Evaluation

We also removed the graphic part of the CFD code and did performance tests on the computational part only. For this we also changed a 2 dimensional Java object distributed array into a 3 dimensional double distributed array and stored fields of the Java object into the collapsed 3rd dimension of double array. This change was to improve performance, because if we are using Java object to communicate between processors, there is an object serialization overhead which is not required for primitive data types. Also we are using HPC implementation of underlying communication to run the code on an SP.

Figure 6.12: CFD with size of $256^2$.

Table 6.5: CFD speedup of HPJava benchmarks as compared with 1 processor HPJava.

CFD
Processors	2	3	4	6	9
97 by 25	1.75	2.33	2.73	4.06	4.91
Processors	2	4	8	16
$128^2$	1.84	3.34	5.43	7.96
$256^2$	2.01	3.90	7.23	12.75

Figure 6.12 shows result of two different versions (HPJava, sequential Java) of CFD with size of 256 by 256. Speedup of HPJava is also summarized in Table 6.5. As we mentioned earlier, we are measuring different size of problems on different number of processors and using the result of the single-processor HPJava version for the reference value. As we are expected speed up of CFD is more scalable then partial differential equation examples which described in the previous section. As we can see on the table we are getting up to 12.75 times speedup (4.68 times speed up on Laplace equation) using 16 processors with problem size of $256^2$.

Table 6.6: CFD performance. All speeds in MFLOPS.

97 x 25
Processors	1	2	3	4	6	9
HPJava	49.09	85.74	114.31	133.79	199.58	241.08
Java	72.03
$128^2$
Processors	1	2	4	8	16
HPJava	51.37	94.75	171.40	278.84	408.86
Java	72.42
$256^2$
Processors	1	2	4	8	16
HPJava	50.04	100.73	195.30	361.72	638.24
Java	72.56

Complete performance results from the CFD code are given in Table 6.6.