Page 1 of 43

1
Nomad Distributed Programming

• A project by
• Morten D. Jørgensen
• Uffe Sørensen
• Claus R. Thrane
• Martin Clemmensen
• Anders Buch
• Thomas L. Kjeldsen

2
Agenda

• Introduction (Uffe)
• Motivation (Uffe)
• The Nomad Approach (Claus)
• Chords (Morten)
• Dist and Task (Morten)
• Architecture (Martin)
• Compiler (Anders)
• Demo (Claus)
• Performance (Thomas)
• Conclusion (Thomas)

3
Problem Statement

• Processing and time requirements
• Distribution
• Scientific computing
• Traditional languages
• Higher abstraction of distribution
• Rapid development

4
Environment

• Foundation
• RMI
• Functionality
• Tuple-Space
• Shared variables
• Task
• Dist

5
Base of Assumption

• Scientists
• Rapid development
• High level of abstraction
• Focus on the problem

6
Motivation

• Java
• Widely used
• Relatively simple syntax
• Cross-platform
• No built-in support for distribution
• C
• Mostly like Java
• More full-featured than Java
• Not stable cross-platform

7
Motivation

• C?
• New model of concurrency
• More intuitive code
• Chords over monitors
• Ada
• Comprehensive language
• Parallel processing (Task)
• Rendezvous between tasks

8
Motivation

• Linda
• Tuple-Space
• Occam
• Low-level constructs

9
The Nomad Approach

• Communication
• Synchronization
• Distribution

10
Multi Computer
11
Communication
Parallel programming models
Adopted in Nomad

• Multi-computer Shared variables
• Tuple-Space

• Shared Variables
• Message passing
• Parallel Objects
• Tuple-Space
• Data-parallel

12
Synchronization

• Threads and processes
• Synchronization mechanisms

• Async (C?)
• Chords (C?)
• Tuple-Space (Linda)

public async demo()
public int take() signal put(int i)
in(myTupl, i, ? j)
13
Distribution
task
dist

• Method inspired
• Shared parameters

• Class inspired
• methods
• Fields (a state)

14
Chords
15
Motivation

• Make intentions more apparent
• Conditions expressed more clearly in chords
  without technical details
• Compiler optimization
• Local and distributed synchronization

16
The Chord

• Lives in classes, like traditional methods
• One traditional method, one or more signals
• Synchronization patterns
• When pending method calls match a pattern, its
  body runs
• If there is no match, the invocations are queued
  up
• If there are several matches, a pattern is
  non-deterministically selected
• Async chords run body in a new thread
• Dist chords run body in a new thread on the
  network

17
Properties

• Equivalent to monitors
• OOP and chords
• No upper bound on invocations

18
Compiler Translation

• Introduces queues for holding (and blocking) the
  threads of methods and for arguments for signals
• Generate code for each method
• Check for matching chords. If match is found
  dequeue signals arguments - invoke scan() -
  execute body - return
• Enqueue (and block) thread
• Generate code for each signal
• Enqueue arguments
• Invoke scan() if signal invocation changed state
• Generate code for scan() method
• Iterate over all chords and wake up thread if
  match is found
• Uses bitmasks to look for matches

19
Example

• Reader-writer lock
• Only one writer blocks other writers and readers
  until release
• Unlimited readers when no writers
• Fair
• A write-request blocks new read-requests until
  writer is given exclusive access

20
Example - Nomad

• class NomadReaderWriter
• NomadReaderWriter()
• free()
• public void Shared() private signal free()
• isShared(1)
• public void Shared() private signal
  isShared(int n)
• isShared(n 1)
• public void ReleaseShared() private signal
  isShared(int n)
• if (n 1) free() else isShared(n - 1)
• public void ReleaseShared() private signal
  exclusiveWaiting(int n)
• if (n 1) freeExclusive() else
  exclusiveWaiting(n - 1)
• public void Exclusive() private signal free()
  
• public void Exclusive() private signal
  isShared(int n)
• exclusiveWaiting(n) wait()

