A Technique for Parallel Reachability Analysis of Java Programs

1
A Technique for Parallel Reachability Analysis of
Java Programs

• Raghuraman R.
• Sridhar Iyer
• G. Sajith

2
Outline of Talk

• Errors in Concurrent Programs
• Reachability Analysis
• A Technique for Parallel RA
• Experimental Results
• Conclusions

3
Errors in Concurrent Programs

• Non-determinism and synchronization involved
• Data access errors (violation of mutex)
• Synchronization errors (deadlocks- waiting for
  shared variables, mesgs)
• Temporal errors (calls to terminated pgm
  components)

4
Reachability Analysis

• Systematic enumeration of all possible
• global states of program execution
• Technique
• Program listing -gt
• Generate abstract representation -gt
• Generate reachability graph -gt
• Analyse for errors -gt
• Error listing

5
RA (contd.)
6
Inefficiencies in RA

• - Exponential complexity (NT)
• - Limited capacity to deal with dynamic
  references
• - Might report spurious errors
• State Explosion tackled by Apportioning
  Technique Iyer 98
• - But time complexity remains

7
The Parallel RA Technique

• Mitigates the time complexity
• Generate state-transition graphs for each thread
• Generate cross-product of all state-transition
  graphs
• Discard invalid tuples from the cross-product

8
System Model

• All threads created at start of program execution
  (each thread at ltentrygt state)
• All threads have the same priority
• Program terminates when all threads reach their
  ltreturngt state
• Synchronization between threads is by using locks

9
Algorithm
10
An Example
Consider an example application with 2 threads
of execution.
Thread 1
Thread 2
11
Example (contd.)

• Construct state-transition graph for each thread
  Ti (all other threads at their entry points) in
  parallel

12
Example (contd.)

• Construct RG as cross-product of all Ris.
• - RG might contain tuples corresponding to
  multiple threads holding the same lock (RG an
  arbitrary cross-product) eg. Tuple 2

13
Example (contd.)

• Detect invalid tuples and discard them
• Remaining tuples correspond to traditional RG

14
Experimental Results
15
Experimental Results (Contd.)
16
Conclusions

• The technique is safe for all programs in the
  chosen model
• Significant reduction in time
• from O(NT)2 to O( (T.K.NT) / P)
• Where Tno. of threads, Nno of states in each
  thread, K no. of locks and Pno. of processors.