November 2015 Results

Here are the results for the fourth order (operators.fv4.c) HPGMG-FV implementation (v0.3). Each machine was allowed to use any amount of memory per node, but three problem sizes were benchmarked: h(max), 2h(max/8), and 4h(max/64). Unfortuantely, due to scheduling and allocation limitations, some machines only evaluated at limited concurrency (%'s). Currently, these machines are ranked based on peak DOF/s (almost invariably h). Nevertheless, we are considering alternate metrics such as the sum, mean, geometric mean, and median. Feedback from the community is welcome.

	System		HPGMG DOF/s			Parallelization			DOF per	Top500
Rank	Name	Site	h	2h	4h	MPI	OMP	ACC	Process	Rank
1	Mira	ALCF	5.00e11	3.13e11	1.07e11	49152	64		36M	5
			3.95e11	2.86e11	1.07e11	49152	64		36M
2	Edison	NERSC	2.96e11	2.46e11	1.27e11	10648	12		128M	34
3	Titan (CPU-only)	OLCF	1.61e11	8.25e10	2.37e10	36864	8		48M	2
4	Hopper	NERSC	7.26e10	5.45e10	2.74e10	21952	6		16M	62
5	SuperMUC (22%)	LRZ	7.25e10	5.25e10	2.80e10	4096	8		54M	20
6	Hazel Hen (7%)	HLRS	1.82e10	8.73e09	2.02e09	1024	12		16M	-
7	SX-ACE (vector)	HLRS	3.24e09	1.77e09	7.51e08	256	1		32M	-
8	Babbage (MIC-only)	NERSC	7.62e08	3.16e08	9.93e07	256	45		8M	-