Para-ksm: Parallelized Memory Deduplication with Data Streaming Accelerator
summary
Para-ksm implements a redesigned KSM that uses Intel Data Streaming Accelerator to offload page comparisons and checksum computation and batches multiple pages per offload to amortize on-chip PCIe latency. On 4th‑gen Xeon with DSA, a naïve DSA-ksm reduces co-runner slowdowns but kills dedupe rates; Para-ksm restores deduplication efficiency, boosting memory-deduplication per CPU cycle by ~31–50% and cutting KSM-induced performance degradation from roughly 1.6–5.8× down to about 1.3–2.7×.
AI-written summary · about
chapters (15)
- 0:00Para-ksm introduction
- 0:40Memory deduplication background
- 1:45Data Streaming Accelerator
- 2:47DSA-KSM design findings
- 3:53Batching to reduce latency
- 5:21KSM batch limitations
- 6:26Para-ksm techniques
- 6:47ParaKSMc candidate batching
- 8:54Batch insertion in trees
- 12:49ParaKSMt tree batching
- 13:52Evaluation setup
- 14:12Performance results
- 15:39Memory savings
- 16:01Duplication efficiency
- 16:23Summary
related talks
43mKernel Samepage Merging (KSM) at Meta and Future Improvements to KSM
42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
31mHMSDK: Enhancing Software Support for CXL Memory
22mAccelerate the Image Distribution for DADI Using Intel DSA- 42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
15mOsprey: Transparent and Efficient Virtual Memory for Secure Computation
27mMitigating speculative execution attacks with ASI - follow up
14mDPAS: A Prompt, Accurate and Safe I/O Completion Method for SSDs