Re: Secret of great tcpdump performance ..

[Rick Jones <rick.jones2 () hp com>]

Michael Krueger wrote:
> On Tue, 22 Jan 2008 19:47:24 +0100, Rick Jones <rick.jones2 () hp com> wrote:
>
> > How many processors do you have, are interrupts from each NIC going 
> > to  seperate processors/cores/whatever (show us the output of  
> > /proc/interrupts), and have you bound each tcpdump to its 
> > corresponding  NICs interrupt CPU?
> I have a single CPU / dual core machine. I think you are right. All of 
> the  interrupts are going to a single core:
> lxvoipmon05:~/perftest # cat /proc/interrupts
>            CPU0       CPU1
>   0:        113  168298887    IO-APIC-edge  timer
>   7:          1          1    IO-APIC-edge  parport0
>   8:          0          2    IO-APIC-edge  rtc
>   9:          0          0   IO-APIC-level  acpi
>  14:          0    6034770    IO-APIC-edge  ide0
>  50:          0   75531924   IO-APIC-level  libata, eth1
>  58:          0     848554   IO-APIC-level  libata, ehci_hcd:usb1
>  66:          0        394   IO-APIC-level  libata, ohci_hcd:usb2
>  74:          0          3   IO-APIC-level  ohci1394
>  82:          0   79599257   IO-APIC-level  eth0
> 177:          0    4140498   IO-APIC-level  eth2
> NMI:          0          0
> LOC:  168325561  168325456
> ERR:          0
> MIS:          0
>
> This was without binding tcpdump to a specific CPU. Anyway, tcpdump was  
> able to capture the traffic on both NICs without dropped packets. My 
> own  app almost immediatly reports dropped packets.
>
> What do I have to do so that interrupts are handled on both cores? Will  
> the interrupts move with the process if I bind them to separate cores?

First, you make certain that there is no irqbalance daemon running. 
Then you do:

echo M > /proc/irq/N/smp_affinity

where M is 1 << CPUnum (ie CPU0 would use 1, CPU1 would use 2, CPU2 
would use 4 etc etc) and N is from the table above.

rick jones
-
This is the tcpdump-workers list.
Visit https://cod.sandelman.ca/ to unsubscribe.