Performance Evaluation of Dense Linear Algebra Kernels using Chameleon and StarPU on AWS

Due to recent advances and investments in cloud computing, public cloud providers now offer GPU-accelerated and compute-optimized Virtual Machine (VM) instances, allowing researchers to execute parallel workloads in virtual heterogeneous clusters in the cloud. This paper evaluates the performance and monetary costs of running dense linear algebra algorithms extracted from the Chameleon package implemented using StarPU on Amazon Elastic Compute Cloud (EC2) instances. We evaluated these metrics with a single powerful/costly instance with four NVIDIA GPUs (fat node) and with a cluster of five less powerful/cheaper instances with a single NVIDIA GPU in each node. Our results showed that most of the linear algebra algorithms achieved better performance and lower monetary costs on the fat node scenario even with one less GPU.