. Author manuscript; available in PMC: 2014 Nov 20.

Published in final edited form as: SIAM J Sci Comput. 2014;36(2):C95–C118. doi: 10.1137/120866580

Table 4.

Test problems on the Amazon EC 2 cluster and corresponding running times in seconds. When we enlarge the problem scale by a factor of 10 and increase the number of cores accordingly, the running time only increases by a factor of 50%. It shows LSRN’s good scalability. Though the CS method takes more iterations, it is faster than LSQR by saving communication cost.

Solver	N_cores	Matrix	m	n	nnz	N_iter	T_iter	T_total

`LSRN` w/ CS	4	`tnimg_4`	1024	4e6	8.4e7	106	34.03	170.4
`LSRN` w/ LSQR	4	`tnimg_4`	1024	4e6	8.4e7	84	41.14	178.6

`LSRN` w/ CS	10	`tnimg_10`	1024	1e7	2.1e8	106	50.37	193.3
`LSRN` w/ LSQR	10	`tnimg_10`	1024	1e7	2.1e8	84	68.72	211.6

`LSRN` w/ CS	20	`tnimg_20`	1024	2e7	4.2e8	106	73.73	220.9
`LSRN` w/ LSQR	20	`tnimg_20`	1024	2e7	4.2e8	84	102.3	249.0

`LSRN` w/ CS	40	`tnimg_40`	1024	4e7	8.4e8	106	102.5	255.6
`LSRN` w/ LSQR	40	`tnimg_40`	1024	4e7	8.4e8	84	137.2	290.2