just keep passing the messages
DESCRIPTION
Presentation at JAXLondon 2011TRANSCRIPT
Copyright © 2011 Russel Winder 1
Just Keep Passing the Messages
Dr Russel Winder
It’z Interactive [email protected]
@russel_winder
Copyright © 2011 Russel Winder 2
Just Keep Passing the Messages
Dr Russel Winder
Concertant [email protected]
@russel_winder
Copyright © 2011 Russel Winder 3
Just Keep Passing the Messages
Dr Russel Winder
Independent [email protected]
@russel_winder
Copyright © 2011 Russel Winder 4
Just Keep Passing the Messages
Prof Russel Winder
Wolverhampton [email protected]
@russel_winder
Copyright © 2011 Russel Winder 5
Aims and Objectives
● Investigate why message passing architectures are the software architectures of the future.
● Have some fun (whilst hopefully learning something).
Copyright © 2011 Russel Winder 6
Audience Aims and Objectives
?
Copyright © 2011 Russel Winder 7
Structure of the Session
● Introduction.
● Do stuff.
● Exit stage (left|right|front|back).
There may well be significantdynamic binding of the session.
Copyright © 2011 Russel Winder 8
Protocol
● Some slides, to kick things off.
● Some programs to really demonstrate things.
● NB Interaction between audience and presenter is probably mandatory.
We reserve the right to (shelve|stash) forlater any interaction that appears to goon longer than seems appropriate.
NB This is a 50min session!
Copyright © 2011 Russel Winder 9
Enter Manic Mode
Copyright © 2011 Russel Winder 10
BEGIN
● History is important:– To know today, we look to yesterday. To know tomorrow,
we see today as yesterday.
http://wiki.answers.com/Q/Why_history_is_important
–
Copyright © 2011 Russel Winder 11
One program at a time.
Multi-tasking
Processes / IPC
Threads / Shared memory
Processes / Threads Processes / RPC
Copyright © 2011 Russel Winder 12
Historical Summary
● Shared-memory multi-threading for applications programming is a total aberration:
– Consequences of operating systems handling of concurrency have been imposed on all programmers.
● Cooperating processes is where applications development sanity is:
– Operating system interprocess communication was slow, hence threads, but this lead directly to the shared-memory, multi-threading quagmire.
– Erlang has shown that processes and message passing can do the job properly even after the “mainstream” had rejected process-based working.
Copyright © 2011 Russel Winder 13
Concurrency vs.
Parallelism
Copyright © 2011 Russel Winder 14
Concurrency
● Running multiple tasks using time-division multiplexing on a single processor.
Copyright © 2011 Russel Winder 15
Parallelism
● Running multiple tasks concurrently.
Note that the word concurrency here is the Englishmeaning whereas the previous slide gave the computing jargon meaning of the word.
Copyright © 2011 Russel Winder 16
Concurrency . . .
● . . . is a technique required for operating systems.
● . . . can sometimes be required for applications but not as much as might be thought.
● Applications can use alternative models, for example event-driven systems.
– Abstract over the control flow rather than manage it with locks, semaphores, monitors, etc.
Copyright © 2011 Russel Winder 17
Parallelism . . .
● . . . is about making an algorithm execute more quickly in the presence of multiple processors.
● . . . is an architectural and design issue for applications.
Copyright © 2011 Russel Winder 18
Concurrency
● The problem with threads is shared memory:– Without writeable shared memory there is no need for
synchronization.
● Why impose a 1960s, operating system driven view of concurrency management on applications programmers?
Copyright © 2011 Russel Winder 19
Mutable Shared Memory is . . .
● . . . anathema to parallelism.
● . . . anathema to concurrency.
● . . . anathema to modern applications programming.
● . . . anathema.
Copyright © 2011 Russel Winder 20
Solution . . .
● . . . do not use mutable shared memory.
Note the caveat here that opens the doorto using shared immutable data.
Copyright © 2011 Russel Winder 21
Operating Systems . . .
● . . . are inherently a concurrency problem.
● Applications on the other hand are not, they should be using higher level abstractions.
Copyright © 2011 Russel Winder 22
ObjectOrientation
Copyright © 2011 Russel Winder 23
Original Model
● Object-based:– A set of (small) closed namespaces, with methods,
exchanging messages requesting services.
