AUTOMATED FPGA FIRMWARE MANAGEMENT IN HPC CLUSTERS
Scientific Bulletin of Naval Academy · 2016
FPGA-based accelerators are increasingly deployed on cluster and grid systems due to their highly flexible architecture.Given the generic nature of a high performance computing system, the firmware and software running on the FPGAs changes dynamically, according to the specifications requested by the launched application.Along with performance monitoring, this reconfiguration process can be automated in order to decrease idle-times on computing nodes and to have a centralized view of the system.Such an architecture would be centered around a client-server model in which the computing nodes run the client component, along with the batch agent.The server component would be located anywhere in the cluster as long as it has the appropriate permission to interact with the batch server.The paper explores the possibility of integrating this reconfiguration model with an existing batch system, without major changes in the way users operate the cluster.