We'll introduce GPU-accelerated unsupervised reinforcement and auxiliary learning (UNREAL) algorithm. Recent state-of-the-art deep reinforcement learning algorithms, such as A3C and UNREAL, are designed to train on a single device with only CPUs. Using GPU acceleration for these algorithms results in low GPU utilization, which means the full performance of the GPU is not reached. Motivated by the architecture changes made by the GA3C algorithm, which gave A3C better GPU acceleration, together with the high learning efficiency of the UNREAL algorithm, we extend GA3C with the auxiliary tasks from UNREAL to create GUNREAL. We show that our GUNREAL system finished training faster than UNREAL and reached higher scores than GA3C.  Back

Recent supercomputers deploy not only many-core accelerators such as GPU but also Non-Volatile Memory (NVM) such as flash memory as an external memory, in order to handle large-scale data processing for a wide range of applications. However, how to construct NVM as local disks at a low cost with large volume for heterogeneous supercomputers is not clear. In order to clarify I/O characteristics between GPU and NVM, we comparatively investigate I/O strategies on GPU and multiple mini SATA SSDs. Our preliminary results exhibit 3.06GB/s of throughput from 8 mini SATA SSDs to GPU by using RAID0 with appropriate stripe size.  Back

Fast processing for extremely large-scale graph is becoming increasingly important. The applicability of acceleration using GPUs including optimization techniques, of the GIM-V graph processing algorithm, which is based on MapReduce, is an open problem. We implemented a multi-GPU-based GIM-V application with load balance optimization between GPU devices. Our experiments on the TSUBAME2.0 supercomputer using 256 nodes (6144 hyper-threaded CPU cores, 768 GPUs) showed that our GPU-based implementation performed 87.04 ME/s on 2^30 (1.07 billion) vertices and 2^34 (17.2 billion) edges, and 1.52 times faster than the CPU-based naive implementation with 2^29 vertices and 2^33 edges.  Back

APGAS programming model abstracts deep memory hierarchy such as distributed memory and GPU device memory by a global view of data and asynchronous operations on massively parallel computing environments. However, how much GPUs accelerate applications using APGAS model remains unclear. We give a comparative performance analysis of APGAS model in X10 on GPUs with MPI on multi-core CPUs using lattice QCD. Our experimental results on TSUBAME2.5 show that our X10 implementation on 32 GPUs exhibits 11.0x speedup over MPI on multi-core CPUs.  Back