Rio Vista

1
Rio Vista
2
Takeaway

• Transactions are considered heavy weight
• Commit at least takes one synchronous disk I/O
• Systems avoid transactions despite its precise
  semantics due to the time penalty
• If disk I/O eliminated then transactions would be
  cheap
• Two problems concerning failure
• Power failure
• Throw some hardware at this problem (100 gets
  you a UPS power supply)
• Software crash
• Reserve a portion of main memory that survives
  crashes

3
RVM revisited
4
Rio File Cache

• UPS backed memory so contents never goes away
• VM protection to prevent OS errors such as wild
  writes to file cache during crash
• Upon crash, file cache data in memory written to
  disk for recovery
• Upshot
• No need for synchronous writes to disk
• Write-back of files to disk can be arbitrarily
  delayed (how does this help?)

5

• Two ways to use Rio file cache
• File write
• Need fsync in the absence of Rio for commitment
• Mmap file into application space
• Need msync in the absence of Rio for commitment

6
Vista recoverable memory

• Exactly similar to Satyas LRVM
• Lives on top of Rio file cache
• How it works
• Undo log for before image
• Commit throw away undo log
• Abort recover before image from undo log
• Implications
• No disk I/O!
• No redo log
• A Disk DB becomes a memory resident DB

7
How Vista works
8
Crash Recovery

• Treat it like an abort to recover old image from
  the undo log (which survives crashes since it is
  in Rio)

9
Vista simplicity

• 700 lines of code (as opposed to 10K lines in
  RVM)
• How?
• No redo logs or log truncation code
• Checkpointing and recovery code simplified
• No group commit optimizations
• Simple semantics like LRVM

10
Performance