How are message-passing concurrent languages better than shared-memory concurrent languages in practice?

The three main contenders for simplifying concurrency are actors, software transactional memory (STM), and automatic parallelization; Scala has implementations of all three.

The three main contenders for simplifying concurrency are actors, software transactional memory (STM), and automatic parallelization; Scala has implementations of all three. Actors Actors find their most notable implementation in the language Erlang, which as far as I know is where the idea started*. Erlang is designed from the ground up around actors.

The idea is that actors themselves are black boxes to each other; they interact only by passing messages. Scala has an implementation of actors in its library, and variants are available in external libraries. In the main library, the black-box-ness is not enforced, but there are easy-to-use methods for passing messages, and Scala makes it easy to create immutable messages (so you don't have to worry that you send a message with some content, and then change the content at some random time).

The advantage of actors is that you don't have to worry about complex shared state, which really simplifies the reasoning involved. Also, you can decompose the problem into smaller pieces than threads and let the actor library figure out how to bundle actors into the appropriate number of threads. The downside is that if you are trying to do something complex, you have a lot of logic to deal with for sending messages, handling errors, and so on, before you know it succeeds.

Software Transactional Memory STM is based on the idea that the most important concurrent operation is to grab some shared state, fiddle with it, and write it back. So it provides a means of doing this; however, if it encounters some problem--which it typically delays detecting until the very end, at which point it checks to make sure the writes all went correctly--it rolls back the changes and returns a failure (or tries again). This is both high-performance (in situations with only moderate contention, since usually everything goes just fine) and robust to most sorts of locking errors, since the STM system can detect problems (and even potentially do things like take access away from a lower-priority request and give it to a higher-priority one).

Unlike actors, it's easier to attempt complex things, as long as you can handle failure. However, you also have to reason correctly about the underlying state; STM prevents rare unintentional deadlocks via failing and retrying, but if you've simply made a logic error and a certain set of steps cannot complete, STM cannot allow it to. Scala has a STM library that is not part of the standard library but is being considered for inclusion.

Clojure and Haskell both have well-developed STM libraries. Automatic Parallelization Automatic parallelization takes the view that you don't want to think about concurrency; you just want stuff to happen fast. So if you have some sort of parallel operation--applying some complex operation to a collection of items, one at a time, and producing some other collection as a result, for instance--you should have routines that automatically do this in parallel.

Scala's collections can be used in this way (there is a . Par method that converts a conventional serial collection into its parallel analog). Many other languages have similar features (Clojure, Matlab, etc.).

Edit: Actually, the Actor model was described back in 1973 and was probably motivated by earlier work in Simula 67 (using coroutines instead of concurrency); in 1978 came the related Communicating Sequential Processes. So Erlang's capabilities were not unique at the time, but the language was uniquely effective at deploying the actor model.

In my opinion, Automatic Parallelization means: passing a serial program as it is without modifications to a compiler/tool produces a parallelized version of the program. The parallelized version has to run at least as fast as the serial version. – Atom Nov 2 '11 at 6:54 @Atom - That is the ideal, but almost no programs can do that successfully, since the analysis required to know whether a parallel approach is both correct and faster is considerable.

One could partially do that in Scala if one set things up (mostly via imports) to always use the parallel version of collections. Maybe I should call it "semi-automatic"; I certainly agree that it's not fully automatic (in Scala or almost anything else). – Rex Kerr Nov 2 '11 at 16:27 +1.

However, use of .par. Doesn't preclude the need to think about concurrency. If you're modifying mutable state, it doesn't protect you in any way.It does help by not having to create threads etc. Though.

– Matthew Farwell Nov 2 '11 at 16:50 @Matthew Farwell - Agreed; it greatly reduces the burden (to "this had better be trivially parallelizable and not rely upon mutable state") but does not eliminate it. – Rex Kerr Nov 2 '11 at 16:54.

In an idiomatic Go program, threads communicate state and data through channels. This can be done without the need for locks. Passing data through a channel to a receiver, implies transfer of ownership of the data.

