Protein alignment algorithms with an efficient backtracking routine on multiple GPUs

BMC Bioinformatics

Table 4 Performance comparison between our solution and the Farrars' implementation

avg. length	# of sequences	CPU, Farrar	1 GPU	4 GPUs
51	4000	0,863	2,581	10,055
	8000	0,875	2,676	10,597
	12000	0,888	2,711	10,740
154	2000	1,677	2,647	10,296
	4000	1,693	2,801	10,980
	6000	1,696	2,865	11,349
257	2000	2,160	2,685	10,544
	4000	2,177	2,835	11,173
	6000	2,187	2,898	11,513
459	2000	2,824	2,709	10,679
	4000	2,834	2,837	11,274
	6000	2,831	2,846	11,370
608	800	2,842	2,358	9,224
	1200	2,871	2,493	9,720
	1600	2,885	2,469	9,770
1103	800	3,154	2,366	9,293
	1200	3,151	2,442	9,752
	1600	3,144	2,478	9,819

Mean values of GCUPS (Giga Cell Updates Per Second) for the Smith-Waterman algorithm applied on different sets of sequences. The GCUPS value is mainly used for score-only versions of the algorithm. Here, the performance of the GPU-based method is understated, because the value does not count in any additional operations (or cells) needed by the backtracking stage while the entire computational time is always considered.

ISSN: 1471-2105