Nomad: Non-Exclusive Memory Tiering via Transactional Page Migration
Talkat 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI '24) · 2024-07-10..07-12·
summary
Nomad implements transactional page migration and page shadowing to keep migrated pages accessible and retain shadow copies in the slow tier, turning exclusive tiering into non-exclusive memory tiering. Migration runs asynchronously via a kprom kernel thread and uses exactly one minor-fault per page to queue candidates; transactions abort if the dirty bit flips. On three testbeds (DDR+Optane PM, FPGA CXL prototype, AMD+Micron) microbenchmarks and real apps show up to 6x speedup over Linux TPP under memory pressure, while TPP can suffer 16x bandwidth loss during migration and Nomad degrades gracefully but sees aborts on write-heavy workloads.
AI-written summary · about
chapters (8)
- 0:00Introduction to Nomad
- 0:31Memory tiering background
- 2:47Limitations of current systems
- 3:41Access tracking and migration
- 6:01Nomad design mechanisms
- 10:12Page shadowing mechanism
- 11:52Evaluation and results
- 14:44Conclusions
related talks
43mLive in a world with multiple memory types
11mAdvanced Parallel Memory Virtualization
48mPersistent Memory Usage within Linux Environment
30mMemory tiering
27mLive Migration With Hardware Acceleration
30mA move_pages() equivalent for physical memory
14mArchTM: Architecture-Aware, High Performance Transaction for Persistent Memory
12mSPHT: Scalable Persistent Hardware Transactions