Optimizing Memory-mapped I/O for Fast Storage Devices
summary
Presents FastMap, a redesigned Linux memory-mapped I/O path that reduces synchronization in the common path and increases device queue depth to make mmap-style access scale on many cores. Evaluates FastMap against Linux mmap using null_blk and an Optane SSD, showing scalability to 80 cores and up to 11.8× higher IOPS and 5.27× higher throughput.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:26Fast storage devices
- 0:52Read/write system calls
- 1:30Memory-mapped I/O
- 2:34Disadvantages of mmio
- 3:00Mmio path scalability
- 5:09FastMap design: 3 main techniques
- 5:42Linux mmio design
- 9:40Linux object-based reverse mappings
- 11:13FastMap full reverse mappings
- 12:26Batched TLB invalidations
- 13:19Other optimizations in the paper
- 13:29Outline
- 14:01Workloads
- 14:42FastMap Scalability
- 16:23FastMap execution time breakdown
- 17:04Kreon key-value store
- 17:29Kreon - 100% inserts
- 18:23Kreon - 100% lookups
- 19:33Conclusions