Power-constrained aware and latency-aware microarchitectural optimizations in many-core processors
- Autores: Sudhanshu Shekhar Jha
- Directores de la Tesis: Jordi Tubella Murgadas (dir. tes.), Antonio González Colás (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2016
- Idioma: español
- Tribunal Calificador de la Tesis: Daniel Ortega Fernández (presid.), Joan Manel Parcerisa Bundó (secret.), Tanausu Ramírez García (voc.)
- Materias:
- Enlaces
- Resumen
As the transistor budgets outpace the power envelope (the power-wall issue), new architectural and microarchitectural techniques are needed to improve, or at least maintain, the power efficiency of next-generation processors. Run-time adaptation, including core, cache and DVFS adaptations, has recently emerged as a promising area to keep the pace for acceptable power efficiency.
However, none of the adaptation techniques proposed so far is able to provide good results when we consider the stringent power budgets that will be common in the next decades, so new techniques that attack the problem from several fronts using different specialized mechanisms are necessary. The combination of different power management mechanisms, however, bring extra levels of complexity, since other factors such as workload behavior and run-time conditions must also be considered to properly allocate power among cores and threads.
To address the power issue, this thesis first proposes Chrysso, an integrated and scalable model-driven power management that quickly selects the best combination of adaptation methods out of different core and uncore micro-architecture adaptations, per-core DVFS, or any combination thereof. Chrysso can quickly search the adaptation space by making performance/power projections to identify Pareto-optimal configurations, effectively pruning the search space. Chrysso achieves 1.9x better chip performance over core-level gating for multi-programmed workloads, and 1.5x higher performance for multi-threaded workloads.
Most existing power management schemes use a centralized approach to regulate power dissipation. Unfortunately, the complexity and overhead of centralized power management increases significantly with core count rendering it in-viable at fine-grain time slices. The work leverages a two-tier hierarchical power manager. This solution is highly scalable with low overhead on a tiled many-core architecture with shared LLC and per-tile DVFS at fine-grain time slices. The global power is first distributed across tiles using GPM and then within a tile (in parallel across all tiles). Additionally, this work also proposes DVFS and cache-aware thread migration (DCTM) to ensure optimum per-tile co-scheduling of compatible threads at runtime over the two-tier hierarchical power manager. DCTM outperforms existing solutions by up to 12% on adaptive many-core tile processor.
With the advancements in the core micro-architectural techniques and technology scaling, the performance gap between the computational component and memory component is increasing significantly (the memory-wall issue). To bridge this gap, the architecture community is pushing forward towards multi-core architecture with on-die near-memory DRAM cache memory (faster than conventional DRAM). Gigascale DRAM Caches poses a problem of how to efficiently manage the tags. The Tags-in-DRAM designs aims at efficiently co-locate tags with data, but it still suffer from high latency especially in multi-way associativity.
The thesis finally proposes Tag Cache mechanism, an on-chip distributed tag caching mechanism with limited space and latency overhead to bypass the tag read operation in multi-way DRAM Caches, thereby reducing hit latency. Each Tag Cache, stored in L2, stores tag information of the most recently used DRAM Cache ways. The Tag Cache is able to exploit temporal locality of the DRAM Cache, thereby contributing to on average 46% of the DRAM Cache hits.
