This preview shows page 1. Sign up to view the full content.
Unformatted text preview: hing Ti writes and Ti ends
reading before Tj ends reading v r w v w • 3) Tj reads & writes nothing Ti writes and Ti ends
reading before Tj ends reading
w – Overlap of Ti copyback with Tj reads because Ti can’t
hurt Tj, but Tj might hurt Ti if its writes start too soon Simple testing of sets More parallelism • Long critical section if copyback slow • Test clearly preceding (no longer changing)
transactions outside critical section Lock Unlock Even more parallelism
• Add ordered testing of condition 3 Space issues
• What if run out of space for sets?
– Abort & retry • What about repeated abort & retry?
– Hold critical section in a retry (ugh) Applic to B-trees
• Models B-trees with lots of entries in each page
(199), uniform key insertion, interior nodes
cacheable and leaf pages not cacheable
• Does consider splitting a leaf, but apparently not
rotating tree to maintain balance (not needed if
inserts are uniform :-)
• Concludes for such B-trees that conﬂict, abort and
restart will be rare (0.07%) Eval
• Thought experiment
– Not fair as really a database theory paper then • Analysis of B-tree not appropriate
– Really should have modeled rotations as “randomness”
is far too unlikely • Very inﬂuential: “optimistic methods” is current
label of anything trying ﬁrst then checking if it
had a conﬂict and undoing it Lock-based Concurrency Control
lookup • U...
View Full Document
- Spring '12