Applying Operations Management Principles On Optimisation Of Scientific Computing Clusters

We apply operations management principles on production scheduling and allocation to computing clusters and their storage resources to increase throughput and reduce lead time of scientific computing jobs. In addition, we study how this approach affects the amount of energy consumed by a computing job comprised of hundreds of calculation tasks. Methodologically we use the design science approach by applying domain knowledge of operations management and efficient resource allocation on the efficient management of the computing resources. Using a test cluster we collected data on CPU and memory utilisation along with energy consumption on different ways of allocating the jobs. We challenge the traditional one job per one processor core method of scheduling scientific clusters with parallel processing and bottleneck management. We observed that by increasing the utilisation rate of the cluster memory increases throughput and decreases energy consumption. We studied also scheduling methods running multiple tasks per CPU core and scheduling based on the amount of free memory available. The test results showed that, at best these methods both decreased energy consumption down to 45% and increased throughput up to 100% compared to the standard practices used in scientific computing. The results are being further tested to eventually support LHC computing of CERN.