Ideal architecture for an image processing cluster

Mark Lucas mlucas at imagelinks.com
Sat Apr 14 12:47:18 PDT 2001


We use an 14 node cluster to serial process remote sensing imagery at 
ImageLinks.  Essentially, we use PVM and segment the imagery for 
intensive CPU processing on the slave nodes.  The jobs are serially 
staged through the master node as we have noticed severe performance 
hits when multiple jobs are kicked off simulaneously.  One would 
expect some performance hit due to the overhead of managing 
additional processes, but not the types of degradation that we have 
seen.  For resampling runs we can perform the work through a 
scheduler, but would ideally want to establish an architecture that 
would support adhoc work from users workstations.

When the production personnel are correcting imagery the same code is 
run as image tiles are processed through the math and displayed to 
the screen.  My ideal architecture would allow each workstation to be 
a master controlling the cluster slaves  (data typically resides on 
the workstation).   I

Is this practical to pursue?

Is the problem a function of PVM or something more basic?

BTW, we are seeing close to linear scaling on the large production 
runs with the scheduling mechanism.  Most of the time the cluster 
sits idle while the interactive work is being performed.

Thanks in advance.

Mark
-- 
**********************
Mark R Lucas
Chief Technical Officer
ImageLinks Inc.
4450 W Eau Gallie Blvd
Suite 164
Melbourne Fl 32934

321 253 0011 (work)
321 253 5559 (fax)

mlucas at imagelinks.com
**********************




More information about the Beowulf mailing list