5 ms·
I bought such machines in an academic high performance computing center a few years ago, with similar budget. Here are a few thoughts: - You definetly want to
by ktpsns 4y ago
I bought such machines in an academic high performance computing center a few years ago, with similar budget. Here are a few thoughts:
- You definetly want to do rack mount. It is more extensible and future proof.
- You will get a server grade mainboard which features industry standards such as IPMI remote managament, redundant power supplies or SAS (which allows for extending hard drives, for instance).
- You should probably put as many RAM into this machine in this machine as you can afford, since this will be a limiting factor, especially if you do "all at the same time"
- With your budget, you can consider buying a few less powerful compute nodes (think of 5k-10k per node) which can be more suitable if you divide the computer to multiple users.
[edit:formatting]
- uniqueuid 4y agoThanks, this is actually for an academic setting, so a good match. I like the "as much ram as you can get" recommendation, and will probably aim for ~512GB which is a cost-effective maximum. The thought of splitting it into multiple machines has occurred to me, but I'm discouraged by anything that creates distributed problems. There is MPI, but it's cumbersome and performance depends highly on the network (I guess I'll have 10G). Batch management such as slurm works, but also has quite some overhead. Some things are near impossible to distribute, i.e. some algorithms with large matrices. That's why a single e.g. 64 core machine seems more attractive. But I'll take to heart all the recommendations for rack-mount - seems this is a way to prevent a lot of headaches.