Results

We timed each of the benchmarks at problem sizes ranging from $21$ to $217$ (131072). Table 2 gives timings for selected problem sizes, averaged over five runs and rounded to two significant figures. We'll analyze the results for the line-fit benchmark in depth, and then briefly discuss the results for the selection and sparse matrix-vector multiplication benchmarks.

Table 2: Running times in milliseconds for three NESL benchmarks using different intermediate language implementations on a 120 MHz 486 PC and a SPARCstation 5/85.

Problem	120 MHz 486				SPARCstation 5/85
size	JDK	JIT	Native	Vinterp	JDK	Native	Vinterp
	Line-Fit
16	9.1	2.6	5.8	0.73	9.5	7.0	0.62
128	17	4.3	7.1	1.3	16	8.0	0.85
1024	84	14	16	6.6	70	15	5.4
8192	580	99	90	50	500	71	40
65536	4800	800	570	460	4000	480	310
	Selection
16	14	5.4	8.8	1.3	15	11	1.3
128	25	8.2	14	2.1	25	16	2.0
1024	74	18	25	4.7	70	29	4.3
8192	350	52	48	19	320	52	21
65536	2600	280	200	130	2300	190	160
	Sparse matrix-vector multiplication
16	2.0	0.62	1.2	0.17	2.0	1.4	0.14
128	3.9	1.1	1.4	0.28	4.1	1.8	0.20
1024	18	3.9	2.9	1.4	17	2.9	0.95
8192	135	25	15	12	120	13	8.2
65536	1100	200	120	110	940	97	68

 

• Line-fit Benchmark
• Selection Benchmark
• Sparse matrix-vector multiplication benchmark