LeanStore

• Describes how the pre-cursor Umbra’s buffer manager works
• Slight modifications result in Umbra’s implementation. Crux of the implementation is in this paper.
• Background contains some other hybrid buffer managers
  • Can be looked into if needed.
• Each optimisation technique described can be used individually
  • Pointer swizzling
  • Page replacement policy
    • Randomly set 10% of pages as cooling
    • Evict once they reach end of FIFO queue
      • Used pages, move out of queue
    • 
  • Latching
    • Better implementation is present in Umbra. Use this only if ref is needed
• Data structures needed for keeping track of state of buffer manager
  • Described well in LeanStore. Not present in Umbra
• Buffer-managed data structures
  • Talks about owning swips and the swip hierarchy in a tree
  • Cites a paper which implements a HT on LeanStore. Might be interesting for spilling state during a join

Any other languages/frameworks/topics you think we should look at? 202308131514
Status: #idea
Tags: OS

Computer Organisation Architectures

Von Neumann Architecture

• Also known as stored program concept
• Has a unified cache

Harvard Architecture

• Cache is split into instruction cache and data cache
• Advantages
  • Data and instructions can be pre-fetched independently
  • Less overhead to enforce protection mechanisms for data
• Problem: The statically split cache, leads to wastage of space

Modified Harvard

• L2 cache is unified, but L1 is still split
• More efficient use of space than Harvard.
• Used by almost all modern systems
• Allows for code to be modified during run-time
  1. Load the instruction as data into L1-data, and modify it
  2. Write the modified instruction from L1-data to L2
  3. Write the instruction from L2 to L1-code
     

References

1. Introduction to OS

LeanStore

• Describes how the pre-cursor Umbra’s buffer manager works
• Slight modifications result in Umbra’s implementation. Crux of the implementation is in this paper.
• Background contains some other hybrid buffer managers
  • Can be looked into if needed.
• Each optimisation technique described can be used individually
  • Pointer swizzling
  • Page replacement policy
    • Randomly set 10% of pages as cooling
    • Evict once they reach end of FIFO queue
      • Used pages, move out of queue
    • 
  • Latching
    • Better implementation is present in Umbra. Use this only if ref is needed
• Data structures needed for keeping track of state of buffer manager
  • Described well in LeanStore. Not present in Umbra
• Buffer-managed data structures
  • Talks about owning swips and the swip hierarchy in a tree
  • Cites a paper which implements a HT on LeanStore. Might be interesting for spilling state during a join