© 2010 IBM Corporation Concurrency Demystified Ajith Ramanath | Senior Developer | JTC-ISL

© 2010 IBM Corporation 2 What’s in store today?  Thinking Performance & Scalability  Erstwhile Java Synchronization  Tools & Diagnostics  Java Memory Model  Concurrency & Non-Blocking Synchronization  Concurrent Building Blocks in Java  A Simple Concurrent Data-structure  Overview of Project Amino  References

© 2010 IBM Corporation 3 Thinking Performance & Scalability…  What are multi-core systems?  Scaling UP (⇑) vs. Scaling OUT (⇔)  What are scalable programs?  Relevance of scalable software

© 2010 IBM Corporation 4 Erstwhile Java Synchronization  Java “synchronized” keyword  Object specific “mutex”  Applies to class and instance Objects  2 or 3 Tier lock strategy  Java semaphore implementation  Public wait() and notify() / notifyAll() methods  Accessible from all Java classes and objects  JSE specification provides semaphore via ‘ java.lang.Object ’ class  Lock contention & ensuing problems  Reduced parallelism & lower CPU utilization  Increased thread context switch time

© 2010 IBM Corporation 5 JLM Report Tools: Java Lock Monitor (JLM)  JVMPI/TI agent for profiling Java monitor entry and exit  Free Download from SourceForgeSourceForge  Available on Linux and Windows (x86_64 && IA32)  Simple to use and Light-weight Java Virtual Machine (JVM)JPROF rtdriver Socket JVMPI Enable / Disable Events Request Events JVM callback Textual commands Reports / Errors

© 2010 IBM Corporation 6 Sample JLM Output

© 2010 IBM Corporation 7  Timer based profiler – works using sampling technique  Tracks CPU ticks and maps them to kernel / application modules  Useful to investigate CPU over/under-utilization  Pick tprof ‘first’ during performance analysis! Tools: tprof JIT module taking the maximum clock ticks

© 2010 IBM Corporation 8 Java Memory Model  Java goofed up…before v5.0:  The ugly DCL paradigm ( A popular anti-pattern in Java!  )  (Im?)mutability  Compiler re-ordering / volatiles & non-volatile stores Immutability ProblemProblem with volatile/non-volatile storesDouble-checked Locking

© 2010 IBM Corporation 9 Concurrent Computing (to the rescue…) Definitions: Concurrent computing is a form of computing in which programs are designed as collections of interacting computational processes that may be executed in parallel A non-blocking algorithm ensures that threads competing for a shared resource do not have their execution indefinitely postponed by mutual exclusion.

© 2010 IBM Corporation10 Concurrent Building Blocks (CBBs) in Java  Atomic variables  Reentrant locks  Collection APIs  Semaphores  Latches  Cyclic Barriers  Future Semantics  Executors  Exchangers Since Java 50 (TIGER) release

© 2010 IBM Corporation11 Java Collection CBBs (A closer look…)  Queue Implementations:  BlockingQueue  Deque  ConcurrentLinkedQueue  Map Implementations:  ConcurrentHashMap  ConcurrentSkipListMap (since Java 60)  Set Implementations:  CopyOnWriteArraySet  ConcurrentSkipListSet (since Java 60)  List Implementations:  CopyOnWriteArrayList

© 2010 IBM Corporation12 A Simple Concurrent API (Treiber’s Algorithm) Class ConcurrentStack { static AtomicReference top = new AtomicReference (); public void push (Object item) { Node newHead = new Node(item); do { Node oldHead = top.get(); newHead.next = oldHead; } while (!top.compareAndSet(oldHead, newHead); } public Object pop() { Node newHead, oldHead; do { oldHead = top.get(); if (oldHead == null) return null; newHead = oldHead.next(); } while (!top.compareAndSet(oldHead, newHead); return oldHead.item; }

© 2010 IBM Corporation13 Project Amino  Building a scalable foundation for multicore programming  Multiplatform - Java and C++  Programming tools and libraries  Collaborative work between IBM research and IBM Software Labs  Available for download now on SouceForge  Open for your participation and contribution!

© 2010 IBM Corporation14 Amino Data Structures  Amino data- structures are designed to have lock-free or wait-free semantics  A lock-free operation allows optimistic concurrency, enables isolated thread failure and will not deadlock  A wait-free operation will always proceed without blocking  Focus on Java  Data structures complement Java’s JUC Stack Vector HashMap QueueDeque Dictionary Tree Priority Queue List Binary srch tree

© 2010 IBM Corporation15 References & Further Reading  Java Concurrency in Practice (by Brian Geotz – Pearson Education)  Oracle Java APIs Download siteOracle Java APIs Download site  High Performance Dynamic Lock-Free Hash Tables and List-Based SetsHigh Performance Dynamic Lock-Free Hash Tables and List-Based Sets  Scalable and Lock-Free Concurrent DictionariesScalable and Lock-Free Concurrent Dictionaries  Lock-free Dynamically Resizable ArraysLock-free Dynamically Resizable Arrays  Lock-free Resizable Hash TablesLock-free Resizable Hash Tables  Fast and Lock-Free Concurrent Priority Queues for Multi-Thread SystemsFast and Lock-Free Concurrent Priority Queues for Multi-Thread Systems  IBM DeveloperWorks series:  Java Theory and Practice: Fixing the Java Memory Model Part 1 and 2Part 12  Finally, use Wikipedia and Google Search Engine for more!

© 2010 IBM Corporation16 I’m done…are there (more?) questions?