Clustered machines partition hardware resources to circumvent the cycle time penalties incurred by large, monolithic structures. This partitioning introduces a long inter-cluster forwarding latency and the potential for load imbalance, both of which degrade IPC and thus counter the cycle time benefits of clustering.

We show that program dataflow can be mapped to clustered machines so as to achieve an IPC rivaling that of an equivalent monolithic machine. That is, the IPC penalties observed by extant schemes are largely an artifact of instruction steering and scheduling policies. Using critical path analysis, we investigate and uncover the main causes for this performance loss. By way of code samples, we illustrate those causes and propose three policies for mitigating them. First, we introduce a new metric, likelihood of criticality, and show how it can halve the performance lost to contention-induced stalls. Second, we develop a stall-over-steer policy that addresses performance lost to inter-cluster forwarding delay. Finally, we show that a proactive load-balancing policy is necessary to improve the distribution of ready instructions among the clusters. Together, these three policies yield performance on 2-, 4- and 8-cluster implementations of an 8-wide machine that is within 2, 4, and 6%, respectively, of the monolithic equivalent.