● Object-oriented:– Object-based plus classes and inheritance
Copyright © 2011 Russel Winder 24
Implementations
● Smalltalk realized the object-oriented model correctly.
● C++ did not: message passing replaced by function call.
C++ destroyed correct appreciation of the object-oriented model of computation.
Copyright © 2011 Russel Winder 25
Non-contentious (sort of) Syllogism
● Object-orientation is about objects passing messages to each other: object-orientation is not about function call in a shared memory context.
● C++ is a programming language based on function call in a shared memory context: C++ does not have objects passing messages to each other.
● C++ is not an object-oriented programming language.
Copyright © 2011 Russel Winder 26
Java
Copyright © 2011 Russel Winder 27
Object-orientation
● Java follows C++:– Function call replaces message passing.
● Java beats C++, had threads 15 years before C++.
● Shared memory multi-threading requires all the locks, semaphores, monitors, etc. and Java has it all.
Java is not an object-oriented language.
Copyright © 2011 Russel Winder 28
Partial Solution for JVM
● Use java.util.concurrent:– Provides many high-level tools.
– Has many low-level tools.
Use jsr166y (Fork/Join) and extra166y (ParallelArray) in preference to using stuff in JDK6.
If you are using the low-level tools then you are lost to the cause of quality application programming.
JDK7 has Fork/Join but not ParallelArray, still have to use extra166y for this – sadly means using jsr166y.
Copyright © 2011 Russel Winder 29
Interesting Aside
● Paul King at SpringOne2GX 2011:
“Who used synchronized?”
Some in audience raise their hands.
“You did it wrong!”
Copyright © 2011 Russel Winder 30
High Performance Computing
(HPC)
aka
Real Computing
Copyright © 2011 Russel Winder 31
Parallelism Rules
● HPC has been doing parallelism for 40+ years.
● Combinations of architectures:– Vector processors
– Multi-bus multiprocessors
– Clusters
Copyright © 2011 Russel Winder 32
HPC Extant Solution
● Message Passing Interface (MPI)
MPI addresses the problem of SPMD or MIMD parallelism in the context of multiple, possibly multicore, systems.
Copyright © 2011 Russel Winder 33
HPC Proposed Solution
● Partitioned Global Address Space (PGAS)
● Champions:– Chapel
– X10
– Fortress
Structure the global address space to allow for multiple processors sharing a single memory and/or to deal with distributed memory systems.
Copyright © 2011 Russel Winder 34
Return to a Better Way
Copyright © 2011 Russel Winder 35
Exit Manic Mode
Copyright © 2011 Russel Winder 36
Real Solutions?
● Actor Model
● Dataflow architectures
● CSP (Communicating Sequential Processes)
Return to a process and message passing view of applications.
Nothing wrong with threads as a tool.The problem is using shared memory.
Copyright © 2011 Russel Winder 37
Return to objects passing messages to each other.
Return to being object-oriented.
Copyright © 2011 Russel Winder 38
Actor Model
● A program is a collection of actors that send messages to each other.
● An actor is a process with a “mailbox” for receiving messages.
● A mailbox is a (thread-safe) queue of messages.
Copyright © 2011 Russel Winder 39
Dataflow Model
● A program is a graph of operators with some data sources and some data sinks.
● An operator is an event-triggered computation with some inputs and some output.
● An operator triggers for a certain state of its inputs.
Copyright © 2011 Russel Winder 40
CSP is . . .
● . . . mathematics:
● . . . but not scary since the mathematics can be hidden in an API, so it just becomes a programming tool.
VMS = X. ( in2p (chocolate X|out1p toffee X)μ → → → → | in1p (toffee X|in1p (chocolate X|in1p STOP X )))→ → → → → α
Copyright © 2011 Russel Winder 41
CSP
● A program is a graph of processes running a sequential computation that take input from input channels and write output to output channels.
● Data exchange down a channel realizes a rendezvous.
Copyright © 2011 Russel Winder 42
Commentary
● Actors are non-deterministic, with chaotic communications and hence complex.
● Dataflow and CSP have much greater determinism with fixed communications channels.
