need help tuning ethernet and disk I/O, dropped packets during disk writes

[Wade Hampton]

G'day,

I need help optimizing a server so it can maximize incoming network traffic
(UDP) and
write the data to disks over a 100T network connection.  My system works
well with
no dropped packets until I turn on disk writes to the SATA disks.  I have
tried tuning
the network parameters but with marginal success.

   Motherboard/CPU:  Supermicro with dual 1000T interfaces, TG3
driver, 3.2GP4 (HT enabled)
   Disks:  dual 80G SATA disks
   Network:  TG3 driver
   O/S:  Fedora Core 4, 2.6.17-1.2142

Tuning options used:
   via kernel:  /proc/sys/net/core/rmem_max   2500000
                    /proc/sys/net/core/wmem_max  2500000
                    /proc/sys/net/core/netdev_max_backlog  20000
   via ethtool, interrupt coalesce itmes:
                   rx-usec  100
                   rx-frames 20
   via ethtool, RX ring size to max allowd by driver:
                   rx 511

Disk:
   originally SW RAID, experienced lots of errors when write was enabled
   changed to single partition, formatted with mke2fs -m 0 -b 4096, far
fewer write errors

I am still getting groups of dropped packets when cron.daily or weekly runs
and at random times.
The errors seem to be groups of missed packets and rarely a single packet.
I did find a similar
problem that someone else reported and am wondering if this could be the
same problem
(both are using Supermicro motherboards and the on-board TG3 driver):
   http://marc.theaimsgroup.com/?l=linux-ide&m=115689223417994&w=2

Any ideas?  Can I lower the priority of writes to disk?  Are there serial
options I can use?  Are
there updates to the tg3 driver I could try (for example really large rx
buffers beyond 511)?

Cheers,
--
Wade Hampton
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/fedora-list/attachments/20061204/5e382e61/attachment-0001.htm>