MIND: In-Network Memory Management for Disaggregated Data Centers

https://www.anuragkhandelwal.com/papers/mind.pdf

• Resource disaggregation

  • High resource utilization, ease to manage, elastic scalability

  • Memory disaggregation

    • Need for memory disaggregation

      • Now: suffer from low utilization, unbalanced usage, energy consumption

      • Timely problem (storage disaggregation is already popular)

    • Goal:

      • Performance (CPU <--> memory): high throughput and low latency

      • Transparent elasticity: flexible resource allocation

    • Existing approach

      • Distributed shared memory (DSM)

        • Good elasticity but not good performance

      • Recent disaggregated memory schemes

        • Compromise elasticity by high throughput (no sharing0

      • Key insight: in-network memory management (MMU)

        • Central location: global view & processing directly in the data path

        • Programmable switching ASIC: flexible processing at line rate

        • Similarity between network functions and memory management

          • 
    • MIND: in-network memory management

      • 

        • Address translation: where is the data

        • Memory protection: is this access permitted or valid?

        • Cache coherence protocol: how to keep data synchronized?

    • Challenges in Building In-network MMU

      • Limited amount of in-network resources

        • Limited size of in-network memory

          • Tens of MB --> not sufficient to store metadata in a traditional way

        • Limited computation capability

          • Switching ASIC --> not sufficient to directly port traditional MMU functions

    • 3 Principles for System Design

      • P1: decouple memory management functionalities

        • Each function has the data structure suitable to its purpose

      • P2: Leverage global view of network

        • Make better decisions for memory management

      • P3: Exploit network-centric hardware primitives

        • Reuse network hardware highly optimized for network functions

    • Challenges

      • Address translation

        • Small: too many entries, huge page: load balancing

        • Apply

          • Static huge partition

          • Fine-grained, range-based translation

      • Memory protection

        • Small: too many entries, huge page: inflexible protection

        • Solve: VMA-granularity protection

      • Cache directory

        • Small cache line: too many entries, large cache line: false sharing

        • Solve: access granularity, coherence granularity (dynamic resizing), with theoretical bound of O(log2)

• Performance evaluation

  • Low contention, high contention

• Conclusion

  • Trade-off between resource elasticity and performance in memory aggregation

  • MIND: in-network MMU by leveraging programmable network

  • Match the performance of prior proposals and provide transparent elasticity