NFS client (10x) performance regression 2.6.14.7 -> 2.6.15

[Jakob Oestergaard <jakob@unthought.net>]

Hi guys,

I just found out... Installed 2.6.16.1 (32-bit) on a spanking new dual
opteron 275 (dual-core) machine, and saw that my link jobs were taking
ages.

I narrowed it down a bit - these are the kernels I have tested:
2.6.13.5:  Good
2.6.14.7:  Good
2.6.15:    Poor
2.6.15.7:  Poor
2.6.16.1:  Poor

Sequential NFS I/O is good on all kernels. Only "ld" shows the problem.

On 2.6.14.7, I can run a large link job creating a 60 MB executable in
15.6 seconds wall-clock time.

On 2.6.15, the same link job takes 2 minutes 28 seconds.

This is almost 10 *times* longer.

Testing with tiobench, I can see no notable difference between the
kernels (!)   It seems that this is very specific to ld.  I am using GNU
ld version 2.15.

The NFS client mounts the working directory using NFS v3 over UDP with
default (32k) rsize/wsize.

Since this machine is not in production yet, I can experiment with
kernel patches on it - I would like to try and narrow this down even
further - any suggestions as to which patches to exclude/include will be
greatly appreciated.

Note; I did double-check all the common problems (ethernet problems,
mount options, ...).

-- 

 / jakob