Once you send a value through a channel, you should not be operating on it anymore as whoever received it now 'owns' it. However, it should be noted that this transfer of 'ownership' is not enforced by the Go runtime in any way. Objects sent through channels are not flagged or marked or anything like that.It is merely a convention.

So you can, if you are so inclined, shoot yourself in the foot by mutating a value you previously sent through a channel. Go's strength lies in that the syntax Go offers (launching of goroutines and the way channels work), makes it a lot easier to write code which functions correctly and thus prevents race conditions and dead locks. Go's clear concurrency mechanics make it very easy to reason about what is going to happen in your program.

As a side note: The standard library in Go does still offer the traditional mutexes and semaphores if you really want to use them. But you obviously do so at your own discretion and risk.

– Rex Kerr Nov 1 '11 at 18:01 As I understand it, Go's approach is based on Tony Hoare's 'CSP' (Communicating Sequential Processes) work. Some info can be found here: usingcsp.com. – jimt Nov 1 '11 at 20:17 Well, the main difference there is in whether the communication is synchronous or not; I think I'd put them in the same general category, even if the theory is not the same behind the two.

Practically, you face the same issues, for the most part; the problems just manifest themselves in different ways (full mailbox vs. waiting forever to send a message). – Rex Kerr Nov 1 '11 at 20:45.

For me, using Scala (Akka) actors has had several advantages over traditional concurrency models: Using a message-passing system like actors gives you a way to easily handle shared state. For example, I will frequently wrap a mutable data structure in an actor, so the only way to access it is through message-passing. Since actors always process one message at a time, this ensures that all operations on the data are thread-safe.

Actors partially eliminate the need to deal with spawning and maintaining threads. Most actor libraries will handle distributing actors across threads, so you only need to worry about starting and stopping actors. Often I will create a series of identical actors, one per physical CPU core, and use a load-balancer actor to evenly distribute messages to them.

Actors can help improve system reliability. I use Akka actors, and one feature is you can create a supervisor for actors, where if an actor crashes the supervisor will automatically create a new instance. This can help prevent situations with threading where a thread crashes and you're stuck with a half-running program.It's also very easy to spin up new actors as needed and work with remote actors running in another application.

You still need a decent understanding of concurrency and multi-threaded programming since deadlocks and race conditions are still possible, but actors make it much easier to identify and solve these issues. I don't know how much these apply to Android apps, but I mostly do server-side programming and using actors has made development much easier.

Scala actors work on shared-nothing principle, so there are no locks (and hence no dead locks)! Actors listen for messages and are invoked by the code that has something for actor to work upon.

Deadlocks may occur in a Actor based system if two actors are waiting for messages from each-other, and they never get the message. However it is easier to reason about this than in a shared memory system where the environment (processor speed, avaliable processors, ... ) can change timings and so forth. Dalnefre.com/wp/2010/08/dining-philosophers-in-humus – oluies Nov 1 '11 at 17:00 thats a great point - thanks for sharing it!

Although I would guess that it would be much more difficult to write such a code that could land you into a deadlock in Scala. Actors passing messages to each other may be an interesting way to program I guess. – aishwarya Nov 1 '11 at 17:10.

I'm appalled that Erlang is not mentioned! (Rex Kerr mentioned Erlang in his excellent answer) No-shared-memory message passing is practically the whole purpose of Erlang. See Wikipedia > Erlang # Projects using Erlang.(Ever heard of CouchDB?

).

I mentioned Erlang in my answer--it's the first language mentioned in the first approach to simplifying concurrency. – Rex Kerr Nov 1 '11 at 18:57 @RexKerr yes I noticed as I was writing my answer, you posted yours. :) Well done.

– Dan Burton Nov 1 '11 at 19:24.

I cant really gove you an answer,but what I can give you is a way to a solution, that is you have to find the anglde that you relate to or peaks your interest. A good paper is one that people get drawn into because it reaches them ln some way.As for me WW11 to me, I think of the holocaust and the effect it had on the survivors, their families and those who stood by and did nothing until it was too late.

Related Questions