Shared Frontend for Manycore Server Processors

Instruction-supplymechanisms, namely the branch predictors and instruction prefetchers, exploit recurring control flow in an application to predict the application’s future control flow and provide the core with a useful instruction stream to execute in a timely manner. Consequently, instruction-supplymechanisms aggressively incorporate control-flow condition, target, and instruction cache access information (i.e., control-flow metadata) to improve performance. Despite their high accuracy, thus performance benefits, these predictors lead to major silicon provisioning due to their metadata storage overhead. The storage overhead is further aggravated by the increasing core counts and more complex software stacks leading to major metadata redundancy: (i) across cores as the metadata of cores running a given server workload significantly overlap, (ii) within a core as the control-flowmetadata maintained by disparate instruction-supplymechanisms overlap significantly. In this thesis, we identify the sources of redundancy in the instruction-supply metadata and provide mechanisms to share metadata across cores and unify metadata for disparate instruction-supply mechanisms. First, homogeneous server workloads running on many cores allow for metadata sharing across cores, as each core executes the same types of requests and exhibits the same control flow. Second, the control-flow metadata maintained by individual instruction-supply mechanisms, despite being at different granularities (i.e., instruction vs. instruction block), overlap significantly, allowing for unifying their metadata. Building on these two observations, we eliminate the storage overhead stemming from metadata redundancy inmanycore server processors through a specialized shared frontend, which enables sharing metadata across cores and unifying metadata within a core without sacrificing the performance benefits provided by private and disparate instruction-supply mechanisms.