[Bioclusters] small simple test lab recommendations

[Chris Dagdigian]

Hi Paul,

Your proposed setup sounds fine. Starting from what you propose the next 
step people usually take when they grow their cluster is to split the 
functionality of admin/master/filserver across multiple machines. No 
need for you to do so, but keep it in mind as the first thing you'd 
probably be required to do if you were to put your cluster into 
production or grow it in size.

You also should think about network topology. Most "compute farm" 
configurations have the compute nodes sitting on a private network. The 
admin/master/nfs nodes have at least 2 network interfaces and are 
connected to both the private compute network and the 'public' network 
that leads to the outside world. Most rackmount systems have at least 2 
network devices so this is usually trivial to setup, but if you are 
starting with small formfactor desktops you may want to figure out if it 
can support multiple NICs.

If gaining practical knowledge is your goal then dealing with 
admin/master/nfs nodes that are multi-homed to several networks is 
something you will want to be able to handle.

When learning clustering the hardware is not as important as the 
software you choose to use and the tools/tricks/technology you use to 
manage the cluster as easily as possible.

Things you'll need to think about:

o will you build your own cluster vs. starting with cluster kits like 
biobrew, ROCKS, warefulf, etc.

o network services (clusters make heavy use of DNS, DHCP, PXE and TFTP 
among others). The #1 admin tool you will need to understand is the 
method in which nodes are reimaged or OS's are loaded automatically onto 
"bare" hardware. Unattended updates, installs and upgrades are something 
you will want to have for a cluster of any size.

o what scheduler/batch software will distribute jobs on the cluster? I 
personally hate cluster kits that remove this choice from the user (good 
argument for ROCKS and its various 'rolls'...) One of the most important 
  choices a person makes about a new cluster is the DRM software suite. 
Grid Engine vs Torque vs PBSPro vs LSF vs XGrid.  { If you want a 
suggestion, go with Grid Engine as it's the best of the freely available 
products)

o What other software tools will you use to make your life easier? 
SystemImager, MRTG, BigBrother, Nagios, SAR, NTOP, Ganglia. Look for 
tools that help with managing, monitoring and reporting.


BioBrew Linx may be a good starting point as it is ROCKS based with Grid 
Engine under the hood and lots of informatics apps pre installed.

If you want to build your own cluster on top of a Linux OS then I'd 
recommend the following software:

  o Any Linux distro you prefer; Suse 9.1 if you want a suggestion
  o Grid Engine 6 for job scheduling
  o SystemImager for OS barebones OS installs & incremental updates
  o BigBrother and Ganglia for cluster monitoring

-Chris





Paul Weiss wrote:

> Hello,
> 
> I'm trying to build a small simple test lab so we can practice
> clustering and administating. My thoughts were a 4 node on pc type
> machines. Maybe 4 Dell small form factor Optiplex 1Ghz 512mb ram 10gb
> hard drive. Use a 5th machine for admin/master/nfs server.
> 
> Any ideas.
> 
> Thank you,
> 
> Paul

-- 
Chris Dagdigian, <dag at sonsorol.org>
BioTeam  - Independent life science IT & informatics consulting
Office: 617-665-6088, Mobile: 617-877-5498, Fax: 425-699-0193
PGP KeyID: 83D4310E iChat/AIM: bioteamdag  Web: http://bioteam.net