Skip to main content
. 2022 Jun 22;38(15):3734–3740. doi: 10.1093/bioinformatics/btac401

Fig. 3.

Fig. 3.

Performance and memory scaling efficiency of matOptimize using the 1M-sample tree. (A) Strong multi-node scaling efficiency of matOptimize. Each node is a GCP instance e2-highcpu-32 consisting of 32 vCPUs and 32 GB memory. The number above each data point corresponds to the actual runtime in minutes: seconds format. (B) The peak memory requirement of matOptimize is small (below 10 GB) and remains roughly constant with the number of CPU threads. This allows matOptimize to exploit all available parallelism on a multicore CPU instance without being limited by the available memory. (C) In comparison, the peak memory requirement of TNT is large (>500 GB) and increases linearly as the parallelism is increased. This limits the amount of parallelism that TNT can exploit—in the example shown, TNT could exploit only up to 8 available vCPUs out of the 40 available on the memory-optimized GCP instance m1-ultramem-40 before running out of memory