[DRBD-user] degraded write performace

Duane Cox duanec at mail.illicom.net
Thu Mar 29 16:24:52 CEST 2007

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


> What kind of NIC? My recollection is that the 2650 has the
> TigonIII (Broadcom) NIC.

That's correct, it's a Broadcom BCM5703CKHB chip using the 'tg3' driver in
the kernel source.

> Personally, I've more trouble with them than
> you can throw a stick at.

Really?  To date, I've not ran across anything or ever had any issues, until
now, which I'm not yet convinced it's network related.

> If you can get 'em, use the Intel or (slightly
> less preferable) the RTL8169. I've had other, less pleasant experiences
> with other chipsets in gig-e at high load.

I think I have a few "add-in cards" that are GigE, but I think they are
based on BROADCOM too.
Maybe it's time to shop for those Intel cards you suggested, let me poke a
few more sticks at it first.

> Are you REALLY REALLY sure that it's not network problems?

No, I keep an open mind about things until I know for sure either way, but
this is what I do know...

Running ethtool -S produces expected results... no errors, no collisions,
packet counters are correct...
The only thing that stands out is this: dma_readq_full: 59681 (and counting
during syncs and high load) on the primary.

Running "iperf -c ip.ip.ip.ip.ip" and "iperf -s" produces 991 MB/sec
transfers at 9000 MTU and 941 MB/sec transfers at 1500 MTU

>
> Have you tried something really simple like a basic TCP benchmark
> between the two machines?

Yes, see above.

How would a network issue effect the write throughput to drbd in a
disconnected state?
I could be wrong, but I don't think that's possible, and at this time, I
don't think it's network related.

Thanks
Duane Cox

>
> --
> Jon Nelson <jnelson-drbd at jamponi.net>
> _______________________________________________
> drbd-user mailing list
> drbd-user at lists.linbit.com
> http://lists.linbit.com/mailman/listinfo/drbd-user
>




More information about the drbd-user mailing list