Accelerating Linpack Performance with Mixed Precision Algorithm on CPU+GPGPU Heterogeneous Cluster

In this paper, the mixed precision algorithm to solve the linear system of equations and the implementation of HPL package are introduced. We use this mixed precision algorithm to improve HPL package on CPU+GPGPU heterogeneous clusters, which is named for GHPL, and give the implementation mechanisms in detail. The experimental results are measured on the platforms of multi-core CPUs and CPU+GPGPU heterogeneous clusters. From the experimental results, we can find out that our GHPL program has good scalability on all the experimental environments and can sustain more than 1.7Teraflops both on the cluster with 16 nodes containing 32 NVIDIA Tesla C1060 GPUs and on the cluster with 8 nodes containing 32 ATI GeForce GTX 295 GPUs, while the average speedup of it with respect to HPL is 3.06 and 2.40 respectively.