irqbalance on HT capable PE4600?

jason andrade jason at rtfmconsult.com
Sun Oct 19 07:50:01 CDT 2003


On Sun, 19 Oct 2003, Arjan van de Ven wrote:

> this batching is clearly more efficient (far less switches) and in
> addition to the explicit steps shown above, it also improves cache
> locality much.
>

nod, this makes a lot of sense.

> >  20:  799452491  799875894   IO-APIC-level  qla2200
> >  22: 2366882687 2368370608   IO-APIC-level  qla2200
> >  24: 1316729851 1317548294   IO-APIC-level  qla2200
> >  26: 1147601730 1148380357   IO-APIC-level  eth2
>
> this behavior may be the least optimal possible ;)
> if the irq's alternate between cpu's it's quite likely that fragments of
> your UDP packets (NFS sends 4Kb or 8Kb UDP packets over ethernet with a
> 1500 byte mtu so those UDP packets get split up) end up on alternating

actually we are sending 1K packets because i have never been able to
get 4 or 8K nfs mounts working on RH (ever) for high loads so everything
is mounted with 1024 read.  i am *really* hoping this is fixed in RHEL/ES3

> CPU's, which means that in order to combine them, half the
> packet (well technically only the metadata) needs to be transported
> from one cpu to the other... which is relatively expensive.

hmm, food for thought.  i'd assumed a SMP server would be better for
nfs serving as i was making an assumption that the network stack
would bind to one cpu and applications might be able to bind to
(hopefully) the other.

with this in mind would you recommend a single cpu machine as a better
alternative or are there other ideas ?

regards,

-jason




More information about the Linux-PowerEdge mailing list