Distributed Process Network Deadlock Detection

1
Distributed Process Network Deadlock Detection
Resolution

• Alex Olson
• Embedded Software
• Spring 2004

2
1. What is Process Networks?

• Process Network (PN) Kahn74
• Concurrent processes communicate only over
  one-way channels (FIFO queues)
• Reads block until enough data is available
• Determinate model of computation
• Queues are unbounded in size
• Program deadlocks only at termination

PN models functional parallelism
3
2. Why make PNs Distributed?

• Why make PNs distributed?
• Multiprocessor (SMP) desktops are expensive, have
  limited number of processors ( 16 )
• SMP is overkill for PN coarse-grain parallelism
  (dont need shared memory)
• PNs are inherently parallelizable
• Perhaps one computer is specialized for
  computation, another specialized for 3D
  visualization, etc

4
3. Implementation of Distributed Process
Networks (DPN)

• Is it Easy?
• Deadlock Detection(Channel is now distributed)
• Load Balancing(Need to consider what subset of
  all processes to put on each sever)
• Dynamic Process Migration(Addition/Deletion of
  Servers)
• Objectives
• Create a high-performance DPN framework
  compatible with the above
• Implement Deadlock Detection

5
3. Extensions of Kahns PN

• Parks 95
• Sets capacity of each arc
• Writes to full queues (artificially) block
• On global deadlock, resize smallest queue
• Prefers incomplete bounded over complete
  unbounded execution
• Is determinacy of model compromised?
• Geilen Basten 03
• Prefers complete unbounded over incomplete
  unbounded
• Solves local deadlock detection/resolution

6
4. PN DPN Implementations

• All PN Implementations
• Map processes onto threads
• Rely on shared memory for deadlock detection
• All DPN Implementations
• Map processes onto threads, multiple servers
  over a network
• Only a few implementations exist
• None do deadlock detection!

7
5. DPN Deadlock Detection

• Network channel more complex than shared-memory
  queue
• Network latency
• Causes different views of channel
• Indeterminate capacity
• How much data is in transit?
• Send / Receive Buffers
• Even TCP doesnt completely guarantee delivery!

8
6. Conclusion

• Distributing PNs has cost performance benefits
• New problems arise in DPNs that didnt exist in
  regular PNs.
• No known DPN framework currently detects
  deadlocks.
• Implementation Plan
• Create simple, high-performance framework using
  Java or C
• Use a distributed global snapshot algorithm for
  deadlock detection
• Tame the channel by counting and acknowledging
  tokens (on top of TCP)

9
7. Future Plans

• Evaluate deadlock detection overhead
• Examine queue sizes vs performance
• Compare DPN vs. PN performance
• Implement Dynamic Load Balancing
• Any questions?

10
8. References
Kahn 74 G. Kahn. "The Semantics of a Simple
Language for Parallel Programming." In J.L.
Rosenfeld, editor, Information Processing 74,
Proceedings, pages 471475, Stockholm, Sweden,
August 1974. North-Holland, Amsterdam, The
Netherlands, 1974. Parks 95 T.M. Parks.
"Bounded Scheduling of Process Networks." PhD
thesis, University of California, EECS Dept.,
Berkeley, CA, December 1995. Technical Memorandum
UCB/ERL M95/105. Geilen 03 M.C.W. Geilen and
T. Basten. "Requirements on the Execution of Kahn
Process Networks." Programming Languages and
Systems, ESOP 2003, Warsaw, Poland, April 7-11,
2003, Proceedings. Lecture Notes in Computer
Science. Berlin, Germany, 2003