Slipstream-Based Steering for Clustered Microarchitectures

Abstract

To harvest increasing levels of ILP while maintaining a fast clock, clustered microarchitectures have been proposed. However, the fast clock enabled by clustering comes at the cost of multiple cycles to communicate values among clusters. A chief performance limiter of a clustered microarchitecture is inter-cluster communication between instructions. Specifically, inter-cluster communication between critical-path instructions is the most harmful. The slipstream paradigm identifies critical-path instructions in the form of effectual instructions.

We propose eliminating virtually all inter-cluster communication among effectual instructions, simply by ensuring that the entire effectual component of the program executes within a cluster. This thesis proposes two execution models: the replication model and the dedicated-cluster model. In the replication model, a copy of the effectual component is executed on each of the clusters and the ineffectual instructions are shared among the clusters. In the dedicated-cluster model, the effectual component is executed on a single cluster (the effectual cluster), while all ineffectual instructions are steered to the remaining clusters. Outcomes of ineffectual instructions are not needed (in hindsight), hence their execution can be exposed to inter-cluster communication latency without significantly impacting overall performance.

IPC of the replication model on dual clusters and quad clusters is virtually independent of inter-cluster communication latency. IPC decreases by 1.3% and 0.8%, on average, for a dual-cluster and quad-cluster microarchitecture, respectively, when inter-cluster communication latency increases from 2 cycles to 16 cycles. In contrast, IPC of the best-performing dependence-based steering decreases by 35% and 55%, on average, for a dual-cluster and quad-cluster microarchitecture, respectively, over the same latency range. For dual clusters and quad clusters with low latencies (fewer than 8 cycles), slipstream-based steering underperforms conventional steering because improved latency tolerance is outweighed by higher contention for execution bandwidth within clusters. However, the balance shifts at higher latencies. For a dual-cluster microarchitecture, dedicated-cluster-based steering outperforms the best conventional steering on average by 10% and 24% at 8 and 16 cycles, respectively. For a quad-cluster microarchitecture, replication-based steering outperforms the best conventional steering on average by 10% and 32% at 8 and 16 cycles, respectively.

Slipstream-based steering desensitizes the IPC performance of a clustered microarchitecture to tens of cycles of inter-cluster communication latency. As feature sizes shrink, it will take multiple cycles to propagate signals across the processor chip. For a clustered microarchitecture, this implies that with further scaling of feature size, the inter-cluster communication latency will increase to the point where microarchitects must manage a distributed system on a chip. Thus, if individual clusters are clocked faster, at the expense of increasing inter-cluster communication latency, performance of a clustered microarchitecture using slipstream-based steering will improve considerably as compared to a clustered microarchitecture using the best conventional steering approach.