Note, this is based on the document prepared last year, but with small updates made to reflect dropped prices on systems. Currently, quad-core intel is the "sweet spot" for inexpensive cluster nodes, at a cost of approximately $1200 per node, or ~<$300 per cpu core.
Context
Given the abundance of inexpensive "desktop grade" CPUs it is absurdly simple to throw togother a "deskside beowulf cluster" for prices which typically in the past would only get you a 2-CPU workstation from a major vendor.
Clearly, a beowulf environment is ONLY suitable for workloads which don't mind being split across multiple CPUs. Such issues are NOT discussed here but basic "good candidate" tasks might include,
- when you run batches of large jobs in parallel, each independant from the other
- when you run a single large job, using MPI parallel architecture
- when you run a single large job, using PVM parallel architecture (older than MPI but still used it seems)
The prices below are NOT to be taken as an "absolute written in stone" but rather as an overview to the rough price range and feasability of a concept such as this.
Also, be warned that a bunch of .. 5 desktop pc Tower cases ... takes up a fair bit of space, so the term "deskside" is somewhat generous. Additionally, they will throw out a not-insignificant amount of heat and noise, plus of course you need a spare outlet, network drop, etc. However, if you have a "back room" with power and a LAN drop, then this might be an option. (Such a cheapo cluster will not be housed in the POLan server room! :-)
It is also taken as a given that a "standard linux cluster distro" (something like "Rocks Cluster - http://www.rocksclusters.org/wordpress/ ) would be used on this gear, to simplify deployment and management issues, which are NOT insignificant in planning a project like this.
Final footnote: this setup has NO redundancy at all. (No mirrored HDDs, etc). Such things are possible with a bit more $$ (not a lot).
Rough Pricing for a 4-box / 16 CPU cluster:
Rough "per node" price:
- $1150.00 pre-tax = Intel QuadCore 2.4ghz / 4gigs RAM / 320gig ATA disk / Dual-GigEther in a "PC Workstation Case"
Thus:...for ~$5000 we could buy,
Four Node Cluster: 3 compute nodes plus one head node @ $1150 each = $4600 8-port gig-ether DLINK switch = $80, add $50 for cables etc.
Grand Total = $4700 with cables/switch etc.
NOTE - ALL PRICES ARE "PRE-TAX". these prices are pulled from local reseller, "Robotnik", so all gear will have 1 year dealer warranty, possibly longer on some parts (such as HDDs which have 5years from Mfr.)
This provides 12 CPU cores @ 2400mhz, 12 gigs RAM / 1 gigs per cpu core, head node with ~300 gigs local storage (NOT mirrored // no redundancy)
If ~300gigs of local disk isn't enough, then more can be added .. $130 per 500 gigs up to a max of 4 drives approx.
if 1 gigs RAM per CPU core isn't enough, more $ can be spent there (although ram price spikes, it is $450 extra per node to upgrade from 4->8 gigs of ram per node)