[Beowulf] Performance characterising a HPC application

[stephen mulcahy]

Mark Hahn wrote:
> I don't follow why that indicts latency - multiple smaller packets don't
> each require a round trip, for instance.  with TCP, I've only
> ever seen jumbo packets resulting in modestly higher bandwidth and often
> noticibly lower CPU overhead.  TCP with 1500B packets will _certainly_
> have multiple packets in flight, so on a lan is not terribly
> latency-sensitive.
> 
> if this traffic is to some hotspot, I'd be more inclined to think that
> small packets are overloading the CPU with interrupt and TCP-stack
> overhead.

True. I guess I'm leaning toward latency because I don't see bandwidth
as being the issue - if the overhead is lots of small packets
overloading the CPU shouldn't I see a lot of system time on some nodes?


>> looking for. Do you eyeball raw tcpdump data or use wireshark to
>> browse it?
> 
> I use perl to browse it ;)

Hehe, fair enough. I am well versed in the art of perl hacking, just not
sure it's the best way to use my analysis time.

> pricing is always squishy for this category of hardware.  you can see
> the myrinet list prices on the myricom website.  for 16 ports:
> 16*(495+75)+6600
> comes to just under $1k/port.  using local (Canadian) HP public-sector
> prices, IB is $1600/port.  I'd be surprised of both couldn't be bettered.

Thanks - interesting to get even this kind of comparison, I'm sure its
in the right ballpark anyways.

-stephen

-- 
Stephen Mulcahy, Applepie Solutions Ltd, Innovation in Business Center,
   GMIT, Dublin Rd, Galway, Ireland.      http://www.aplpi.com