LIBRA : Multi-mode on-chip network arbitration for locality-oblivious task placement

Abstract

As multi-hop on-chip networks (OCNs) are used to provide access to shared resources, such as memory controllers, for chip multiprocessors, guaranteeing global fairness among the nodes is important in future OCNs, regardless of their location. However, we show how placement of the tasks with respect to the location of the memory controllers can result in a significant performance variation, up to 30in our evaluation. Since current operating systems are oblivious to OCN and its topology, we provide architectural support for locality-oblivious task placement through on-chip network arbitration. We propose virtualcontention-based arbitration (VCA) that introduces virtual contention ?? i.e., even if contention does not occur, VCA takes historical non-contention into account when performing arbitration in the future and overcomes the limitation ofrealcontention-based arbitration. In addition, since arbitration is often on the router critical path (and arbiters such as VCA will further increase the complexity), we propose a hybridarbiter that switches between a simple arbiter and a complex arbiter. The hybrid arbiter is enabled by the observation that VCA only impacts the overall performance and global fairness at a high load. Thus, using a simple arbiter such as a round-robin arbiter at a low load has minimal impact on the overall performance, while effectively hiding the latency. The combination of VCA and the hybrid arbiter resultsin LIBRA (Locality-obliviousBandwidthRegulatoryArbiter). Our results show that LIBRA improves fairness by 2.73and 1.77compared with round-robin and ideal age-based arbiter while having minimal impact on the router cycle time. Compared with existing QoS schemes that provide global fairness, LIBRA uses 38% less hardware than the QoS scheme requiring the least hardware.