TL;DR
This paper introduces a novel technique for managing Resistive RAM (RRAM) Last Level Caches (LLCs) that dynamically adjusts the cache size to match application needs, saving significant energy. Crucially, it also addresses RRAM's limited write endurance by distributing writes evenly across the cache, thereby extending its lifespan.
What Problem Does This Paper Address?
Modern processors rely on large Last Level Caches (LLCs) to keep up with increasing core counts and program demands. While traditional SRAM caches are fast and durable, they consume a lot of power due to leakage. Non-volatile memories (NVMs) like RRAM offer a compelling alternative with high density and low leakage. However, NVMs, particularly RRAM, suffer from high write latency and, more critically, limited write endurance. Existing energy-saving techniques that dynamically turn off cache portions can exacerbate this write endurance issue by concentrating writes on a few active blocks, drastically reducing the cache's lifespan. This paper tackles the challenge of leveraging NVMs for energy-efficient caches without sacrificing their longevity.
Methods
The core of the proposed solution is a dynamic cache reconfiguration strategy based on cache coloring. The cache is logically divided into multiple "colors" (groups of cache lines). The system monitors the application's working set size and dynamically adjusts the number of active cache colors. When the working set is small, fewer colors are activated, and the rest are powered down to save leakage energy.
To address RRAM's write endurance limitation, the technique incorporates a wear-leveling mechanism. It tracks the write count for each color. When reconfiguring the cache (turning colors on or off), it prioritizes turning off colors that have experienced the most writes and turning on colors that have experienced the fewest. This ensures that writes are distributed more evenly across the entire cache, preventing premature wear-out of specific blocks.
The system uses a profiling unit to estimate the application's miss rate under different cache configurations, allowing it to make informed decisions about how many colors to activate without significant performance degradation. The reconfiguration process is performed at large intervals (e.g., every 10 million instructions) to amortize the overhead.
Datasets
The evaluation uses the SPEC2006 benchmark suite with "ref" inputs. The paper does not introduce new benchmarks but uses these standard workloads to simulate various application behaviors.
Key Insights and Ideas
- Dynamic Cache Sizing is Key: The paper demonstrates that adapting cache size to workload requirements is crucial for energy savings, especially with high-density NVMs.
- Write Endurance is a Critical Bottleneck for NVM Caches: Simply applying existing energy-saving techniques to NVMs can be detrimental to their lifespan due to limited write endurance.
- Cache Coloring Enables Fine-Grained Control: The cache coloring scheme provides a flexible mechanism to reconfigure the cache at a granularity that balances effectiveness and overhead.
- Wear-Leveling is Essential for NVM Cache Longevity: Actively managing write distribution across cache regions is paramount for achieving practical lifetimes with NVMs.
- RRAM's Leakage Advantage Outweighs Write Latency for Many Workloads: For applications with large working sets, the energy savings from RRAM's low leakage can compensate for its higher write latency, leading to performance improvements.
Key Results
- Significant Energy Savings: The proposed technique achieves an average of 17.55% saving in memory subsystem energy compared to a baseline 8MB RRAM LLC.
- Minimal Performance Impact: The average speedup is 0.99x, indicating negligible performance loss. The increase in miss rate (MPKI) is also very small (0.09).
- Improved Cache Lifetime: The technique provides an average improvement in cache lifetime of 1.33x.
- Comparison with SRAM: An SRAM cache of similar area (but much smaller capacity, 1MB) results in an energy loss of 186.13% and a performance degradation of 0.96x, highlighting the benefits of RRAM's capacity.
- Endurance-Aware vs. Unaware: An "Endurance_Unaware" version of the algorithm showed similar energy savings but a lifetime improvement of only 0.99x, underscoring the importance of the wear-leveling component.
Limitations and Future Work
- Granularity of Wear-Leveling: The technique operates at the granularity of cache colors, not individual cache sets or blocks. This means it might not perfectly level writes if a block with high write activity is not in a color with high write activity. The authors believe this is not a common issue.
- Cache Bypass for Streaming Workloads: For specific streaming benchmarks (libquantum, milc), the technique degraded lifetime. The authors suggest that techniques like cache bypass could further enhance lifetime for such workloads, which is planned for future work.
- Multicore Processors: The paper suggests extending the technique to multicore processors as a future direction.
- Reducing Write Traffic: Integrating methods to reduce write traffic (e.g., write buffers, filter caches) and error correction mechanisms could further enhance RRAM cache lifetime.
Summary updated