On 27 Feb 2005, at 6:37 pm, Elia Stupka wrote: > I've been hoping to be able to access large amounts of memory on > affordable servers for a while, in reality, though, the 4GB OS limit > has hardly been the issue since unfortunately the cost of memory is > still very high and hardware vendors seldom offer more than 4GBs per > processor. > > The Sun Opterons are the only ones (among the mainstream vendors) that > offered us a 4-way option with 32 GBs. The Apple G5s are still limited > to 8GB (4 per processor, probably when Tiger will be truly released > they will finally offer more memory slots?), IBM Opterons offer 16GB > (still only 4GB per processor), the blade versions are always limited > in memory, etc... then you are left with the usual suspects (Power5s, > etc.) who have been dealing with more memory for a long time, but at a > nasty price... > > ...as long as it costs more to equip hardware with good amounts of > memory than it costs to buy the hardware, the refinement of 64-bit OS > for access to large amounts of memory can't take off properly, can it? There's always the option of a mixed architecture cluster; buy lots of cheap 32-bit boxes for the vast majority of compute tasks which can run in a 4GB address space, and then spend the money saved to buy one or two large and much more expensive machines to handle the very large memory tasks. This is the approach we have taken; we have about 500 dual processor Pentium IV machines, with only 2GB per processor. Almost all our workflow goes through those machines, but the LSF cluster also contains a couple of huge memory machines for the other stuff - two SGI Altix 350 machines, with 192 GB of RAM each, one with 4 CPUs, the other with 16. Obviously, 192GB of memory is *seriously* expensive, and not what a lot of groups would be able to afford, but smaller Altix boxes are really quite affordable [ as Itanic machines go ], and if you want a 64GB machine with, say, 4 CPUs, they're only mildly expensive :-) The main issue with several types of machine in the same cluster is of course that the users then have to specify their requirements to the DRM so that it schedules to the right kind of machine. And I don't know about you guys, but ours are notoriously bad at estimating what their jobs' requirements actually are. Tim -- Dr Tim Cutts Informatics Systems Group, Wellcome Trust Sanger Institute GPG: 1024D/E3134233 FE3D 6C73 BBD6 726A A3F5 860B 3CDD 3F56 E313 4233