Rob van der Heij writes:
> Malcolm Beattie wrote:
>
> > Start with
> >     # echo 262144 > /proc/sys/net/core/wmem_max
> >     # echo 262144 > /proc/sys/net/core/rmem_max
> > to raise the maximum read/write buffer size internally from 64K to 256K.
>
> And you did not even ask whether he's running on VM or not.
> The virtual machine not being dispatched by CP in time can be a
> cause for packet loss and require retransmission, in which case
> a large tcp window might even make you go slower. Given enough
> competition the same would apply to LPARs with shared CPUs.

Good point. My bad.

> I'm not sure where your 256K comes from, but I would be very
> reluctant to spread this kind of rules-of-thumb.

I did some investigation and test of high performance networking
with Gig-Eth (non-390) and raising that limit helped. I also did
some tests with 390 networking performance (vctc and iucv under vm)
and tweaked some other things. I hadn't thought carefully enough
that the rmem_max/wmem_max change wouldn't be a generic speedup.

> reluctant to spread this kind of rules-of-thumb. I'm afraid
> QDIO performance turns out to be more complicated than this.

Yeah, I know... :-)

--Malcolm

--
Malcolm Beattie <[EMAIL PROTECTED]>
Linux Technical Consultant
IBM EMEA Enterprise Server Group...
...from home, speaking only for myself

Reply via email to