Thinking about HPC Infrastructure

re: Thinking about HPC Infrastructure

piano tuning cypress tx — Wed, 28 Jan 2009 00:01:09 GMT

Женщин умных не бывает. Есть прелесть какие глупенькие и ужас какие дуры.

What Lies Beneath: Setting up underlying HPC tools

Port 25 — Fri, 22 Dec 2006 21:55:45 GMT

After reading thru the previous blogs on HPC, someone might ask “What are some of the core components of HPC ?”. After all, once you’ve seen the outside of a Maserati or a Pantera DeTomaso, you’re not going to be satisfied just by ogling at it. Even after

re: Thinking about HPC Infrastructure

kishi — Mon, 04 Dec 2006 20:58:33 GMT

Good points, especially about the cabling. HPC actually does use almost as much, if not more than SAN. True - filesystem isn't so cookie-cutter however, it is an area where a lot of the OEM's are turning their focus towards more and more

re: Thinking about HPC Infrastructure

einhverfr — Sat, 02 Dec 2006 06:47:35 GMT

First of all, premature optimization is the root of all evil. One cannot come up with cookie cutter approaches to all aspects of HPC planning unless you have a pretty good idea of how the cluster is likely to be used.

However, many of the areas you address (most notably cabling and HVAC) are relatively easy to plan for ahead of time if you are aware of them. Generally the HVAC recommendation is to build in spare capacity. This is done by taking the total PSU capacity, assuming it is maxed out all the time and that all electricity is converted into heat. So if you have a small HPC cluster with 10 systems running on single 300W PSU's each, you plan on cooling the room as if you had a single 3000W space heater running in it at full blast.

Of course we all know that the PSU's aren't at full capacity all the time, that some energy gets emitted as noise, etc. but this gives some room for error, unplanned heat sources, etc.

Another point is the cable. I suspect that HPC in many deployments uses more cable per U than even a SAN, though this would depend on the network topology of each. Although I note that you recommend a SAN attached to an HPC cluster, so I guess that means lots and lots and lots of cable...

However... The filesystem optimization isn't so cookie-cutter. There may be cases where having a well engineered storage network (including both NAS and SAN technologies) may be a general win, and there may be cases where having some local storage may actually be beneficial.

On a side note, I haven't found much Microsoft support for OSD. Am I wrong? I would love to move my own SAN node offerings to OSD but I don't think the Microoft iSCSI drivers support that yet. I hope it is on the roadmap.