Break On Through to the Other Side: Pooling Memory Elastically with RamRyder
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
RamRyder is a software-defined elastic memory system that maps guest physical pages to specific DRAM channels (obtained via BIOS/UEFI partitioning and reverse engineering) to control bandwidth and capacity independently. It exposes channels as channel NUMA nodes, uses CXL devices for capacity/bandwidth extension, runs on AMD servers with 12 DDR5 channels per socket, and implements lazy page-migration for runtime adjustments. It boosts cluster capacity and bandwidth utilization by 28.6% and 43.2%, adds <5% overhead from software interleaving, and can miss very short bursts due to seconds-long redistribution.
AI-written summary · about
chapters (12)
- 0:00RamRyder and memory elasticity
- 1:25Memory underutilization in cloud
- 3:52Overprovisioning causes and limits
- 6:02RamRyder goals and insight
- 7:49RamRyder architecture
- 8:53Bandwidth allocation with channel NUMA
- 10:42Independent bandwidth and capacity
- 11:49Runtime adjustment of resources
- 13:19Evaluation setup and benefits
- 14:47Utilization with complementary workloads
- 15:34Overheads and optimizations
- 16:37Conclusion and open-source
related talks
17mPreview: Persistent Memory
16mJohnny Cache: the End of DRAM Cache Conflicts (in Tiered Main Memory Systems)
17mBreakthrough Data-Centric Computing with a New Memory Tier
35mDeveloping Software for Persistent Memory
43mPromise and Pitfalls of Persistent Memory
15mImproving Performance of Flash Based Key-Value Stores Using Storage Class Memory...
25mDisaggregating Persistent Memory and Controlling Them Remotely
01h 07mEmerging Embdedded Memories and Storage