Partha Dasgupta Arizona State University The MILAN Project

2. Parallel Processing with Windows NT Networks • Partha DasguptaArizona State University • The MILAN Project • New York University Arizona State University • Funding Sources: DARPA/Rome Laboratory, NSF, Intel, and Microsoft Collborators: Zvi M. Kedem Donald McLaughlin Shantanu Sardesai Rahul Thombre

3. ECLIPSE CalypsoLinux 1.0 Malaxis Chime + Joint Research ofArizona State University and New York University

4. The Platforms • Calypso • Language independent parallel processing • Shared memory and fault tolerance. • Chime • CC++ based parallel processing • Shared memory, fault tolerance • Malaxis • DSM package for Windows NT • Read/write locking, barriers • Milan • A metacomputing platform • Coalesces features from the above systems to a general purpose computing platform

5. Unix to Windows NT • Port a system program or middleware from Unix to Windows NT. How? • Just change the system calls? • Does not work. • Change programming and design styles to NT-centric: • no signals in NT • use structured event handling (no such thing in Unix) • use threads (useful) • integrate with windows messages or MFC • remote execution support is weak • Learn NT-centrism, and NT lingo

6. NT Terminology • MSDN is not a network • Developer’s library contains books • Resource Kit is not about resources • Huh? • SDK, DDK, checked build • Service Pack • OSR2 • Remote access does not let you execute anything remotely • Use a Share? • You mean remote mount? No, I mean map network drive • Memory can be reserved or committed or both. • Synchronization primitives - never mind...

7. What is • Yet another parallel processing system, which runs on a distributed network of microcomputers: • Shared Memory • Novel execution and memory management strategy • Fault Tolerant: • Machines may stop and start dynamically without affecting the execution • Automatic Load Balancing: • Manages slow and fast machines Provides near optimal thread assignments (measured) • Execution strategy hidden from programmer: • No message passing, process management, data partitioning • Low-overhead mechanisms

8. manager Eager Scheduling TIES worker worker Key Techniques in Calypso • Eager Scheduling • Manager - worker architecture • Provides fault-tolerant and load-shared executions with minimal overhead • Two-phase Idempotent Execution Strategy • Distributed memory management strategy • Stops side effects due to failures • Ensures idempotence of results, in spite of duplicate executions • These techniques developed in previous joint theoretical research

9. A 9 1 3 5 8 10 12 B 2 6 11 9 Worker interrupted C 4 7 9 Worker crashed time 50 100 150 200 250 300 350 400 Eager Scheduling • Workers contact the manager for work after finishing previous assignment, if any • When there is unfinished work, the manager has the option of assigning an unfinished thread to a “willing” worker regardless of who is already working on that thread • An example of Round Robin Eager Scheduling: • 3 machines: fast, slow and transient • 12 threads of equal length (50 secs)

10. Chime • Chime is a programming system and runtime environment for parallel processing • The first system to incorporate standard parallel language support on a network of workstations: • Nested Parallelism, Parallel statements • Language-defined scoping of variables • Synchronization support • Transparent shared memory • Chime supports the “shared memory” constructs of CC++ • Adds fault tolerance…. • Adds load balancing…. …. with low overhead A “distributed” cactus stack

11. Chime Software Architecture

12. Chime Execution Trace

13. Malaxis • A DSM Package • Uses NT threads and memory mapping and protection features • Uses barrier synchronization, memory XOR-ing and intelligent monitoring of page/lock requests to prevent page shuttling • Programmer support: • Spawning processes on remote machines • Mapping shared segments • Barrier Synchronization • Read and Write locks (abstract, advisory)

14. Milan • A metacomputing platform • Creates a system image of a large computer on a set of workstations • Smart scheduling • bunching • job recall • pre-emption • Shared memory • Fault tolerant

15. Using Windows NT • The needs of our implementations: • User Level page fault handling • Getting and setting thread contexts • Getting and setting stack contents • Asynchronous notification and exception handling • Networking support • Process/Thread control • Windows NT provides all of the above

16. Memory Handling • Windows NT memory handling is elegant and powerful (After you understand the terminology) • States of memory: • committed • reserved • guarded • Protection and allocation is done by: • VirtualAlloc • VirtualProtect • Access violations generate exceptions • Needed reprogramming Calypso - for the better

17. Exception Handling • All exceptions are delivered to an exception handler, defined in the current scope of execution. • Great, for programmers - nice and structured • Not good for middleware solutions…. • How can I execute another persons code, with my exception handlers? • I cannot change the exception handler, from within my exception handler. • In our case, we found reasonable workarounds - but don’t have general solutions to the above problems.

18. Threads • Good, consistent, kernel threads. • Easy to use • works great • plethora of synchronization constructs (too many, in fact) • Threads are useful for: • Threads inside middleware - wow! • Handling distributed shared memory (callbacks, caching, memory service) • Process migration - a thread can set up the main process • Segregating functionality (assign a thread per job)

19. Process and Stack Migration • Migration is used by our system for several purposes: • Cactus stacks • Checkpointing • Pre-emptive scheduling (produces better turnaround times in dynamic environments) • When a thread has to be migrated: • Another thread suspends it and gets its context • The context is a checkpoint • The context is sent to the target machine • A thread sets the context of a suspended thread with the new context and resumes it. Stack has to be reset too. • IT WORKS

20. Other Features • Networking • winsock is like sockets, no surprises • Remote execution • our approach: Use a daemon process • NT approach: use a starter service • Execution Monitor (GUI) • External process, that controls and displays state of the distributed computation

21. Performance • Program: Ray Trace, generates a nice picture • Equipment: Pentium-90, running Windows NT (Calypso tests)Pentium Pro 200, running Windows NT (Chime tests) • Tests conducted • Speedup • Speedup in case of mixed speed machines • Speedup in case of crashing and recovering machines • Micro-tests (migration, stack creation) • Not all tests will be shown now.

22. Calypso Performance Performance is comparable to Unix systems

23. Chime Performance Chime has higher network overhead than Calypso

24. In Retrospect • NT has some strong points, things that are better than Unix • Threads • Exception Handling • Memory Management • Program development tools • (very good, especially the debugger) • Documentation • A few shortcomings • no signals • no remote execution facility • terrible terminology

25. Status • Operational prototype systems • Calypso on Windows NT / Windows 95 released • A prototype of Chime implementing most of the “parallel part” of Compositional C++ on an unreliable network of workstations • Ongoing research • Distributed scheduling and resource management (for MILAN) • Quality of service • Better integration with NT (MFC support, remote services, global scheduling…)

26. Acknowledgements • Co-PI • Zvi M. Kedem • Calypso • Arash Baratloo, Mehmet Karaul • Calypso NT • Donald McLaughlin and Shantanu Sardesai • Chime • Shantanu Sardesai • Calypso Linux • Arash Baratloo

27. Questions ?

28. done?

29. Done? Review request for SP&E