Table 1.
Strong scaling performance of NEURON (seconds) on BG/P in Virtual Node mode for simulation runs lasting 200 ms.
| Cores | Cells | Conn. | MPI_Allgather | MPI_ISend | DCMF_Multicast | Record-Replay | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| Comp | Run | Comp | Run | Comp | Run | Comp | Run | |||
| 8 K | 2 M | 1 K | 6.76 | 7.84 | 22.3 | 24.9 | 7.30 | 7.74 | 6.30 | 6.99 |
| 16 K | 2 M | 1 k | 4.01 | 5.44 | 11.5 | 13.5 | 3.67 | 4.00 | 3.17 | 3.62 |
| 32 K | 2 M | 1 k | 3.00 | 4.69 | 5.90 | 7.40 | 1.84 | 2.07 | 1.58 | 1.91 |
| 64 K | 2 M | 1 k | 2.98 | 5.05 | 2.87 | 4.20 | 0.957 | 1.25 | 0.808 | 1.03 |
| 128 K | 2 M | 1 k | - | - | - | - | 0.497 | 0.64 | 0.417 | 0.608 |
| 8 K | 1/4 M | 10 k | 5.23 | 5.64 | 18.5 | 21.5 | 6.50 | 6.90 | 5.81 | 6.58 |
| 16 K | 1/4 M | 10 k | 3.16 | 3.78 | 11.1 | 14.6 | 3.76 | 4.18 | 3.35 | 4.07 |
| 32 K | 1/4 M | 10 k | 2.04 | 3.11 | 6.37 | 9.37 | 2.04 | 2.39 | 1.79 | 2.64 |
| 64 K | 1/4 M | 10 k | 1.58 | 3.47 | 6.19 | 6.99 | 1.36 | 1.42 | 0.948 | 1.77 |
| 128 K | 1/4 M | 10 k | - | - | - | - | 0.573 | 0.799 | 0.494 | 1.39 |
Comp and Run columns are in seconds. The Conn column refers to the average number of connections per cell. The MPI_Allgather method used a buffer size of 10 compressed spikes. The MPI_Isend and DCMF_Multicast methods utilize the two-phase and two sub-interval variants. The Record-Replay method utilizes the one sub-interval variant.