Re: Linux v2.6.17-rc6

[Jakob Oestergaard]

On Tue, Jun 06, 2006 at 12:35:28PM -0400, Bill Davidsen wrote:
> Trond Myklebust wrote:
> >On Tue, 2006-06-06 at 14:07 +0200, Klaus S. Madsen wrote:
> >>Hi,
> >>
> >>We still experience the NFS client slow down reported by Jakob
> >>Østergaard in http://lkml.org/lkml/2006/3/31/82, even with 2.6.17-rc6.
> >>
> >>Trond Myklebust have created a patch which we have verified solves this
> >>problem for 2.6.16, 2.6.17-rc4 and 2.6.17-rc6. The patch is available
> >>from http://lkml.org/lkml/2006/4/24/320, and as an attachment to
> >>bugzilla bug 6557.
> >
> >The patch is already queued up for inclusion in 2.6.18. I'm not planning
> >on submitting it for 2.6.17 since it is not a critical bug.
> 
> I guess that depends on how much it slows down and how much you depend 
> on the speed of NFS. I have all of my machines sharing some local 
> binaries and docs, but the files are typically small and the network is 
> gigE, so I doubt it will hurt me. 

It hurts writing, not reading.  And only certain cases of writing.

But when it hurts it hurts quite a bit; slowdown is 10-20 times and gigE
most likely won't matter.

> On the other hand I do know people 
> running workstations with virtually everything NFS mounted, working with 
> large image files.
> 
> The initial bug report makes it look as if it's about two orders of 
> magnitude slower, but doesn't quantify the effect on more common 
> sequential access operations.

Sequential read and write seems to be fine. At least from my testing.

This was what made this so weird; normal testing with dd or bonnie etc.
showed all was fine. But link jobs over NFS were unreasonably slow.
Turned out you needed a somewhat special read/write pattern (similar to
ld) to trigger the slowdown.

So I guess it's probably mostly people who do development over NFS who
are hurt - and then only those who produce rather large executables.

> If this becomes an issue in 2.6.17, I hope it will show up in -stable 
> before 2.6.18, the current development cycle is a bit, um, protracted... 
> lately.

Me too :)

-- 

 / jakob

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/