Poor performance in NFS Cluster over UDP

[David_854]

Hello Everyone,
We have the following environment.
Two ES45 with a DEGXA PCI Gigabit Cards each as follows:
name = bcm0
category = network
sub_category = Ethernet
model = DEGXA
hardware_rev = BCM5703_C0
firmware_rev =
MAC_address = 00-10-18-0C-4E-39
MTU_size = 1500
media_speed = 1000
media_selection = Automatic


The NFS Environment we have is clustered running only samba.
The performance in the NFS Cluster is decreasing every day until it
reaches a point in which we have to reboot the cluster once a month to
clear out the cache and buffer. Then the services come back fine.
The following errors are constantly logged on to the /var/adm/messages
in relation to the


RFS3_COMMIT
RFS3_READ
RFS3_WRITE
RFS3_GETATTR, client address = xx.xx.xx.xx, errno 70
rfs_dispatch: sendreply failed IP address: xx.xx.xx.xx
rfs_dispatch: sendreply failed IP address: xx.xx.xx.xx
NFS server: stale file handle fs(2554,616540) file 857277 gen 32771


Our environment consists of over 40 TB of files in data being accross
the NFS.
Question:
Is there a problem with the retransmit of large file systems, since
most of our file systems are over 2 gig in size.


Also, while doing some tests internally we noticed the following in
MB/sec and while reading a file on being served by the CFS.


In server 1 the reads were at 11 and the writes at 25
In server 2 the reads were at 35 and the writes at 25


Questions:
Can someone tell me what is the high-performance blocksize for NFS V3.
If there is a diagram online that will make us understand this better
it will be greatly appreciated.


TIA


David