kern/131360: [nfs] poor scaling behavior of the NFS server under load

[Peter Keel]

The following reply was made to PR kern/131360; it has been noted by GNATS.

From: Peter Keel <freebsd at discordia.ch>
To: bug-followup at FreeBSD.org, martin at email.aon.at
Cc:  
Subject: kern/131360: [nfs] poor scaling behavior of the NFS server under
	load
Date: Thu, 26 Feb 2009 16:50:18 +0100

 We experience exactly the same problem. With FreeBSD 7.0 everything was 
 alright, but with the upgrade to 7.1 performance became abyssmal. Normally
 it seems to work for some time; but after a few hours, the load climbs 
 and the 4 nfsd-processes each use 100% CPU (on a 4way-system). The system 
 only serves as nfs-root (ro) for about 20 systems, so there's not too much
 nfs-traffic going on. 
 
 I wasn't able to capture it when every nfsd was using 100% CPU (because, well,
 some thousands of users depend on it); this is what it looks when it's 
 half-working (hundreds of timeouts on the clients): 
 37907 root        1   4    0  4604K  1112K -      2  46:29 33.59% nfsd
 37908 root        1   4    0  4604K  1112K -      2  15:29 17.38% nfsd
 37909 root        1   4    0  4604K  1112K -      3   5:59  5.66% nfsd
 37910 root        1   4    0  4604K  1112K -      2   2:38  0.00% nfsd
 
 We're suspecting the ULE scheduler as responsible for the mess. Right
 now we're testing a kernel with the 4BSD scheduler; we'll know more
 next week.
 
 But now it looks like this:
  1040 root        1   4    0  4604K  1068K -      3   5:49  6.49% nfsd
  1039 root        1   4    0  4604K  1068K -      3   2:20  1.86% nfsd
  1042 root        1   4    0  4604K  1068K -      1   0:48  0.05% nfsd
  1041 root        1   4    0  4604K  1068K -      2   0:15  0.00% nfsd
 
 Promising.
 
 Regards
 Peter
 -- 
 "Those who give up essential liberties for temporary safety deserve 
 neither liberty nor safety." -- Benjamin Franklin
 "It's also true that those who would give up privacy for security are 
 likely to end up with neither." -- Bruce Schneier