Network-Attached Accelerators: Host-Independent Accelerators for Future HPC Systems
Authors: Sarah Marie Neuwirth (University of Heidelberg), Dirk Frey (University of Heidelberg), Ulrich Bruening (University of Heidelberg)
Abstract: The emergence of accelerator technology in current supercomputing systems is changing the landscape of supercomputing architectures. Accelerators like GPGPUs and coprocessors are optimized for parallel computation while being more energy efficient. Today's accelerators come with some limitations. They require a local host CPU to configure and operate them. This limits the number of accelerators per host. Another problem is the unbalanced communication between distributed accelerators.
Network-attached accelerators are an architectural approach for scaling the number of accelerators and host CPUs independently. The design enables remote initialization, control of the accelerator devices, and host-independent accelerator-to-accelerator direct communication. Workloads can be dynamically assigned to CPUs and accelerators at run-time in an N to M ratio. An operative prototype implementation, based on the Intel Xeon Phi coprocessor and the EXTOLL NIC, is used to evaluate the latency, bandwidth, performance of the MPI communication, and communication time of the LAMMPS molecular dynamics simulator.
Two-page extended abstract: pdf