On Thu, 8 Mar 2001 kuznet@xxxxxxxxxxxxx wrote:
> Hello!
>
> > In my experience upping it can increase NFS performance a lot.
>
> Indeed, 256K is too low. Even TCP needs ~512K window on gige.
>
> Actually, gentle hint to Bob was to estimate minimal threshold,
> when evicton completely stops on myrinet. This can be taken as
> new default, adjusted to lower values for low memory configurations.
>
> Alexey
>
I tried a little yesterday, but here's more info.
Upping the threshold does not help a great deal.
I went all the way to 4Meg and still saw
similar behavior of reassem fails.
Between 1 and 2
echo "1048576" > /proc/sys/net/ipv4/ipfrag_high_thresh
between 2 and 3
echo "2097152" > /proc/sys/net/ipv4/ipfrag_high_thresh
between 3 and 4
echo "4194304" > /proc/sys/net/ipv4/ipfrag_high_thresh
Then I ran a netperf udp test of about the same length.
grep "reassembles failed" /tmp/1 /tmp/2 /tmp/3 /tmp/4;
/tmp/1: 608545 packet reassembles failed
/tmp/2: 881139 packet reassembles failed
/tmp/3: 1128584 packet reassembles failed
/tmp/4: 1322635 packet reassembles failed
I think the culprit here is the packet receive errors from
UDP. These are socket overflows I think. I'm trying to
track it down. When I used DaveM's zero-copy patches on
a linux-2.4.0 kernel, most, if not all, of these
packet receive errors went away.
grep "packet receive errors" /tmp/1 /tmp/2 /tmp/3 /tmp/4
/tmp/1: 621834 packet receive errors
/tmp/2: 826133 packet receive errors
/tmp/3: 985257 packet receive errors
/tmp/4: 1140693 packet receive errors
|