Enabling Big Memory workloads to run
summary
MemVerge demonstrates a user-space software-defined memory platform that augments DRAM with persistent memory and integrates with Loki/OpenStack, using automatic tiering and copy-on-write snapshots. Actual results: a 300 million-key Redis snapshot that normally needs 15 minutes to dump is restored in 0.5 seconds with PMem, metaSPAdes runs on a 96-core node with 2 TB RAM instead of a 50-node cluster, and some workloads drop wall-clock time ~25%. Hardware claim: persistent memory uses ~10% the power of DRAM.
AI-written summary · about
chapters (8)
- 0:00Start
- 5:00Breaking Down the Memory Wall and
- 8:04The Need for Big Memory Software
- 16:05Bigger Memory Calls for Better Protection
- 18:03Lower Time-to-Results for Scientific Computing
- 19:59Rolling Back Becomes instant
- 23:36NASTRAN workload memory and disk usage
- 24:15NASTRAN Test results (seconds)
related talks
18mCommunity Software Powers the Machine
29mThe Toy-SRAM Project
14mKeynote: Building and Sustaining Momentum in Open-Source Storage
20mEnhance the Performance of QEMU RVV Load/Store Implementation
16mSoftware-driven Hardware in the Data Center: LibreBMC, Renode, RowHammer...
33mIntroduction to Redis
17mComposable Data Management Systems
06mDay 1 Keynote