Copyright © 2011 Russel Winder 43
Implementations
● Actor Model:
– JVM: GPars, Scala, Akka, Fantom
– Native: C++/Just::Thread Pro, D
– Alternative: Erlang
● Dataflow Model:
– JVM: GPars, Pervasive DataRush
– Native: C++/Just::Thread Pro
● CSP:
– JVM: GPars, JCSP
– Native: C++CSP2
Copyright © 2011 Russel Winder 44
Acknowledgements
Copyright © 2011 Russel Winder 45
First Example Problem
● Something small, so the code is small.
● Something not too “toy”.
● Something with good parallelism.– Embarrassingly parallel to allow checking of scaling.
Copyright © 2011 Russel Winder 46
Copyright © 2011 Russel Winder 47
What is the Value of ?
● Easy, it's known exactly, it's (obviously).
It's simples†
† Aleksandr Orlov
Copyright © 2011 Russel Winder 48
Approximating
● What is it's value represented as a floating point number?
– We can only obtain an approximation.
– A plethora of possible algorithms to choose from, a popular one is to employ the following integral equation.
4=∫0
1 1
1x2dx
Copyright © 2011 Russel Winder 49
One Possible Algorithm
● Use quadrature to estimate the value of the integral – which is the area under the curve.
=4n∑i=1
n 1
1i−0.5n
2
With n = 3 not much to do, but potentially lots of error.
Embarrassingly parallel.
Copyright © 2011 Russel Winder 50
Major Minor Hardware Problem
● Multiple hyperthreads per core on multicore processors can be a serious waste of time.
Ed: Rant about chip manufacturers andoperating systems elided to avoid persecutionprosecution.
Copyright © 2011 Russel Winder 51
Second Example Problem
● Sleeping Barber Problem– A barber shop has a cutting chair and some waiting chairs.
The barber sleeps in the cutting chair if there are no customers. If a customer enters, the customer checks the cutting chair and wakes the barber if the barber is asleep in the chair, sits in the chair and gets a hair cut. If the entering customer sees the cutting chair in use, the customer checks to see if there is a waiting chair free. If there is the customer sits and waits, otherwise the customer leaves dissatisfied. On finishing a customer cut, the customer leaves satisfied (we assume), and the barber checks for waiting customers. If there are any waiting customers, the customer moves to the cutting chair. If there are no waiting customers, the barber returns to sleeping in the cutting chair.
Copyright © 2011 Russel Winder 52
Sleeping Barber Problem . . .
● . . . is an interesting recasting of a process synchronization problem in operating systems.
● . . . is due to Edsgar Dykstra
http://en.wikipedia.org/wiki/Sleeping_barber_problem
Copyright © 2011 Russel Winder 53
If the examples haven't been shown yet, now is the time to show them!
Copyright © 2011 Russel Winder 54
Summary
● Multiprocessor programming is now the norm – even if you don't actually need it.
● Hardware is rapidly heading towards distributed memory architectures, instead of shared memory multi-threading.
● Shared memory multi-threading requires locks, semaphores, monitors, etc. and programmers find it hard to get that stuff right.
● Actor Model, Dataflow Model, and CSP are higher level abstractions of managed control flow that are easier for programmers to deal with.
Copyright © 2011 Russel Winder 55
Summary of the Summary
● Shared-memory multi-threading is like stacks, you know it's there, but you just don't worry about it.
Copyright © 2011 Russel Winder 56
Summary of the Summary of the Summary
● If you think locks, semaphores, monitors, etc. are important for your work, you are either working in the concurrency frameworks business (*) OR you are doing it wrong.
(*) Which includes operating systems.
Copyright © 2011 Russel Winder 57
END
for ( person in roomContents.collect { x -> if ( x instanceof Human ) x } ) { person << "Please leave peaceably."}println "Goodbye all."
Copyright © 2010 Russel Winder 58
Blatant Advertising
Python for Rookies
Sarah Mount, James Shuttleworth andRussel Winder
Thomson Learning
Buy these books!Buy these books!
Now called Cengage Learning.
Developing Java Software Third Edition
Russel Winder and Graham Roberts
Wiley
Copyright © 2011 Russel Winder 59
Just Keep Passing the Messages
Dr Russel Winder
It’z Interactive [email protected]
@russel_winder