Due to their high parallelism graphics processing units (GPUs) and GPU-based clusters have gained popularity in high-performance computing. However, data transfer in GPU-based clusters remains a challenging problem, due to the disjoint memory of GPU and host. . New technologies, such as GPUDirect RDMA, help to improve data transfer among multiple GPUs and enable new data transfer and communication models for distributed GPUs.  In my talk I will introduce and discuss the assets and drawbacks of three communication alternatives for distributed GPUs. GPI2 is a communication library for one-sided communication. It was designed for heterogeneous memory structures and therefore allows an easy integration of GPUs.  The communication in GPI2 is controlled from the host, which requires frequent context switches between host and GPU. Therefore, GPI on the GPU and GGAS deal with GPU controlled communication.  GPI on the GPU uses a remote put/get model, similar to GPI2, but the communication requests are issued from the GPU.  GGAS (Global GPU Address Spaces) creates a global Address space over distributed GPUs, which allow communication with simple remote load/store instructions from the GPU.

Bio: Lena Oden is a PhD Student at the University of Heidelberg in Germany since 2011. There, she works in the department of computer architecture. She also works at the Fraunhofer Institute for Industrial Mathematics, where she has got a full time fellowship.  Before, she studied electrical engineering at the RWTH Aachen University. Her main research topic is communication and data transfer in heterogeneous systems.