Benchmarks

Linpack from top500.org

Linpack Highly Parallel Computing Benchmark

The Linpack Highly Parallel Computing benchmark is designed to explore the scalability of parallel systems. In this benchmark, we are allowed to pick the size of the matrix (N), and hence the amount of work done by each processor. The available physical memory in the system under test provides a rough upper bound on the value for N.

Results of the Abacus:

CPUs
Nmax
Rmax (GFLOPS)
N1/2
1
5000
0.3176
-
32
14000
5.573
5000
34 14000 5.568 6000
36 14500 6.070 5500

Nmax : The matrix size (N) that gives the highest Gflops for some number of CPUs.

Rmax : The Gflops achieved for the Nmax size matrix.

N1/2 : The matrix size that achieves half the rated Rmax Gflops

Results of other Supercomputers reported in 1996:

Systems
CPUs
Nmax
Rmax (GFLOPS)
N1/2
T3E (300 MHz)
16
19,968
6.3
2208
HP Exemplar
S-Class
16
13,320
7.8
1044
Fujitsu VPP500/16,10ns
16 21,120 23.6 3360
Cray C90
(240 MHz)
16 13,312 20.7 700


PVMPOV benchmark

PVMPOV measures the performance by Ray Tracing tasks, it uses PVM to perform message passing.

click to enlarge~
Skyvase - The image we going to render

we use following options for the benchmarking:

pvmpov +iskyvase.pov +w1600 +h1200 \
+FT +vl -x -d +a0.300 -q9 \
-mv2.0 -b1000 -nw48 -nh36 -nt34 \
-L/home/carlos/pvmpov3_1e_1/povray31/include

which means we render a 1600x1200 image by the information in iskyvase.pov using 34 CPUs, and split the image into rendering blocks sized 48 x 36.

Results:

No. of CPUs Rendering Time(s)
1 413
2 209
4 107
8 51
16 31
32 21
34 20


Content Based Image Retrieval System

Results:

No of CPUs Retrieval Time(s)
1 199.56
2 124.27
4 46.73
8 26.64
16 14.32
32 9.17
33 9.04
34 9.53

*The system use DCT histogram intersaction to perform content based search, and the database contains 4707 JPEG encoded images.


Home