Towards a Noise-Resilient TAGE via PURE

摘要

Modern server workloads exhibit large instruction footprints and complex control-flow, leading to significant mispredictions even for the state-of-the-art TAGE-SC-L. We demonstrate that prediction quality is not primarily limited by TAGE’s intrinsic representational power or capacity, but by indiscriminate allocations that waste predictor capacity and impede online learning. Our analysis reveals that if shielded from unnecessary allocations, a standard 64KB TAGE can achieve performance approaching that of an infinite-capacity TAGE.Driven by this observation, we propose PURE (Pattern-Utility Rejection of Entry Allocation), a mechanism based on input sanitization. Unlike existing profile-guided solutions that bypass the predictor by offloading targeted branches to auxiliary models, PURE uses offline profiles to identify and filter noise-dominated patterns before they enter the predictor. This approach improves TAGE’s learning efficiency and preserves capacity for useful patterns across branches of varying prediction difficulty. Across 12 widely used server applications, PURE reduces MPKI by an average of 25.2% (7.8% to 41.7%) over a 64KB TAGE-SC-L, outperforming state-of-the-art profile-guided schemes by 9.3% with 22KB of metadata.