Tyan S2466 MPX integrated ethernet interrupt happy

I have an S2466 with an integrated 3COM 3C905C ethernet controller. What I've noticed when streaming tons of data accross my internal LAN is that the ethernet driver appears to be sucking up tons of cycles.

A quick investigation of /proc/interrupts shows-

  0:     195949     192414    IO-APIC-edge  timer
  1:       5126       5129    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
  5:          0          0   IO-APIC-level  usb-ohci
  8:          1          0    IO-APIC-edge  rtc
  9:    4248819    4245969   IO-APIC-level  eth0
 10:     144865     145243   IO-APIC-level  usb-ohci, nvidia, EMU10K1
 12:      48546      48865    IO-APIC-edge  PS/2 Mouse
 14:     122395     121652    IO-APIC-edge  ide0
 15:      26286      26498    IO-APIC-edge  ide1
NMI:          0          0 
LOC:     388204     388212 
ERR:          0
MIS:          4

So, approximately 8.5 million ethernet interrupts. The system is noticably slower when streaming ethernet data, and it's sucking up a lot more processing time than on my other much slower other box. This box is running a stock 2.4.18 kernel from kernel.org (i.e. no custom hacks). It's running 2 1800+ XPs.

I'm wondering if anyone has seen this and has a quick solution. If not I'll take a look through the 3com code when I get back from vacation next Sunday. (Leaving in a few hours)

Thanks,
Shane Nay.

That all looks normal. Could you be more specific about the performance problems? How much slower? Output from top' and ps'? Any nasty messages in the system log?

Looking at the ethernet driver won't help, I expect - it's as efficient as most any other driver. If there is a problem, it lies elsewhere...

A kernel profile would be illuminating. Build the kernel with as few modules as possible, boot the machine with `profile=1' and play with readprofile.

-

The table is just the number of ints, see below for one of my boxen which is running 200-250 users and has idle of ~92% over three days.

newssvr17:twister$ cat /proc/interrupts
CPU0 CPU1

  0:   14112007   14110727    IO-APIC-edge  timer
  1:        473        467    IO-APIC-edge  keyboard
  2:          0          0          XT-PIC  cascade
  8:          0          1    IO-APIC-edge  rtc
 12:        104        105    IO-APIC-edge  PS/2 Mouse
 14:          3          2    IO-APIC-edge  ide0
 18:   88268186   88280808   IO-APIC-level  eth0
 20:    6655721    6658714   IO-APIC-level  ips
 27:   74212944   74213812   IO-APIC-level  PCnet/FAST III 79C975
 28:         14          2   IO-APIC-level  aic7xxx
 29:         10          6   IO-APIC-level  aic7xxx
NMI:          0          0 
LOC:   28223102   28223059 
ERR:          0
MIS:          0

bill davidsen <[email protected]>
  CTO, TMR Associates, Inc
Doing interesting things with little computers since 1979.

No nasties in the syslog. System usage from top averages 50% per CPU. (By system I mean kernel usage, bounces between 70 and 30 per CPU)

Gotcha, yes that could definetly be. This was extraordinarily unscientific. I was thinking maybe someone had seen something else like this on this platform. So, I'll write a test that isolates ethernet from other stuff like ide performance, etc.

Will do when I return.

Thanks,
Shane Nay.
(Someone else noted their /proc/interrupts for the same machine, but my box had only been up for 1.5 hours :)