Folding accross racks. How? Possible?

Here you can find out about our Folding Team! Our goal: to understand protein folding, protein aggregation, and related diseases
Post Reply
TraceHagan
Legit Fanatic
Legit Fanatic
Posts: 164
Joined: Tue Jul 07, 2009 1:28 pm

Folding accross racks. How? Possible?

Post by TraceHagan »

I'm going to be coming into some older p4 based servers (i.e. single core HT, dual socket/processor machines).
Is there an easy way to boot these as a single system/OS and run a single F@H instance across them?


Thanks,

Trace
Image
User avatar
kenc51
Legit Extremist
Legit Extremist
Posts: 5167
Joined: Thu Jun 23, 2005 1:56 pm
Location: Dublin, Republic of Ireland
Contact:

Re: Folding accross racks. How? Possible?

Post by kenc51 »

spitter
Legit Extremist
Legit Extremist
Posts: 252
Joined: Mon Jul 13, 2009 1:33 am

Re: Folding accross racks. How? Possible?

Post by spitter »

http://foldingforum.org/viewtopic.php?f ... er#p183739
WUs are large blocks of computation which only need to be synchronized rarely. They can be assigned as uniprocessor assignments to each of the nodes in a cluster as independent machines or to individual donor's machines. Each WU contains an entire protein. If you were assigned half of a protein (as in the case of -smp on a duo) you cannot process more than one step of that half-protein without knowing what changed in the other half of the protein during the same timestep. Breaking up a protein into fractional parts as in SMP assignments is severely limited by the need to constantly interchange a lot of data but it does allow for cooperative processing at the expense of a huge amount of data interchange.

Clusters work well as a single system when portions of a single process consist of relatively large computation blocks that need relatively rare synchronization. When working on a fraction of a single protein from a single SMP assignment, FAH has relatively small computation blocks that need constant synchronization. In a cluster, network latency is a huge impediment to getting anything done on SMP assignments. Latency in the interchange of data between cores is manageable but shows up when something like the GPU client delays a single thread.

For that reason, there is no FAH client that is designed to run on clusters. FAH, itself is a massive cluster, and each of us represents one or a few nodes, but you can't subdivide the work designed for a single node.
Post Reply