Actors, a Unifying Pattern for Scalable Concurrency | C4 2006

Slide 1: actors a unifying design pattern for scalable concurrency www.iolanguage.com steve@dekorte.com

Slide 2: talk overview what is an actor? concurrency trends problems and solutions the big picture

Slide 3: what is an actor? an informal definition an object with an asynchronous message queue and an execution context for processing that queue encapsulates state, instructions and execution (a CSP is a process level actor)

Slide 4: concurrency trends a quick look

Slide 5: clock speed leveling off 5,000 ? 4,000 ? 2,000 800 120 50 1990 1995 2000 2005 2010 2015

Slide 6: cores per machine increasing exponentially 80 ? 24 ? 8 4 2 2002 2004 2006 2008 2011

Slide 7: clusters massive scaling typical scale memcached 10^2 gfs 10^4 mmog 10^3 p2p 10^5 @home 10^6

Slide 8: trends are cores and clusters ideal concurrency model will naturally scale across both

Slide 9: traditional concurrency model preemptive threads with shared memory and coordination via locks

Slide 10: problem nondeterminism

Slide 11: For concurrent programming to become mainstream, we must discard threads as a programming model. Nondeterminism should be judiciously and carefully introduced where needed, and it should be explicit in programs. - Ed Lee, The Problem with Threads Berkeley CS Tech Report

Slide 12: traditional concurrency model threads can directly change one another’s state “spaghetti concurrency” thread A thread B

Slide 13: actor/csp model only a thread can directly change it’s own state thread A thread B queue queue same model works across machines

Slide 14: a natural extension actors are the object paradigm extended to execution objects encapsulate state and instructions actors encapsulate state, instructions and execution

Slide 15: what this looks like in Io any object becomes an actor when sent an async message account deposit(10.00) // sync message account @deposit(10.00) // future message account @@deposit(10.00) // async message Io is a hybrid actor language

Slide 16: problem asynchronous programming

Slide 17: a future an object returned from an async message which becomes the result when it is available. If it is accessed before the result is ready, it blocks the calling execution context.

Slide 18: futures sync programming with async messages by decoupling messages from return values allows lazy synchronization and automatic deadlock detection

Slide 19: another look account deposit(10.00) // sync message account @deposit(10.00) // future message account @@deposit(10.00) // async message

Slide 20: another example // waits for result data := url fetch // returns a future immediately data := url @fetch data setFutureDelegate(self)

Slide 21: the undiscovered country lots of interesting concurrent coordinating patterns can be composed from futures or other forms of async return messages c := urls cursor c setConcurrency(500) c setDelegate(self) c foreach(fetch)

Slide 22: problem distributed programming

Slide 23: transparent distributed objects unified local and remote messaging peers append(DO at(ip, port, key)) ... c := peers cursor c setConcurrency(50) c setDelegate(self) c foreach(search(query)) eliminates protocol hassles but doesn’t mean we can ignore the network, etc

Slide 24: problem the high concurrency memory bottleneck

Slide 25: execution contexts memory usage and maximum concurrency bytes max per GB process/os thread 1,000,000s 100s* stackfull coro 10,000s 10,000s stackless coro or 100s 1,000,000s** continuation *webkit etc use os threads.. **but thread related state may exceed stack size

Slide 26: synonyms synonyms user level thread os thread lightweight thread kernel thread microthread native thread green thread coroutine coro fiber

Slide 27: taxonomy execution context / thread user level thread os thread large fixed stack size nondeterministic hw registers swapping coroutine continuation linked activation records stackfull coro stackless coro small fixed stack size variable stack size deterministic hw registers swapping

Slide 28: what this means in practice connections

Slide 29: user level threads aren’t preemptive, so what about blocking ops? avoid them by using async sockets and async file i/o

Slide 30: async issues ease of use? pause the user level thread while waiting on the async i/o request cpu bound blocking? use explicit yield() where needed

Slide 31: conclusion use user level threads for scaling concurrency on a given core and one os threads or processes per core for scaling across cores

Slide 32: the big picture “powers of 10” each level follows the actor pattern of encapsulating state, instructions and execution and communicating via async queued messaging

Slide 33: actor user thread

Slide 34: actor actor actor actor csp actor actor actor actor os process actor actor actor actor actor actor actor actor

Slide 35: csp csp csp csp csp csp csp csp machine csp csp csp csp csp csp csp csp

Slide 36: machine machine machine machine machine machine machine machine cluster machine machine machine machine machine machine machine machine

Slide 37: some fun speculation what about cores? a prediction based on this pattern

Slide 38: traditional SISD architecture works, but clock speed growth is slowing and silicon is cheap core bus memory

Slide 39: current MISD architecture bus bottleneck - memory performance per core drops as core count increases core core bus memory

Slide 40: future MIMD architecture? the actor pattern on the hardware level a connection machine on a chip core SISD SISD SISD SI SISD memory SISD SISD SISD SISD SI SISD SISD SISD SISD SI

Slide 41: automatic MIMD distribution SISD SISD SISD SISD SI array SISD SISD SISD SISD SI array shard array shard array shard array shard SISD SISD SISD SISD SI

Slide 42: this talk, in a nutshell solution problem solution concurrency nondeterminism actors/csp distributed and async transparent distributed actors/csp programming objects and futures high concurrency memory bottleneck user level threads many cores bus bottleneck MIMD www.iolanguage.com steve@dekorte.com