processes like ksoftirqd_CPU1, CPU2, CPU3 which don't eat up any cpu time, so that in the end I have something like CPU0 90% full, CPU1, CPU 2 and CPU3 2% full.

I wanted to ask you if it's possible or if there's any way to make the system treat interrupts with both CPU's I have (which appear as 4 because of hyperthreading).

I'm currently running a linux kernel 2.6.11 with IRQ Balancing enabled.

It's probably an issue with the applications being used that don't utilize more than one cpu... but that's my guess.

Nice try skyraven, but I think that your wrong. ksoftirqd is nuisance in the 2.4.x kernels. This kernel mutant hogs cpu time. A search on Google shows that ksoftirqd is known to hog 99% of CPU resource, even on important servers that serve enterprises. ksoftirqd makes my ***mobile*** laptop spin the harddrive madly and becomes unresponsive with 99% CPU usage. That pisses me off. The first time it happened I thought I was being attack via malscript javascript in mozilla or something. I have a 2.2Ghz Celeron and 256MB ram and a harddrive that can sustain 6MB per/second. That's more than enough firepower for surfing, emailing, IRC, etc. Normally my systems is 99.7% idle while surfing, but when ksoftirqd comes to life the processor is 0% idle and ksoftirqd hogging 99% processor time. System unresponsive for seconds at a time. Systems admins and others are having "hard biscuits/tough biscuits" with this phenomenon.

Good thing that I already have 2.6.x kernel hand.

ksoftirqd is simply one of the peculiarities in Linux. You can see IMPORTANT help info on this page, also about a potential method of disabling ksoftirqd (found it via google):

http://www.linux.org/docs/ldp/howto/...s-HOWTO-5.html

Stuff like this is negative to the adoption of Linux in the desktop and server arena. It could eventually permit the *BSDs and other Unices to overrun Linux due to these "surprise peculiarities".

Mobile Celeron.

ksoftirqd wakes up and executes qeued jobs-look at the website above. That's what I forgot

in a way i think i agree with trickykid though without knowing more it's just a guess
softirq is just Linux version of what we would call interupt "bottom half" and all kernels have them in some form or another but the point is the hardware interupt requests are being made by a driver and if it's the same driver the bottom half functionality might really benefit from being executed over and over on the same cpu. this is called cpu affinity and it means that cpu can keep the code needed to execute the softirq in mmu cache and do it alot faster that spreading it around. The other cpus would have to eject what they have in cache and go to RAM to read functions and data while cpu0 is good to go without doing that. If it was spread around you would have memory thrashing.

you mean that your kernel has (CONFIG_IRQBALANCE) enabled?
If no, you will map the irq per CPU, and get scenarios like this one.

Regards,
Breno Leitao