[OpenAFS] Re: Expected performance

[Andy Cobaugh]

On 2011-05-20 at 14:10, Andrew Deason ( adeason@sinenomine.net ) said:
> On Fri, 20 May 2011 13:51:05 -0400 (EDT)
> Andy Cobaugh <phalenor@gmail.com> wrote:
>
>>> ...or it's that you're writing to the same disk twice as much. If the
>>> cache and /vicepX are on the same disk, it seems pretty intuitive that
>>> it's going to be slower.
>>
>> It was with memcache.
>
> Well _I_ wasn't talking about memcache. :)

Of course. When I'm talking about performance, I'm almost never talking 
about disk cache ;)

>> mal and badger are slightly different hardware, but the tests above
>> show that we get very similar performance between all server and
>> client combinations except the case where client and server are on the
>> same machine.
>>
>> Maybe I'm missing something here?
>
> I think to some extent it can still be that they're just using the same
> hardware resources, so some performance loss is to be expected (if the
> network wasn't the bottleneck for separate machines). I'm not sure if
> that can explain that degree of difference, though. I believe Rx in the
> past has had some odd behavior that really low RTT, but any known fixes
> there should have been in 1.6 for awhile.
>
> I expect a similar thing happens on 1.4? Though of course the baseline
> performance is probably different, so it's not really comparing the same
> thing.

I think there were differences with 1.4, but it's been a while since that 
particular machine has run 1.4 that I don't remember exactly.

One would think that a modern 8-core box with 32GB of memory would provide 
for enough 'isolation' between server and client. Maybe I just don't know 
enough about what resources are involved in that case.

I'm still curious what's actually causing that much of a performance loss.

--andy