Christian Weisgerber wrote: > Claudio Jeker <[email protected]> wrote: > > > Here is a diff to support the newer Broadcom chips seen in Dell and > > HP > > servers. This was tested against a BCM57765, BCM5721, and BCM5720. > > This needs a lot of testing on ANY bge(4). The diff will most > > probably > > only apply to a very current -current. > > This patch is very bad. > > It totally kills NFS performance in at least one of my usage cases. > I mount a CVS repository by NFS and run cvs up against it. With > this patch on the NFS server (BCM5761 A1), cvs up slows down by a > factor of ten or so. I'm at a loss to explain how this happens; I > don't see anything suspicious in netstat -I or -s. Ping doesn't > show any unusual latency, bulk TCP traffic doesn't seem to be > affected. The overall effect is perfectly reproducible, though. > > I get the same effect if this patch is present only on a client > machine (BCM5702/5703 A2). > > -- > Christian "naddy" Weisgerber [email protected] When I see the above, it usually implies that the receive side is dropping packets when it receives a burst of traffic. (Especially when it is also transmitting. Having simultaneous traffic in both directions is something NFS RPCs generate more often than most net traffics.)
Another culprit for this I've seen is TSO (if the driver is doing it). If possible, you could try disabling TSO and checksum offload. Good luck with it, rick
