NUDA: Non-Uniform Directory Architecture for Scalable Chip Multiprocessors

Chip multiprocessors (CMPs) involve directory storage overhead if cache coherence is realized via sharer tracking. This work proposes a novel framework dubbed non-uniform directory architecture (NUDA), by leveraging our two insights in that the number of "active" directory entries required to stay on chip is usually small for a short execution time window due to high directory locality, and that the fraction of interrogated directory entries drops as the core count rises. Unlike earlier storage overhead reduction techniques that require all cached LLC blocks to have their directory entries fully on chip, NUDA dynamically buffers only most active directory vectors (DVs) on chip while keeping DVs of all LLC blocks in a backing store at low level storage. NUDA attains its superior efficiency via an inventive criticality-aware replacement policy (CARP) for on-chip buffer management and effective prefetching to pre-activate vectors (PAVE) for upcoming coherence interrogations. We have evaluated NUDA by gem5 simulation for 64-core CMPs under PARSEC and SPLASH benchmarks, demonstrating that CARP and PAVE enhance on-chip directory storage efficiency significantly. NUDA with a small on-chip buffer for DVs exhibits negligible performance degradation (to stay within 2.6 percent) compared to a full on-chip directory, while outperforming its previous counterparts for directory area reduction when on-chip directory budget is provisioned scarcely for high scalability.