A Technique for Improving Lifetime of Non-Volatile Caches Using Write-Minimization

Sparsh Mittal, Jeffrey S. VetterJournal of Low Power Electronics and Applications 2016 · DOI

TL;DR

This paper introduces ENLIVE, a novel technique that significantly extends the lifespan of non-volatile memory (NVM) caches by using a small, fast SRAM "HotStore" to buffer frequently written data. By intelligently migrating hot blocks to the HotStore, ENLIVE dramatically reduces writes to the NVM, leading to lifetime improvements of up to 15.79x while also boosting performance and energy efficiency.

What Problem Does This Paper Address?

As computing systems pack more cores, the demand for larger last-level caches (LLCs) to feed these cores has exploded. Traditionally, these caches use SRAM, which offers high performance and endurance but suffers from high leakage power and density limitations. Non-volatile memories (NVMs) like ReRAM and STT-RAM offer a compelling alternative with high density and low leakage, but they have a critical drawback: significantly lower write endurance. This limited write endurance, coupled with typical cache management policies that can concentrate writes on a few blocks (write-variation), leads to premature cache failure. Existing techniques often focus on wear-leveling (distributing writes) rather than reducing the total number of writes, which is insufficient to overcome the fundamental endurance limitations of NVMs.

Methods

ENLIVE tackles the write-endurance problem by introducing a small, dedicated SRAM buffer called the "HotStore." The core idea is to identify frequently written blocks within the NVM cache and temporarily move them to this HotStore. When a write operation targets a block already in the HotStore, it's handled by the fast SRAM, bypassing the NVM entirely. This not only speeds up writes but, more importantly, drastically reduces the number of writes that reach the NVM, thus preserving its limited endurance.

The paper details the microarchitectural schemes for managing the HotStore:

  • Insertion: Blocks are promoted to the HotStore if their write count exceeds a threshold ($\lambda$) and if there's space or if they are "hotter" than a potential replacement.
  • Eviction: When the HotStore is full, a replacement policy (like Least Number of Writes - LNW) is used to select a candidate for eviction. If the incoming block is deemed more write-intensive, the evicted block is written back to the NVM cache.
  • HotStore Management: The HotStore itself doesn't require a separate tag directory; it leverages the L2 cache's tag directory, simplifying the design. The paper also discusses two replacement policies: LNW (which performed better) and NRU.

Datasets

The evaluation uses the SPEC CPU2006 suite and six benchmarks from the High-Performance Computing (HPC) field. These benchmarks are combined into multiprogrammed workloads for dual, quad, and eight-core systems. The paper does not introduce new benchmarks but rather uses a comprehensive set of existing ones to simulate realistic system behavior.

Key Insights and Ideas

  • Write-Minimization is Key: Unlike wear-leveling techniques that merely spread writes, ENLIVE directly reduces the total number of writes to the NVM, which is crucial for overcoming its inherent endurance limitations.
  • SRAM as a "Hot Block" Buffer: Leveraging the speed and high endurance of SRAM for frequently written blocks is an effective strategy to protect the NVM.
  • Synergy of HotStore and NVM: The HotStore acts as a high-performance, high-endurance buffer, allowing the NVM to be used more efficiently for less frequently written data.
  • Security Implications: ENLIVE's ability to mitigate repeated address attacks (RAA) is a significant benefit, as it makes NVM caches much more resilient to malicious attempts to wear them out.
  • Small Overhead, Big Gains: The proposed HotStore is small (e.g., ~2% of L2 cache size), incurring minimal area overhead while delivering substantial lifetime improvements.

Key Results

  • Lifetime Improvement: ENLIVE achieved significant lifetime improvements: 8.47x for dual-core, 14.67x for quad-core, and 15.79x for eight-core systems. This is notably higher than competing techniques like PoLF and LastingNVCache.
  • Performance: ENLIVE slightly improves performance due to faster writes from the HotStore.
  • Energy Efficiency: ENLIVE incurs minimal energy loss (around 2% for 8-core systems) and avoids the large energy losses associated with techniques that increase writes to main memory.
  • Write Reduction: ENLIVE significantly reduces writes to the NVM cache, with an average reduction of 10.3% to 16.7% in writes per kilo-instruction (WPKI), despite the HotStore only holding a small fraction of the L2 cache's blocks.
  • Parameter Sensitivity: The technique shows robustness across various parameters like HotStore size, threshold values, and associativity, allowing for trade-offs between overhead and lifetime improvement.

Limitations and Future Work

  • HotStore Redundancy: The paper acknowledges that data in the HotStore is also stored in the NVM LLC, leading to some redundancy. However, due to the small size of the HotStore, this redundancy is considered small.
  • Future Work: The authors plan to explore synergistically integrating ENLIVE with other write-reduction mechanisms (like cache compression and bit-level write reduction) and wear-leveling techniques. They also aim to investigate prefetching for NVM caches and explore the use of NVMs for GPU caches, considering the growing importance of power efficiency in GPUs. Additionally, they plan to address soft-error mitigation in NVM caches.

Summary updated

Abstract

While non-volatile memories (NVMs) provide high-density and low-leakage, they also have low write-endurance. This, along with the write-variation introduced by the cache management policies, can lead to very small cache lifetime. In this paper, we propose ENLIVE, a technique for ENhancing the LIfetime of non-Volatile cachEs. Our technique uses a small SRAM (static random access memory) storage, called HotStore. ENLIVE detects frequently written blocks and transfers them to the HotStore so that they can be accessed with smaller latency and energy. This also reduces the number of writes to the NVM cache which improves its lifetime. We present microarchitectural schemes for managing the HotStore. Simulations have been performed using an x86-64 simulator and benchmarks from SPEC2006 suite. We observe that ENLIVE provides higher improvement in lifetime and better performance and energy efficiency than two state-of-the-art techniques for improving NVM cache lifetime. ENLIVE provides 8.47×, 14.67× and 15.79× improvement in lifetime or two, four and eight core systems, respectively. In addition, it works well for a range of system and algorithm parameters and incurs only small overhead.
BibTeX
@article{mittal2016technique,
  author = {Sparsh Mittal and Jeffrey S. Vetter},
  title = {{A Technique for Improving Lifetime of Non-Volatile Caches Using Write-Minimization}},
  journal = {Journal of Low Power Electronics and Applications},
  year = {2016},
  doi = {10.3390/jlpea6010001},
}
Figure