[gmx-users] MD simulations on a remote workstation

[Justin Lemkul]

On 4/2/14, 6:29 PM, Amjad Farooq wrote:
> Justin,
>
> I am trying to set up a single system image using PC as the master node and
> two workstations as slave nodes (essentially a supercompuer; more nodes
> will be added later), so that I can have everything together, instead of
> running my simulations individually on each workstation. But, all nodes
> will share the master hard disk. That is where all I/O will take place but
> processing will be split between different nodes.
>

OK, that's a very different situation than what you originally described, which 
basically was "computer 1 needs to talk to computer 2."  It's much more 
complicated than that :)

> The problem:
> The slaves nodes are being networked with the master node using 1-Gb
> ethernet cards.
>
> The Question:
> Is it worth upgrading to 10-Gb ethernet cards? How much improvement can I
> expect?
>

No.  If you want decent performance out of a cluster of any size, Infiniband is 
the way to go.  Gb ethernet is generally not up to the task here.

-Justin

-- 
==================================================

Justin A. Lemkul, Ph.D.
Ruth L. Kirschstein NRSA Postdoctoral Fellow

Department of Pharmaceutical Sciences
School of Pharmacy
Health Sciences Facility II, Room 601
University of Maryland, Baltimore
20 Penn St.
Baltimore, MD 21201

jalemkul at outerbanks.umaryland.edu | (410) 706-7441
http://mackerell.umaryland.edu/~jalemkul

==================================================