The Five-Minute Rule 20 Years Later (and How Flash Memory Changes the Rules) - HP Labs, 2009

This article was written before SSD’s became common in laptops.

  • Flash usage
    • OS/File systems will use flash as extended RAM
      • No log so they need to write more quickly to persistent storage
    • DBMSes will use flash as extended persistent storage
      • System-wide checkpoints that flush the log and buffers
  • B-Trees should use different size of I/O for flash than they currently due with disc (due to seek time)
  • The flash / disk trade-off is interesting since the brake even point for pages is a little over 2 hours
  • The author notes that original reports prediction for 20 years into the future was “amazingly accurate” for RAM versus disk (5 hours predicted / 6 actual)
  • The author also discusses the issue of energy efficiency as part of the trade-off
    • Not only in battery powered devices but also in server farms
  • Future Work
    • CPU should be measured in terms of cache line replacements, not instructions
      • I found this interesting as a lot of experiments have shown cache misses to be a major bottleneck in computation throughput
    • In a database, short access times to data on flash allow for less multi-programming
      • I note that often we overlapped work with I/O in databases
      • The author notes that less multi-programming could reduce lock contention and could possibly allow for coarser grained locking.
    • GC could also benefit from flash, some objects may live for long periods of time but not need to be accessed frequently