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

Table 1

Communication versus computation percentage time for Parscalebush model.

Cells	Number of procs.	2	4	8	16	32
500	MPI_Allgather%	0.81	4.98	6.79	6.84	18.02
	Computation time%	99.18	95.01	93.2	93.15	81.97
1000	MPI_Allgather%	2.04	1.69	3.14	7.81	14.26
	Computation time%	97.95	98.3	96.85	92.18	85.73
2000	MPI_Allgather%	2.71	1.18	4.28	3.64	12.88
	Computation time%	97.28	98.81	95.71	96.35	87.11
4000	MPI_Allgather%	1.38	2.8	5.83	4.97	14.29
	Computation time%	98.62	97.19	94.16	95.02	85.7