On Wed, Dec 06, 2000 at 01:13:00AM -0500, Andrew Gillham wrote:
> Tracy J. Di Marco White writes:
> > 
> > NetBSD and Linux are the NFS clients.  Linux is slower, I believe they've
> > only gotten v2 to work.  I've gotten the NetBSD boxes to be slightly faster,
> > I've changed -r & -w to 16K.  I'm not sure if that's led to improvement
> > on the Linux boxes, I'll be checking benchmarks on them later today.
> 
> Here are some numbers from bonnie on my machine.  The NFS server is a
> 1.5_ALPHA2 system with a PII-400, 128MB, an epic ethernet (100FD), and
> a 20GB UDMA/33 drive.  The client is a Duron 950, 256MB, fxp (100FD), and
> it is running 1.5L with UBC.
> The input numbers are not accurate due to having 256MB and UBC.  I will test
> some more with larger files.
> Anyway, I tested 64k/32k/16k/8k udp v3 mounts, a 16k udp v2, and a "default"
> tcp mount. (a "default" v2 mount performed similar to the 16k v2 mount)
> The difference between a v2 mount and a tcp mount was astonishing.  I did
> not expect to see a 5x difference!
> 
>               -------Sequential Output-------- ---Sequential Input-- --Random--
>               -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
> Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
> 65k       100  5580 12.9  5784  7.2  5138 11.7 47708 102.7 230898 101.4 800.6 19.1
> 32k       100  6192 14.3  6268  7.7  5307 12.6 47860 102.1 230648 100.9 2088.3 39.0
> 16k       100  6392 15.0  6381  7.5  5315 10.4 47880 101.3 231486 100.8 2111.3 48.4
> 8k        100  3187  9.2  3188  4.4  3194  5.1 47842 112.8 234803 106.5 2614.4 47.8
> 16kv2     100  1793  4.7  1789  2.4  1938  5.2  8111 22.0 234653 99.2 1752.2 37.9
> tcpmnt    100  8772 23.9  9463 11.6  5969 11.6 47381 98.2 223368 98.6 1816.6 39.3

I don't understand why TCP is faster. I would expect UDP to be faster as
the protocol overhead is lower ...
However I noticed the same thing with Solaris, so maybe it's just the NFS
protocol wich is different between TCP and UDP.
I hope the problem with TCP mount will be fixed one day, so that we can
default to TCP mounts again :)

--
Manuel Bouyer <bouyer@antioche.eu.org>
--