Newsgroups: comp.parallel From: "Greg Lindahl" Subject: Re: Good news for Beowulf Organization: Date: 28 Apr 1997 13:53:39 GMT Message-ID: <5k2a53$cif@server1.ctc.com> heirich@beazles.asd.sgi.com (Alan Heirich) writes: > I would be interested in knowing whether anyone out there has made > an effort to run a Beowulf (or other roll-your-own-shared-nothing-cluster) > for data base applications. We mostly talk about Beowulfs as servers for > cpu bound applications. And there are lots of scientific codes that run > effectively on them. And in particular, cost-effectively. But what about > commercial apps? Seems to me the pocketbook argument should play well > with the bean counter types, as well as scientists on a budget. A guy at Sun told me that a majority of Sun's sales these days is Ultra2 systems -- sounds like many commercial customers are running commercial apps on piles of sparcs. I'm not allowed to talk about any specifics about what I used to do until recently on Wall Street, but that's the general idea. You can point to several webservers that are piles of PC's -- sites with relatively small numbers of pages simply replicate them onto a pile of PC's, and other sites like AOL are using a pile of PC's (dunno what type) with the pages themselves on a Network Appliance fileserver. But the economies shift back and forth -- Network Appliance was talking about putting a http daemon on their box, so if you are just serving plain pages, it would probably be efficient to go back to a monolithic architecture. As for commercial databases like Oracle, I believe they do have a distributed memory version of the db. Combine that with fast networking such as Myrinet or SCI, and it could do quite well on a pile of PC's. I'm not sure which kinds of database workloads work well on such a platform, compared to shared memory platforms. But anyone with a database on an IBM SP2 could certainly use a pile o PC's instead, and I hear of big corporate kinds of databases on SP2's. No single database that I have personal experience with was big enough to need such a setup; I had a bunch of "small" databases, where small is 5-30 gigabytes. And the definition of small gets bigger every day as equipment gets faster&cheaper, while the trading volume of the stock market and many other fundamental business elements grow slowly. There's nothing really new in any of what I've said. Most companies are capable of examining the price/performance curve, and even if you're locked in to Solaris/Sparc, a savvy business will naturally buy piles Ultra2 boxes for any business task which is separable into sufficiently small chunks. The intel equivalent is a dual-PPro system. -- greg [ now working for the Legion project at UVa -- which is all about efficient, secure, fault-tolerant use of piles of separate machines. ] -- Articles to parallel@ctc.com (Administrative: bigrigg@ctc.com) Archive: http://www.hensa.ac.uk/parallel/internet/usenet/comp.parallel