I have an enduser that wants to purchase a Paralogic Beowolfe cluster in order to perform parallel computations of a major flow diagram's points. He says that using the beowolfe and linux INSTEAD of a Sun or Sgi will save money and achieve the same results. Problem is that Paralogic is only a year or so old. The setup is something like: 8 nodes of Pentiun III - 500 (512 MB Cache), 256 MB ECC SDRAM, 6.4 gb hd, with a single host node of 13.1 GB HD, 3D openGL, and 3dfx Voodo video... etc. Can anyone give me some advice or stories of experience in this area. Seems my users department is short on budget yet deep in work (like everyone now-adays). If I need to provide further details, I can. Thanks

We've been investigating a Linux/Beowolf cluster at my company. We've talked with several people using such setups for the radiation application, and they're all very happy. If you're concerned about Paralogic's age, consider Carrera or Microway.

1 question and 1 comment. the Beowulf concept was supposed to be a do-it-yourself supercomputer but I guess knowadays people will find the smallest excuse to buy something. lots of people have built their own Beowulf clusters (and this one is really small) why aren't your clients considering building their own? Comment: go for the gusto the PIII/500 is already obsolete, if thats what this company is offerring then they're probably not worth it. Why not one of these new gigahertz class machines. also your client will need more per node memory (get as much as will fit especially if they're using a commercial code) and the host node could do with more hd capacity unless they have more storage somewhere else.

We have a 64 processor cluster from Paralogic, and we are extremely pleased. They have been very supportive, and the prices are pretty good. Don't let the one year in business scare you. Doug Eadline is a very well respected member of the beowulf community (I have been on the mailing list for the last year). We run Cobalt (http://www.va.afrl.af.mil/vaa/vaac/COBALT/) on the cluster with a linear speedup on 64 processors.

One person commented that they are not with the times because they are not selling the fastest processor. This is not true. When you are talking about parallel computing, what matter is price/performance. If you work the numbers with a P933 vs say a P800, it turns out it is better to buy more of the slightly slower processor. 500Mhz is too slow, though - there web page is just out of date - they do sell faster.

(1). Well, in principle yes, it is the price/performance ratio. (2). But, don't you think that you need 64 spare parts of everything? (3). At some point, you will be replacing hard drive every week. And at that point, it is hard to decide whether you should use the already obsolete spare parts or the new hardware available at that time. (4). So, from reliability point of view, 64 processors approach is going to create more problems. And it is hard to prove that the peak price/performance ratio can be maintained for the life of the system. (5). And you also need more space 64 times, more cooling capacity to house the system. (6). But this approach does open up one effective way to use the obsolete processors(low speed CPU), as long as the power outlet of the building can handle it. So the old computer life can be extended through the parallel system. (7). It is just a different point of view. Your approach is one good approach, but not the optimum one, not even a good one from reliability point of view.

i agree with you John. The price/performance benefit he'll get up front by buying a system w/ older processors is going to be lost probably many times over by the cost of slow runs later on. moreover he'll have to replace his equipment more often. Isuppose there's an optimum balance between the extremes but I don't think it'd be very far from the highest performance end due to the relatively low cost of computing compared to salaries of engineers whose time would be wasted by slower computing.

(1). Talking about the salary, I was amazed by the labor cost to replace a rubber timing belt in my Honda car yesterday. (2). If your Honda car use a rubber timing belt, you must replace it before 90,000mile service, otherwise, it will cause great damage when the belt fail. (3). The dealer chargeed $49. for the timing belt, but the total cost is about $490 which include the labor cost of over $312. (4). I also had the two axles replaced at the same time. And they also did 60,000mile service of changing oil, filter, spark plug etc. (5). So, in about one afternoon of time (say 5 hours), the total labor cost was over $600, which give you some idea of the labor cost of $120/hour. (6). Anyway, people love to have workstations and network instead of the old main frame computer. In this way, you have more jobs and work to do. (7). I think, the initial idea of using a rubber timing belt was to save cost. But in the end, the user must pay more to get the mechanic install the belt. ( Well, he also has to remove the other components before he can reach the timing belt. So, there are more work to do.) Also, timing belt failure will not give you early warning, so the design is bad. (8). If the Air Force is trying to save the cost, it is likely that it will run into the situation of Russian sub problem.