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

Table 2

Runtime comparison of MPI_Allgather versus RMA_Allgather in Parbulbnet.

Parbulbnet: cells = 2525
Tstop	1000 ms		2000 ms		5000 ms
Procs.	MPI	RMA	MPI	RMA	MPI	RMA
2	117.94	104.25	247.78	220.48	635.64	580.52
4	52.97	44.86	113.49	101.45	283.26	269.73
8	19.31	15.33	37.33	34.54	93.47	86.25
16	8.82	7.61	17.26	15.03	41.45	36.92
32	6.06	5.87	10.64	9.16	26.32	23.38