Re: (NFS) 2.4.18 XFS 1.1 : Gave up on XFS - too many Oops

["Jameel Akari" <Jameel_Akari@xxxxxxx>]

Note that an Ethernet duplex mismatch will result in a very slow transfer
rate indeed.. I consistently see 47KB/sec when we find a 100Mb half-duplex
/ full-duplex mismatch, pretty much independant of protocol.  I've found it
to be painfully common among Cisco switches, from 2924 to 6500s.

Have you tried tweaking the rsize and wsize parameters on your NFS mounts
as well?
--
Jameel Akari
UNIX Admin
First Albany Corp


                                                                                
                                      
                    Andi Kleen                                                  
                                      
                    <ak@xxxxxxx>            To:     Walter R Fletcher 
<fletcher@xxxxxxxx>                             
                    Sent by:                cc:     linux-xfs@xxxxxxxxxxx       
                                      
                    owner-linux-xfs@o       Subject:     Re: 2.4.18 XFS 1.1 : 
Gave up on XFS - too many Oops          
                    ss.sgi.com                                                  
                                      
                                                                                
                                      
                                                                                
                                      
                    07/24/2002 01:27                                            
                                      
                    PM                                                          
                                      
                                                                                
                                      
                                                                                
                                      




>      Initially I used the SGI XFS ISO (RH7.1 I think) to set
> the system up.  The system ran OK except for frequent rpc timeout
> problems reported by the NFS clients but only during write operations.
> The problem could be made to happen by setting processes running on
> each client which wrote data to the NFS server at the fastest rate
> possible, i.e.:
>
>     dd  if=/dev/zero  count=<bignumber>  bs=<various>  of=/big/file

[...]

A common problem with many older linux kernels is that NFS is using a 64K
socket buffer only. This can cause packets to get dropped when the buffer
queue is overloaded on the NFS server. What you can do is to increase
/proc/sys/net/core/[rw]mem_{default,max}. This has to be done before the
kernel nfs server is initialized, e.g. in the NFS server start script
before the daemons are started.
Newer kernels should do the tuning automatically. Same can be done on
the clients.

Another common cause for dropped packets is a Ethernet half/full duplex
mismatch between switch and server. May be worth checking for.

-Andi