Results

Using the methodology above, we can run the stream-wall test (part of the BCS package) from the command prompt. To determine STIME we run:

$ time /usr/bps/stream-wall

real 0m1.616s 

user 0m1.490s 

sys 0m0.120s 

To determine CTIME1 and CTIME2 the following is run:

$ time /usr/bps/bin/stream-wall& time /usr/bps/bin/stream-wall

real 0m2.839s

user 0m2.660s

sys 0m0.170s

real 0m2.844s

user 0m5.360s

sys 0m0.330s

Therefore, STIME is 1.6s, CTIME1 is 2.8s and CTIME2 is 2.8s. Using the formula shown above, we get a speedup of 1.14! This result indicates that if we run two programs that have a similar memory access patterns to stream, we will see very little performance gain on a dual node when running two programs.

In order to test more ``real world'' programs the NAS parallel test suite was employed. The test script (see below) was run for both the GNU compilers and the Intel compilers. The speed-up for each test is listed below. An average speedup for the whole suite is also given.

Test	GNU2.96	Intel 6.0
CG	1.1	1.05
BT	1.45	1.34
EP	1.99	2
FT	1.55	1.55
LU	1.4	1.17
IS	1.78	1.7
SP	1.56	1.2
MG	1.34	1.04
Average	1.52	1.38

Speed-up for concurrent NAS programs on a dual SMP node