- The assumption of an ideal linear scale-up is wrong because the cluster is not homogeneous. The quick way around it would be to assume that the CPU clocking is proportional to the nodes' effective speed. As shown below, this is clearly wrong.
- The timings reported in the previous tests were those reported by the program fairly soon after the start-up phase of the simulation. For the tests reported here the simulation time was extended to 10000 steps (20 psec) and the time reported is the minimum observed from the whole run.
- In the previous tests, the network connection of node number 9 (aspera) involved an additional hub. This significantly changed the effective bandwidth of this node's connection to the rest of the cluster. For this test, aspera was directly connected to the cluster's main switch.
Estimation of an 'ideal' scale-up : Server is a PIV at 2.6 GHz, the newest nodes are Celerons at 2.4 GHz, and the oldest nodes are PIIIs at 733 MHz. Based on that, it would be expected that the server and celerons would have comparable performance. But, they don't : running NAMD stand-alone on these machines gave the following results
Machine | CPU | Clocking | Days per nanosecond of simulation | Relative speed |
Server | PIV | 2.6 GHz | 11.7721 | |
New nodes | Celerons | 2.4 GHz | 17.2784 | |
Old nodes | PIII | 733 MHz |