Efficient Computation of May-Happen-in-Parallel Information for Concurrent Java Programs

1
Efficient Computation of May-Happen-in-Parallel
Information for Concurrent Java Programs

• Rajkishore Barik
• rajbarik_at_in.ibm.com

2
Introduction

• May-happen-in-parallel (MHP) analysis
• Statically determines statements that may
  execute in parallel in a multithreaded program
• Application
• Detecting synchronization anomalies
• Program optimization
• Debuggers and program understanding tools
• Data flow analysis
• Enforce ordering among shared accesses
• Complexity
• NP-complete (Taylor)

3
Related Work

• B4 analysis (Callahan and Subhlok)
• Ada
• Interprocedural B4 analysis (Duesterwald and
  Soffa)
• Non-concurrency analysis (Masticola and Ryder)
• May-Happen-in-Parallel analysis (Naumovich and
  Avrunin)
• Java
• May-Happen-in-Parallel analysis (Naumovich and
  Avrunin)
• Data flow analysis
• Constructs start, join, notify, wait, notifyAll,
  locks
• Number of runtime threads need to be explicitly
  enumerated during static analysis
• Worst-case complexity (pN)3 where p is the
  number of runtime threads and N is the number of
  PEG nodes per thread
• Happens-before, happens-after and
  happens-in-parallel (Sura et al.)
• Data flow analysis
• Constructs start, join, locks

4
Example

• 1 class Shared int field0
• 2 class Main
• 3 static Shared s
• 4 public static void main(String args)
• 5 s new Shared()
• 6 s.field
• 7 Thread ta new Thread10
• 8 for(int i0ilt10i)
• 9 tai new Task1()
• 10 tai.start()
• 11
• 12 s.field
• 13 Thread t2 new Task1()
• 14 t2.start()
• 15 s.field
• 16
• 17
• 18 class Task1 extends Thread

5
Thread Model

• Abstract Thread (ti)
• Compile-time entity that corresponds to
  invocation of Threadstart in a context
• isUniqueti holds if ti has unique runtime
  correspondence
• Intra-thread control flow graph,
• ICFG(ti) ltV(ti), E(ti)gt
• V(ti) USE(ti) 4 ASS(ti) 4 NEW(ti)
• 4 BEGIN(ti) 4 END(ti ) 4 ENTRY(ti )
• 4 EXIT(ti) 4 CSTART(ti) 4 CJOIN(ti)
• 4 CALL(ti) 4 ACQUIRE(ti) 4 RELEASE(ti)
• E(ti) contains both intra and inter procedural
  control flow edges
• locksvi denotes the set of objects that are
  locked while executing vi cV(ti)

6
Program Representation

• Must-join abstract thread (tj)
• CJOIN(ti,tj) postdom CSTART(tk,tj) and ti tk
• Thread Creation Tree (TCT)
• Encodes start relationship among abstract threads
• Nodes are abstract threads Edges represent
  thread creation
• Nodes in TCT are colored black if it is not a
  must-join abstract thread
• yca(ti,tj) denotes the youngest common ancestor
  of ti and tj in TCT
• canc(ti,tj) denotes the child of ti that is an
  ancestor of tj

7
MHP Computation

• MHP Computation
• Thread-level MHP (Èt)
• Node-level MHP (Èn)
• Thread-level MHP
• ti Èt tj true, if ti c anc(tj ) or tj c
  anc(ti )

8
MHP Computation

9
MHP Computation

10
MHP Computation

11
MHP Computation

• Node-level MHP

12
MHP Computation

13
MHP Computation

14
MHP Computation
15
MHP Computation Algorithm

• 1. Identify abstract threads and their ICFGs.
• 2. for every abstract thread ti do
• 3. Compute postdom for every node in Vi.
• 4. Compute reachability information.
• 5. for every child abstract thread tj of ti do
• 6. Compute start-dominance information w.r.to
  tj.
• 7. end for
• 8. Add a node to TCT.
• 9. end for
• 10. Compute must-join chains and yca information.
• 11. for all abstract thread ti do
• 12 for all abstract thread tj do
• 13 for all vim cVi do
• 14 for all vjn cVj do
• 15 Determine vim Èvjn
• 16 end for
• 17 end for
• 18 end for
• 19 end for

16
Complexity

• Worst-case complexity is (kN)2, where
• k is the number of abstract threads and
• N is the number of ICFG nodes per thread

17
Experimental Setup

• Infrastructure
• ERCO (ETH Research COmpiler at ETH, Zurich)
• Pentium IV at 2.66GHz on RedHat Linux
• Benchmarks
• Java Grande Forum
• Philo, tsp, elevator, sor and mtrt
• Comparison
• Naumovich et al, our approach

18
Experimental Results
19
Experimental Results

20
Conclusion

• A new thread model based on context and flow
  sensitive analysis is presented.
• MHP computation is split into thread-level and
  node-level based on TCT and is obtained
  efficiently.
• TCT depicts interaction among threads and can
  perform various thread structure analysis.
• Our MHP algorithm is on an average 1.77x faster
  than Naumovich et al.