While cloud computing is widely adopted in many application domains, it is not yet the case for the high performance computing (HPC) domain. HPC traditionally runs on homogeneous, high-cost servers with fast networking providing for predictable performance; while bare-metal cloud offerings is promising, the underlying hardware is heterogeneous, with slower network connection, making it difficult to predict performance and hence tune applications. In this paper we consider performance modelling message passing interface (MPI)-based applications, being a major class of HPC applications. In particular, we present a queueing network performance model to account for computation and communication contentions on the underlying heterogeneous, relatively slow-interconnect architecture of the cloud bare-metal servers. The proposed model uses a non-linear problem solver to enhance the parameters acquired by profiling. We utilise our model to conduct an initial study of the performance of two benchmarks from SPECMPI-2007 suite and two NASA Parallel kernels, executing on a small cluster with varying number of multicore servers ranging from 2 to 8. Comparing the predicted and actual execution times of workloads with different number of processes shows 86% average accuracy for the benchmarks used. |
