Acquire and deploy hardware
Here is our recommendation of how to build a computer cluster that will work well for molecular docking and cheminformatics. This document can be used whether you already own hardware, or whether you are planning to buy new. All recommendations are the best we know as of Feb 2014. Things do change, but this advice should be ok through 2015.
About buying CPU and disk
We are currently buying CPU from Silicon Mechanics and Dell and disk from Silicon Mechanics and HP. We recommend buying two different kinds of machines in a modular fashion: head nodes to which disk enclosures may be attached and cpu nodes which contain large numbers of cores. We are currently buying enclosures holding 12 SAS disks of 4TB each for 48 TB raw for around $8000 or about 6 raw GB per dollar. Formatted RAID6 this works out to 36TB or 4.5 formatted GB per dollar. We like the HP P822 high performance RAID controller. Compare this to what we were paying just a year ago in spring 2013: 25 TB for $10,000 or 2.5 TB per dollar unformatted. An amazing development in the last 12 months.
For CPU, we like the C6145 from Dell. For around $20,000 you get 2 machines in a 2U form each with 64 cores and 256 GB memory and a pair of RAID1 formatted disks each. A single 42U rack could hold 2560 cores and still have room for a switch. Of course, this would cost you $400,000, pull 28 kW and need 5 T of cooling. Amazing density at commodity prices.
About setting up a network
$250 for a Managed 24 port GigE switch with VPN support. About $250. Will do for most circumstances. Allows you to run both private and public networks.
Minimum requirements for legacy equipment
At the very least, you really need 24 cores at 1GB/core and 10TB of RAID10 formatted disk space to get started. There have been substantial improvements in disk price/performance and CPU-core/dollar performance in the past 24 months that we really recommend getting new equipment if you can. If you do not have a 24 core / 48 GB / 4 TB disk system to run the hypervisor and core services, then you will need to reserve 3 older 2, 4 or 8 core machines to run central services. You can use two cheap switches to run private and public networks separately.
If you have more money
Our recommended CPU machine costs $20K (128 cores). Our recommended disk machine costs $20K (96 TB raw). We recommend just buying these two machines in a proportion that makes sense for your research. A single switch will handle an entire rack. If you have a lot more money, ask us. Don't forget about cooling, UPS and power.
Return to So you want to set up a lab.