Index of /download/dist/upc-examples/psearch

[ICO]NameLast modifiedSize

[PARENTDIR]Parent Directory  -
[   ]uts.pdf2022-10-28 13:54 63K
[   ]sha1.upc2022-10-28 13:54 6.3K
[TXT]sha1.h2022-10-28 13:54 538
[   ]psearch.upc2022-10-28 13:54 14K
[   ]harness.conf2022-10-28 13:54 678
[TXT]README2022-10-28 13:54 1.6K
[   ]Makefile2022-10-28 13:54 329

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

This is the unbalanced tree search benchmark in upc.

To compile:
  <upc-compiler> -o psearch psearch.upc sha1.c

It may be necessary to change the extension of psearch.upc to .c,
depending on the compiler.

To execute:
  psearch <way to specify number of threads> 

This will search 3200 trees using the specified number of threads and
should report a total size of 50,045 nodes if everything is working
correctly.  These trees are relatively well balanced, and the test
is primarily to insure that everything is working correctly.  If you 
want to generate a larger amount of work, increase the number of trees:

  psearch <numthreads> -n 10000

will search 10000 trees.


The benchmark goal is to get good performance on highly unbalanced 
trees.  The canonical benchmark run for this case is:

  psearch <numthreads> -m 8 -q 0.124999

This will search 3200 unbalanced trees and should report a total
size of 5,529,089 nodes.  This is the setting that I'm using to
test various machines.  

It may be worthwhile to vary the steal chunk size, since that is the
main way to deal with variations in the ratio of processor speed to
communication speed.  By default it is set to 20, but try it at a very
fine grain setting, say 4, and at a course grain setting, say 100:

  psearch <numthreads> -m 8 -q 0.124999 -c 4
  psearch <numthreads> -m 8 -q 0.124999 -c 100

When using a large number of processors it is likely that the
chunksize should go up slightly for maximum performance.

Additional description of the problem can be found in the file
uts.pdf

Jan Prins
[email protected]
May 2003