[Beowulf] Cluster for Finite Element Analysis
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Vincent Diepeveen diep at xs4all.nlMon Apr 25 05:34:52 PDT 2005
- Previous message: [Beowulf] CFP: APPT 2005 (deadline: May 1, 2005)
- Next message: [Beowulf] Cluster for Finite Element Analysis
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Hello Ghanta, a direct question: what is your budget for this and would it be a problem if there is sometimes a loss of data (in case the raid array crashes)? Please realize that very small clusters (like a node or 2) require real good software to be faster than a single dual machine. A cheap quad opteron might be faster than a 4 node network, the 4 node network having a higher price. A 8 node network build from real cheap machines with gigabit ethernet, will cause the total bandwidth over the network to be less than 200MB a second. Count at 50MB a second. Each shipment of data from a node to another one will cost CPU time, as gigabit ethernet doesn't use DMA like the highend networks. Also please consider that with 8 nodes and each node a small harddisk, that the ODDS are real real big that one of the disks will have a problem after a few months of operation. With 8 machines the chance is way way higher for that than with just 1 machine. A quad opteron is like $9000 only and even the most stupid software will be fast at it. Adding a disk array of 2 terabyte is real easy (for example using a fast raid-5 diskpack). It will be very hard to beat such a system with a network. Only real good software will and the network will be quite a bit more expensive in order to beat it. A good 8 node NETWORK is already $13000 or so (but then you really have something). And if each node is like $3000 (assuming a lot of RAM inside each node), then for like $40k (including a bunch of diskpacks) you have an awesome system with 16 processors that kicks the hell out of anything for you if the software can work over the network very well. Any cluster of $2500 will be slower for you than a $2500 dual opteron practical of course. Such contests are nice for university professors, but worthless for serious calculation work that require the network. There is a major difference between theory and practice. "If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?" Seymour Cray Vincent At 05:09 PM 4/22/2005 -0400, you wrote: >Hi, > >I am beginner and just started to learn about cluster technology. My goal is >to build a cluster that could be used for FEA ultimately. I have searched >the web and it confused me more on how to do this. I was advised that this >was a good place to ask questions and I would be getting some real help. > >1. Did anybody built a cluster for FEA application? > >2. What type of cluster should I select for FEA ? > >3. What Clustering software and other software packages should I be >considering? > >4. What are my minimum hardware requirements? > >5. Is there a set procedure on how I could proceed in building this cluster? > >Any advice on this issue will be of great help. Thank you for taking time to >help me out > >Regards > > >Srini > > >_______________________________________________ >Beowulf mailing list, Beowulf at beowulf.org >To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf > >
- Previous message: [Beowulf] CFP: APPT 2005 (deadline: May 1, 2005)
- Next message: [Beowulf] Cluster for Finite Element Analysis
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
