Optimizing NEURON Simulation Environment Using Remote Memory Access with Recursive Doubling on Distributed Memory Systems

Table 3

Runtime comparison of MPI_Allgather versus RMA_Allgather in Parscalebush.

Parscalebush: tstop = 5000 ms
Cells	1000		2000		5000
Procs.	MPI	RMA	MPI	RMA	MPI	RMA
2	766.32	705.22	1973.5	1824.25	7449.68	6825.24
4	277.83	250.49	792.49	732.85	3092.29	2725.24
8	115.3	102.28	293.66	266.8	1185.74	1097.65
16	59.65	51.56	120.23	112.45	477.9	420.35
32	35.24	31.88	73.56	66.48	213.99	196.78