Version 1.9 Build 1556

News	FAQ
Search	Home

Next: Results Up: NOTE 232 - Parallelization Test Results Previous: Introduction

Tests

The two tests are described as follows:

Parallelized deconvolution over spectral channels using the AIPS++function bigimagertest to do a Clark CLEAN on a subset of the M33 data set supplied by Dave Westpfahl. specifically

niters := 50000; nthread := 5; processors := nthread-1;
print spaste('2kx2kx4 - ', niters, ' iterations - ', processors, 
             ' worker threads');
bigimagertest(size=2048, cleanniter=niters, imagertask='pimager', nchan=4,
              numthreads=nthread, forcenew=T);

with nthread being varied 2,3, and 5.

A simple parallel iteration through the complete M33 uv-data set using the AIPS++ function pimager.tryparread.
```
pim := pimager ("M33-A.MS", numprocs=i)
pim.tryparread("M33-A.MS", numloops=128)
```
Where numprocs was varied from 2,3,5, 9, 17, 33, 65, and 129. The run with numprocs set to 129 was only done on the Origin.

Both sets tests were run with and without collecting Pablo statistics. Pablo is the parallel I/O profiling library developed at UIUC. For more details about Pablo visit their web site at www-pablo.cs.uiuc.edu.

Next: Results Up: NOTE 232 - Parallelization Test Results Previous: Introduction Contents