[tech] morwong NFS speed problem

[Adrian Chadd]

On Tue, Jul 12, 2005, Alastair Irvine wrote:
> On Tue, 12 July, 2005 at 02:15:24PM +0800, Adrian Chadd wrote:
> > Alastair: get some hard figures. "pretty crap" is very useless.
> > Some throughputs with testing examples would be nice.
> 
> Fine.  (Although I should not have to justify myself here unless my
> judgement counts for nothing.  I've been around long enough for "pretty
> crap" to indicate something serious.)

I'm talking about being objective. This isn't an attack against you.
This is trying to figure out what is "crap" so we know, when we change
something, that its actually had a positive affect.

>   morwong% time cat {~50MiB file} > /dev/null
>   0.006u 0.872s 12:04.65 0.1%     0+2k 5900+0io 0pf+0w
> 
>   mussel% time cat {~50MiB file} > /dev/null
>   0.006u 0.364s 0:05.01 7.1%      0+0k 0+0io 0pf+0w

The trouble with this is:

* you haven't taken into account local buffering;
* you didn't tell us which share this was;
* you only did the test once

> And also,
> 
>   mussel% time cat {~2MiB file} > /dev/null
>   0.000u 0.014s 0:00.34 2.9%      0+0k 0+0io 0pf+0w
> 
>   morwong% time cat {~2MiB file} > /dev/null
>   0.003u 0.053s 0:31.50 0.1%      0+2k 238+2io 1pf+0w

Again, buffer cache, no share information, one test.
We don't know how busy things are.

The other things to take into account:

* mussel is a 700ish MHz PIII
* morwong is a less fast alpha (as leighton said, a piece of poop)
* mussel has >512mb ram
* morwong has what? 256mb?

Is mutt on morwong compiled 32 or 64 bit?

> For those who haven't used the csh "time" command before, those first three
> fields in the output are userspace time-in-seconds, kernel time-in-seconds
> and elapsed time in minutes:seconds.  The sixth field is supposed to be
> number of blocks read/written during I/O operations (I can't figure out why
> this is 0 on Linux).
> 
> And for the record, the "pretty crap" measurement came from running mutt
> on a ~8MiB mailbox.  It loaded very slowly on morwong and quickly on
> mussel.

What would be nifty here is a pcap of mussel and morwong, with
timestamps, of said mailbox loading. It'd be big, it'd be messy,
but you could then use one of the nice GUI tools to look through the
packet exchange and see how long each of the NFS ops are taking.





Adrian