- Goal: Show IO related problems and possible solutions
  - not really user oriented
  - things are mostly good, "more extreme" cases are having problems
  - landscape changing, new types of IO
- speaking for myself, not the project
- just ideas, partially not fully developed, wont implement everything myself
- ask questions

- previous talk mentioned improvments
  - mostly lock-free pin/unpin
  - dirty buffer management
  - sorted checkpoints
- lotsa other improvements

- Explain: buffer mapping
- Explain: clock sweep

- Problem: lookups are slow
- Problem: lookups scale badly
- Problem: truncation, dropping tables, creating databases, etc. needs to flush data
- Problem: locks for extension, truncation
- Solution: cache locality
- Solution: ordered data structure
- Solution: shared open relation map

- Problem: backends are doing most write IO
  - mention pg_stat_database et al.
- Solution: replace bgwriter -> sweeper

- Problem: clock sweep is expensive
- Solution: sweeper?
- Solution: continous sweep based on usage ratio
- Solution: different data structure?

- Problem: wrong buffers are being replaced
- Solution #1: more intelligent usagecount management
- Solution #2: different datastructure

- Problem: too much random IO
- Solution: batch neighbouring writes

- Problem: double buffering
- Solution: O_DIRECT?
- Hard: latency around writes changes drastically
- Hard: Portability sucks
- Solution: Use buffered IO but hint aggressively?
- Hard: restart performance characteristics