BAC on HPC Systems Sample Clauses

BAC on HPC Systems. BAC computational method is based on classical MD simulations where the unit of parallelism involved is a few hundred cores, but runs hundreds to thousands of these at a time in order to perform “high throughput” calculations in support of drug discovery and personalised medicine (for advising on drug treatment for patients based on their genomic profile). In order to guarantee reproducibility of predictions, ensembles of replica MD calculations are performed for each method. Combining this approach with the necessity to screen multiple compounds, makes BAC a perfect example of replica computing pattern application, where a large number of copies of simulations are needed to produce statistically robust results. All replicas simulations, which may be parallel runs themselves, execute independently of each other, allowing BAC to reach excellent scalability also at high number of cores. It is a computational workflow which is data intensive and data driven in the sense of needing a large amount of data to initialise and produce tens or more of terabytes of output per run. In the case of replica compute applications such as BAC, weak scaling analysis provide useful information to demonstrate the ability to solve larger systems as the resources available increase. Figure 6 shows the result of a weak scaling analysis, performed by UCL on the Blue Waters (NCSA) supercomputer. The study consisted in the screening of sixteen drug candidates concurrently using thousands of multi-stage pipelines. The results demonstrated the almost perfect scaling of the code up to 33,280 cores under the conditions tested. The overhead measured, consisted in the adaptive sampling of the method used, and small overhead from the HTBAC framework. The latter depends on the number of protocols generated and increase linearly, however, as showed in the results presented, is always negligible compared to the total time of execution. Similar scaling has been demonstrated on other platforms such as Titan (ORNL) and for different protocols.