Jan 19/22 buffer management

Memory management in DBs.

• As databases designer, you don't trust the os.
• Includes Memory management
• So we generally implement by ourselves (garbage collection, no malloc, new and free, delete --expensive)

Typically on startup.

• DB allocate huge area of RAM. (GBs.)
• Take and inserts all the ram into the systems buffer managers as a set of fixed size pages.

What are these pages used for?

• speed i/o (why reread rewrite to the disk if used very often).
• Temp storage for running data proc. algorithms.

Question: should buffer cache and scratch pool be seperate?

• pros: perhaps use storage more efficiently (keeping them together using the same set of ram)
• cons: the system must ensure one side not being starved for RAM.

In homework, we are going to put them together. Our System has no separation to get RAM, can call:

• getHandle(DBtable, pagenum) -- for buffer cache (handle is more like a pointer)

• getHandle(). -- for scratch

Key thing: DB working set often large than all available ram.

The question: how to handle this?

• page data in and out. (write out some data to disk to free some space)
• should be transparent to user! (the low level system should be automatically deal with this.)

So, what do we do when we "page data"?

1. We choose an existing (used) page to evict (move out of the dbms).
2. If the page is dirty (data havs been changed, but not written back to disk). write its contents to disk.
3. (fill contents of page (if necessary) first) -- Give a handle to the page back to requester

What are common eviction policies?

• LRU (can be challenge to implement)
  • associate a timestamp with each page.
  • whenever a page is accessed, set timestamp to current CLOCK value (basically a counter); inc clock
  • when need to evict pages, choose page with smallest timestamp

You have to index the pages in RAM via (several ways):

1. TS- timestampe (you don't want to scan all pages) --- really bad considering updating (inserting/deleting) the data structure (priority queue for example)
2. By file, pagenum
3. via a "super fast" pointer or handle. --- if some one called getHandle, they want ram which they can get easy access to.

Clock algorithm

• Attempts to address slowness of updates of TS

• idea: logically organize pages around the of a "clock"

• Clock has a second hand pointing at "current" page.

• Each page has a "DNE" (do-not evict -- I am not allowed to evict that page.) bit

• When request eviction, sweep second hand until find page with DNE='false'

• whenever the second hand passes page, set DNE = "false" (but not evict it)

• Whenever access a page, set DNE = "true"

• generally a good approximation of LRU but much easier to implement.

• all you have to do is to change a bit (while in LRU delete and reinsert in a data structure which is expensive).

Jan 22 Monday

LRU:
good: temporal locality, its "optimal"
bad:

• expensive; simplest data structure to implement is priority queue.
• classical failure case:
  1. vulnerable to attack
  2. vulnerable to real-life access pattern: repeated file scan. -- 100 pages to read in ram, 98 pages of ram available, will REWRITE the first two pages when reading.
  -- solved by allowing pinning pages, will only have first 98 pages cached in RAM, the left 2 pages will never being read.

clock
good: approximate LRU, less cost
bad: eviction may require full sweep of arm (but is it bad? you are not going to pay this cost unless you make eviction)

homework related:

smart-pointer: nowadays good coding styles, use smart pointers instead of raw pointer. For homework, will need the raw pointers e.g for page.;