Supplementary Figure 10: Scalability of the parallel computing of THUNDER.

Three dirty datasets were used to measure the computing time of a) 3D refinement and b) 3D classification of the CNG dataset, c) 3D refinement and d) 3D classification of the proteasome dataset, e) 3D refinement and f) 3D classification of the β-galactosidase dataset. Each job was submitted to 8, 16, 32 and 64 computing nodes, respectively. The computing time versus the number of computing nodes shows nearly ideal linear trend (red curves). Each computing node has two E5–2680v3 CPU and 128GB DDR4 RAM. The current version of THUNDER needs to load the entire dataset into RAM to reduce the pressure on the storage I/O, which requires that the sum of the available RAM spaces from all computing nodes is larger than the size of the dataset. Due this limitation, the β-galactosidase job on 8 computing nodes failed, and hence was not tested. The data loading strategy will be changed in the next version of THUNDER to avoid this problem.