21
Example - Java

• class JavaReaderWriter
• private int isShared 0
• private boolean writing false
• private int waitingWriters 0
• public synchronized void Shared()
• while (writing waitingWriters gt 0) wait()
  
• isShared
• public synchronized void ReleaseShared()
• isShared--
• if (isShared 0) notifyAll()
• public synchronized void Exclusive()
• waitingWriters
• while (isShared gt 0 writing) wait()
• waitingWriters--
• writing true

22
Task Construct

• Entrypoint method
• wait() and active()
• Supports async, dist and normal methods

23
Dist vs. Task

• Similarities
• Distributed
• Parameters
• Explicit synchronization

• Differences
• Methods
• Uses
• ?? Dist SETI_at_home
• ?? Task Server
• Easy to model Dist (and other models) using Task

24
Architecture

• Topology
• Implementation
• Scheduler
• Tuple-Space
• Shared variables
• Using the RTE

25
Topology

• Worker
• RMI registry server
• Executes jobs from its queue
• Server
• RMI registry server
• Scheduler
• Shared variables
• Tuple-Space
• Runner (user program)

26
Implementation

• Scheduler
• Maintains a list of workers
• Synchronous calls
• Returns when the job has been assigned to a
  worker
• Worker choice is round robin-style
• Adding and removing workers can be done at
  runtime (when not executing jobs)
• Alternate worker choice can easily be implemented

27
Implementation

• Tuple-Space
• Maintains a list of tuples (which are also lists)
• Synchronous calls
• Linear search time
• Blocks if a desired tuple is not found

28
Implementation

• Shared Variables
• Synchronous calls
• Returns a remote reference (stub)
• Garbage collection provided by the JVM.

29
Using the RTE

• Creating and submitting Jobs for execution
• Implementing the IJob interface
• Submitting via the SchedulerAccess class
• Using Shared Variables and the Tuple-Space
• Submitting desired shared objects via. the
  SharedMemAccess class.
• Adding and getting tuples from the Tuple-Space
  via. the TupleSpaceAccess class.

30
The Compiler

• Syntactic analysis / SableCC
• Contextual analysis
• Code-generation
• Usage of the Environment

31
Compiler Overview
32
Accessing the Environment
33
Accessing the Environment

• Tuple space out
• public void caseATsOut(ATsOut node)
• int listname _varName
• createAJasminStmt(1,0,
• "new java/util/ArrayList\n"
• "dup\n"
• "invokenonvirtual
  java/util/ArrayList/ltinitgt()V\n"
• "astore " listname "\n"
• ).apply(this)
• AListExp exps ( AListExp )
  node.getListExp()
• write("aload "listname"\n")
• doAutoBoxing(exps.getExp())
• write("invokevirtual java/util/ArrayList/add
  (Ljava/lang/Object)Z\n")
• write("pop\n")
• Object list exps.getListExpTail().toArra
  y()
• for (int i 0 i lt list.length i)

34
Performance

• Test setup and results
• Tuple-Space
• Distributed methods
• Improvements

35
Tuple-Space

• Method
• Put and take a tuple is considered one operation
• Tuple-Space and test program on separate hosts
• Average of 5 tests
• Test setup
• Server
• AMD2600, 512Mb Ram
• Linux 2.6, Sun JVM 1.5
• User program
• Intel P3, 1Ghz, 384Mb Ram
• Linux 2.4 kernel, Sun JVM 1.5

36
Tuple-Space Overview
37
Tuple-Space Details
38
Distributed Methods

• Method
• Dist method does nothing
• Average of 5 tests
• 1 to 4 workers
• Test setup
• Server
• AMD2600, 512Mb Ram
• Linux 2.6, Sun JVM 1.5
• Workers
• Private PCs ranging from 1 to 2.6Ghz

39
Distributed Methods Overview
40
Distributed Methods Details
41
Improvements

• Hash elements in the Tuple-Space
• Scheduling based on worker load

42
Conclusion

• Building the nomad compiler
• Language features
• Chords (new and interesting approach)
• Dist methods (idea mixing with chords)
• Tuple-Space (intuitive and handy)
• The objective of Nomad

43
Demonstration
Chords Tuple-Space Dist