Results April 6, 2006

Tpetra

Tpetra (Templated Linear Algebra Services Package) is a software developed at Sandia National Laboratories. We run this algorithm on T2000 using the shared memory standard OpenMP. The results obtained are shown in Table 1. You can find the analysis on the bottom of this page.

	MIOPS
		10	100	1000	10000	100000	1000000
	1 thread	12.244	34.5855	31.8034	32.2643	26.2351	25.1227
	2 threads	2.81033	17.9475	50.2481	62.7786	52.0425	48.2585
	4 threads	1.2174	11.4638	62.9097	115.496	101.905	92.4524
	8 threads	0.766131	7.80281	61.1201	187.29	194.294	169.859
	16 threads	0.629272	3.45052	32.5254	186.942	278.437	239.742
	32 threads	0.627869	0.240922	2.50485	24.0638	173.12	205.873


	SPEEDUP
		10	100	1000	10000	100000	1000000
	2 threads	0.23	0.52	1.58	1.95	1.98	1.92
	4 threads	0.10	0.33	1.98	3.58	3.88	3.68
	8 threads	0.06	0.23	1.92	5.80	7.41	6.76
	16 threads	0.05	0.10	1.02	5.79	10.61	9.54
	32 threads	0.05	0.01	0.08	0.75	6.60	8.19



	EFFICIENCY
		10	100	1000	10000	100000	1000000
	2	0.11	0.26	0.79	0.97	0.99	0.96
	4	0.02	0.08	0.49	0.89	0.97	0.92
	8	0.01	0.03	0.24	0.73	0.93	0.85
	16	0.00	0.01	0.06	0.36	0.66	0.60
	32	0.00	0.00	0.00	0.02	0.21	0.26


Table 1: First results obtained using Tpetra

Analysis

We obtained interesting speedups results without much effort. This is a remarkable fact that shows that the machine is behaving like a parallel computer. We can see that Tpetra package runs up to 10.6 times faster in parallel mode than in serial mode when we use 16 threads and the matrices have 100000 rows and columns.

The performance of the machine increased up to 16 threads but then deceased. This is a result that we were not expecting. We supposed that the speedup will continue growing or at least will keep in the same level. This problem was solved later (see timeline).

Maximum SWaP was obtained at 16 threads and NEQ equal to 100000 (New !). If we consider a performance of 278.437 MIOPS, a space of 2 RU (rack units), and a power consumption of 300 Watts.