From owner-chemistry _-at-_)ccl.net Sat Oct 1 12:08:00 2005 From: "Perry E. Metzger perry.:.piermont.com" To: CCL Subject: CCL: W:hardware for computational chemistry calculations Message-Id: <-29450-051001104957-18075-PSdynCfptnDCkWeF/srY3A-,-server.ccl.net> X-Original-From: "Perry E. Metzger" Content-Type: text/plain; charset=us-ascii Date: Sat, 01 Oct 2005 10:49:52 -0400 MIME-Version: 1.0 Sent to CCL by: "Perry E. Metzger" [perry(~)piermont.com] "Perry E. Metzger perry===piermont.com" writes: > Lets say you have a computation that is I/O bound on access to a 10G > file. Right now, an additional 10G of DRAM will cost ~$1200. Er, I'm not being quite 100% careful there. The price can end up being twice that or more depending on speed and configuration, and only rare AMD64 server motherboards will allow you to put enough in. The principle, though, is important (and of course, every year the prices will be half what they were and the densities twice what they were, so the principle will continue to apply). The way you should figure this out is: if I can keep my entire working set of files in RAM (something the buffer cache mechanisms can easily do on Linux or NetBSD or FreeBSD), will it speed things up so much that I don't need to buy enough more machines that the price tradeoff for a "crazy amount of memory" is actually not crazy at all. With time, for most problems, the needle points more and more often towards "get more RAM" than "get a faster disk". The obvious exception is stuff where getting things onto disk is mandatory (database servers), and where your working set is so large (right now past, say, 16G, soon much larger) that it is literally not going to be practical to get a machine with that much memory. Below that size, you owe it to yourself to do the calculation... Perry