Optimizing Multicore and Persistent Memory Systems

Current Trends in Multicore and Persistent Memory Systems

Recent advancements in multicore and persistent memory systems are addressing critical challenges such as cache contention, write endurance in Non-Volatile Memory (NVM), and the performance of read-only transactions. Innovations in cache management are focusing on balancing write traffic to extend the lifespan of NVM cells and preventing malicious cache bank-aware denial-of-service attacks. These approaches aim to enhance system throughput and predictability, particularly in real-time environments. Additionally, there is a notable shift towards optimizing read-only transactions in persistent hardware, leveraging advanced hardware transactional memory features to achieve significant performance boosts. The integration of analytical and transactional workloads is also being addressed with new indexing techniques that support efficient concurrent updates, thereby improving the performance of hybrid transactional/analytical processing (HTAP) systems.

Noteworthy Developments

  • A per-bank bandwidth regulation approach for shared last-level caches shows substantial performance improvements in real-time systems under cache bank-aware DoS attacks.
  • A sample-based blocking technique for the Last Level Cache effectively balances write traffic in NVM, extending memory cell lifespan.
  • DUMBO, a design for Persistent Hardware Transactions, significantly enhances the performance of read-only transactions on hardware transactional memory.
  • CUBIT, a concurrent updatable bitmap indexing technique, offers substantial improvements in throughput and latency for hybrid transactional/analytical processing workloads.

Sources

Per-Bank Bandwidth Regulation of Shared Last-Level Cache for Real-Time Systems

LLC Intra-set Write Balancing

DUMBO: Making durable read-only transactions fly on hardware transactional memory

CUBIT: Concurrent Updatable Bitmap Indexing

Built with on top of