Sunday, July 18, 2010

Three ways to make an object thread-safe


There are basically three approaches you can take to make an object such as RGBThread thread-safe:

  1. Synchronize critical sections
  2. Make it immutable
  3. Use a thread-safe wrapper 
design techniques and thread safety

    Why worry about thread safety?

    Given the structure of the JVM, local variables, method parameters, and return values are inherently "thread-safe." But instance variables and class variables will only be thread-safe if you design your class appropriately

    -- Bill Venners

    Preventing Memery Locks on the Multi CPU machine

    Problem
       Multithreaded apps create new objects at the same time
       New objects are always created in the EDEN space
       During object creation, memory is locked
       On a multi CPU machine (threads run concurrently) there can be contention

    Solution
        Allow each thread to have a private piece of the EDEN
        space
    Thread Local Allocation Buffer
        -XX:+UseTLAB
        -XX:TLABSize=
        -XX:+ResizeTLAB
        (On by default on multi CPU machines and newer JDK)
    Analyse TLAB usage
        -XX:+PrintTLAB
    JDK 1.5 and higher (GC ergonomics)
        Dynamic sizing algorithm, tuned to each thread


    INSIDE THE JAVA VIRTUAL MACHINE / Filip Hanik

    Sun recommended JVM Settings

    GC Settings
       -XX:+UseConcMarkSweepGC
       -XX:+CMSIncrementalMode
       -XX:+CMSIncrementalPacing
       -XX:CMSIncrementalDutyCycleMin=0
       -XX:+CMSIncrementalDutyCycle=10
       -XX:+UseParNewGC
       -XX:+CMSPermGenSweepingEnabled
    To analyze what is going on
       -XX:+PrintGCDetails
       -XX:+PrintGCTimeStamps
       -XX:-TraceClassUnloading
                                       
    Minor Notes
    -XX:+UseParallelGC <> -XX:+UseParNewGC
    -XX:ParallelGCThreads=
        Use with ParallelGC setting
    If you have 4 cpus and 1 JVM
        Set value to 4
    If you have 4 cpus and 2 JVM
        Set value to 2
    If you have 4 cpus and 6 JVM
        Set value to 2


    INSIDE THE JAVA VIRTUAL MACHINE / Filip Hanik

    Thread Safety Rules

    Safety Rules

    Sometimes, a set of data is interdependent. For example, we might have two fields corresponding to a street address and a zip code. Changing an address might involve changing both of these fields. If the zip code is changed without a corresponding change to the street address, the data may be inconsistent or incoherent. Such a set of operations, which must be done as a unit -- i.e., either all of the operations are executed or none are -- in order to ensure consistency of the data, is called a transaction. The property of "doing all or none" is called atomicity. A system in which all transactions are atomic is transaction-safe.
    The following rule suffices to ensure that your system is transaction-safe:
      All (potentially changeable) shared data is accessed only through the synchronized methods of a single object; no interdependent piece can be accessed independently.
    Note that this means that shared data cannot be returned by these methods for access by other methods. If shared data is to be returned, a (non-shared) copy must be made. Further, if interdependent values are to be returned (i.e., a portion of the shared data is to be used by other methods), all of the relevant values must be returned in a single transaction.
    For example, the address and zip code of the previous example should not be returned by two separate method calls if they are to be assumed consistent.
    public class AddressData {
    private String streetAddress;
    private String zipCode;
    public AddressData( String streetAddress, String zipCode) {
        this.setAddress( streetAddress, zipCode );
        ....
    }
    public synchronized void setAddress( String streetAddress,
                                         String zipCode) {
        // validity checks
        ....
        // set fields
        ....
    }
    public synchronized String getStreetAddress() { // problematic!
        return this.streetAddress;
    }
    public synchronized String getZipCode() { // problematic!
        return this.zipCode;
    }
    }
    If this class definition were used, e.g. for
    printMailingLabel( address.getStreetAddress(),
                       address.getZipCode() );
    it would in principle be possible to get an inconsistent address. For example, between the calls to address.getStreetAddress() and address.getZipCode(), it is possible that a call to address.setAddress could occur. In this case, getStreetAddress would return the old street address, while getZipCode() would return the new zip code.
    Instead, getStreetAddress() and getZipCode() should be replaced by a single synchronized method which returns a copy of the fields of the AddressData object:
    public synchronized SimpleAddressData getAddress() {
        return new SimpleAddressData( this.streetAddress,
                                      this.zipCode );
    }
    The SimpleAddressData class can contain just public streetAddress and zipCode fields, without accessors. It is being used solely to return the two objects at the same time.

    Introduction to Interactive Programming by Lynn Andrea Stein