Transactional Collection Classes.ppt

TCC Lecture 7 EE 313: Transactional Collection Classes Brian D. Carlstrom, Austen McDonald, Michael Carbin Christos Kozyrakis, Kunle Olukotun Computer Systems Laboratory Stanford University Transactional Memory Promise of Transactional Memory (TM) Make parallel programming easier Better performance through concurrent execution How does TM make parallel programming easier? Program with large atomic regions Keep the performance of fine-grained locking Transactional Collection Classes Transactional versions of Map, SortedMap, Queue, … Avoid unnecessary data dependency violations Provide scalability while allowing access to shared data Evaluating Transactional Memory Past evaluations Convert fine-grained locks to fine-grained transactions Convert barrier style applications with little communication Past results TM can compete given similar programmer effort What happens when we use longer transactions? TM hash table micro-benchmark comparison Old: Many short transactions that each do only one Map operation TM SPECjbb2000 benchmark comparison Old: Measures JVM scalability, but app rarely has communication 1 thread per warehouse, 1% inter-warehouse transactions Unwanted data dependencies limit scaling Data structure bookkeeping causing serialization Frequent HashMap and TreeMap violations updating size and modification counts With short transactions Enough parallelism from operations that do not conflict to make up for the ones that do conflict With long transactions Too much lost work from conflicting operations How can we eliminate unwanted dependencies? Reducing unwanted dependencies Custom hash table Don’t need size or modCount? Build stripped down Map Disadvantage: Do not want to custom build data structures Open-nested transactions Allows a child transaction to commit before parent Disadvantage: Lose transactional atomicity Segmented hash tables Use ConcurrentHashMap (or similar approaches) Compiler and Runtime Support for Efficient STM, Intel, PLDI 2006 Di