Is your application really strictly CPU bound or is there a memory/IO component?
You might be setting your bar a bit low. You could potentially get a full rack of 1U servers for less than that (40 U * 8 cores * 2 CPUs). I'd recommend looking at the Supermicro Twin Squared if you want to stick in half rack configuration. You'll get better power efficiency and screwless maintenance/replacement. Dell has a similar product for a similar price, but not as easily maintainable. The twin squared has 4 servers in 2 U and each is on an independent sled. I have a favorite supermicro dealer who will rack the whole thing, cable it to my specifications, do an acceptance test, and ship it as one unit for a relatively low install fee. They even label the machines and cables and color code according to our specs. Finally, they give us a spreadsheet with all of the IPs, mac addresses, etc. as part of the deliverable. Let me know if you want a vendor reference. (shipping an entire rack as one unit is optional) Also, we too, use rocks. If you plan to run a CPU intensive load, you can tell them to run something like Linpack across all nodes for, for instance, 4 hours before delivering. We did something similar for an Infiniband cluster. Also, that no node should be more than ~15% below the average Linpack numbers for the cluster. You can specify similar memory benchmarks. Linpack is a good, general purpose, HPC benchmark. On 11/16/2010 7:22 PM, Hugh Brown wrote: > Hi everyone -- I work at a university in a smallish department, and > I'm helping a faculty member craft an RFP for a smallish cluster: > about $250k budget, taking up half a rack or so, and around 500 cores. > We plan to manage/install the cluster using ROCKS. I expect bids from > the usual medium/big vendors. It isn't huge, but it's a couple of > firsts for me: > > * First cluster (woot) > * First RFP of this size > > I'm concerned about acceptance testing, especially after hearing about > the problems with the Intrepid cluster[1] at the latest LISA > conference, and I wanted to ask what other admins have done about > this. > > My questions: > > * Have you included acceptance testing for large hardware purchases > before? Why or why not? > > * What have you specified in the tests? Do you have a template you > can share? How specific did you have to be in your RFP about what > constituted failure? > > * What resistance, if any, have you found from the vendors? What > about management? (I heard stories at LISA about acceptance testing > being whittled down in negotiations in return for concessions on price > or other things.) If you had to negotiate, what did you give > up/trade? > > * Have you had to reject anything, or send anything back to be fixed? > What reaction did you get from the vendor? > > I've found one paper[2] on this subject -- very helpful -- but I'm > hoping that someone can chime in on the larger/squishier (to use Adam > Moskowitz' term) issues around this. > > I realize these are basic questions, but I hope I'm on the right side > of the "how do I do this" vs. "please do my work for me" line. > > Thanks, > Hugh > > Footnotes: > [1] http://www.usenix.org/events/lisa10/tech/full_papers/Wilson.pdf > [2] http://www2.cs.uh.edu/~openuh/hpcc07/papers/10-Muller.pdf > -- > Hugh Brown > http://saintaardvarkthecarpeted.com > Because the plural of Anecdote is Myth. > > Footnotes: > > > _______________________________________________ > Discuss mailing list > [email protected] > https://lists.lopsa.org/cgi-bin/mailman/listinfo/discuss > This list provided by the League of Professional System Administrators > http://lopsa.org/ _______________________________________________ Discuss mailing list [email protected] https://lists.lopsa.org/cgi-bin/mailman/listinfo/discuss This list provided by the League of Professional System Administrators http://lopsa.org/
