Lock Reservation: Java Locks Can Mostly Do Without Atomic Operations Kawachiya, Koseki,

1
Lock Reservation Java Locks Can Mostly Do
Without Atomic OperationsKawachiya, Koseki,
OnoderaIBM Research, Tokyo Research Lab,
JapanOOPSLA 2002

• Jon Steelman
• cs8803g
• 26 November 2002

2
What is paper seeking to accomplish?

• Increase in threaded performance
• Why important?
• Built-in support for multi-threaded programming
• Java programs perform many lock operations
• Atomic operations are very expensive

3
What is being exploited?

• Thread locality
• How is thread locality being exploited?
• Novel algorithm called lock reservation

4
Thread Locality- What is it?

• Sequence of threads in temporal order that
  acquire the lock.
• Dominant locker long repetition of a specific
  thread

5
Exploitable Thread Locality of Java Locks
Even for multi-threaded programs, more than 75
of lock operations were performed by the initial
lockers in first repetition.
6
Lock Reservation

• Exploits Thread Locality and reserves locks for
  threads
• Algorithm that can built on any existing locking
  algorithm that uses a lockword and allows use of
  one bit in the lockword, q.v. the fast Tasuki
  lock used in IBMs jdk 1.3.1
• Bit used for to represent lock reservation status
  and named LRV

7
Lock Reservation

• If objects lock reserved for thread, thread
  acquires lock with a few instructions and no
  atomic operations.
• If lock reserved for another thread, reservation
  cancelled and falls back to conventional locking
  algorithm
• If objects lock not reserved, system uses the
  conventional locking algorithm
• When LRV is set, lockword in reserve mode and
  lockword structure defined by the lock
  reservation algorithm. Otherwise, lockword is in
  base mode.

8
Lockword Structure
LRV bit
Thread ID
Recursion Cnt
1
(defined by base locking algorithm)
0
Lockword semantics in reserve mode
A
0
1
Reserved for Thread A, but not held
A
gt0
1
Reserved for and held by Thread A
0
0
1
Reserved anonymously
9
Lock State Transitions
10
Notes on State

• Owner thread can read-modify-write the lockword
  without atomic operations
• Once reservation is cancelled, lockword will
  never be reserved again
• Cancellation is most expensive operation but
  ratio to lock operations is less than 0.05 in
  actual lock-intensive programs

11
Performance Improvements
12
Lock Statistics
13
Observations

• Significantly improves performance if reservation
  succeeds while maintaining comparable performance
  if reservation fails
• Since lock reservation is a runtime technique, it
  is complementary to compiler optimizations like
  escape analysis and recursive lock elimination

14
Future Extensions

• Lowering cost of cancellations
• Avoiding cancellations, for example, by starting
  in base mode and avoiding reserve mode
• Dynamic profiling
• Static analysis
• When cancellation cost lowered, pursuing an
  algorithm enhancement allowing repeated lock
  reservations on an object