Last night was our Clojure event with Rich Hickey. BTW, if you were there, regular meeting of the Western Mass Developers Group are every other Thursday at Panera Bread in Hadley (7pm) — most of the time, we just sit around and talk — very informal — hope to see you there.
Back to the presentation — hopefully the slides/video/etc will be up somewhere soon — Rich focussed on the concurrency features of Clojure (Vars, Refs, and Agents). First he showed us the state of concurrency support in Java/C# today (locking) and identified the problem as having direct references to mutable objects. Clojure offers only immutable atoms and datastructures so that’s how he addresses one part of the problem. However, since the mutability is how the world works, Clojure models change as a change to what a reference points to. So, if you view the world through refs, it will appear to change. Then he introduced the controlled ways in which refs can change.
1. Vars are essentially thread local variables. They are a reference whose changes can only be seen in a single thread — each thread gets an independent and isolated view of the variable. Like thread-local variables — they are a very simple way to take code that looks like it’s manipulating global state and give each thread its own isolated copy.
2. Refs are mutable references to immutable objects. They can only be changed inside a transaction and all changes to all refs in a transaction appear to have happened at the time the transaction ends (to other threads). Inside the transaction, the changes are immediate. When two transactions are open and change the same ref, the second one to finish is automatically retried (so you should have no side-effects inside of transactions). All Refs read inside of a transaction will return the value that they had at the start of the transaction (so are consistent with each other).
3. Agents (which I have explained before) are for asynchronous changes to a reference. Each agent hold a value and you send actions to it to change the value. Only one action will happen to an agent at a time and you can read the current state of the agent any time. Rich spent a little time to caution us against any comparison to Erlang’s actors. They are a different concept with different tradeoffs. Agents have the advantage that they can be read from any thread conveniently (just dereference) whereas actors require sending a message to read (which is asynchronous). Clearly, Erlang’s benefit is location transparency (for distributed code) — which is what it was designed for. Rich hinted that Clojure might have an Actor concept, but it would not be unified with Agents.
What was new to me with agents is that there are two types of message sending mechanisms — send and send-off (which appear to be undocumented right now) — Rich used send-off which dedicates a thread to the agent (rather than run the action from a thread-pool). This is how you have to do it if agents block at all (which ants do because they have a small sleep).
Then, he showed us an ant simulation — I think he will make the code available. In short, there is a square world of cells, each one holds a Ref to what is in the cell (food, ant, pheromones, home base). Ants are represented by agents, which are sent a behave message. Behave picks a random action for the ant to do (move, turn, pick up food, drop food) and then mutates the world of cells in a transaction, then it sends itself another behave action. There is an agent responsible for painting the world, and another which evaporates pheromones in the background.
Anyway, the demo was impressive — since painting makes a copy of the world in a transaction, it has a completely consistent view of the world. Refs make sure that all changes to them are in transactions, so you have language support for where you need cooperation (contrasted to locking, which is not enforced). Agents also help you model the real world in a way that a coordinated single-looped simulation (calling behave on ants in a loop, then paint over and over again) could not.
And clojure’s agent mutating mechanism (sending a function to it), means that agents don’t have to have any knowledge of the messages that might be sent to it (again contrasted to Erlang Actors). Finally, various messages can take different times to complete and that would be represented in the simulation — some ants might complete several small actions in the time it took another to complete one (which would not be the case in a behave loop).
I’ll have more on this when the slides, code, etc are available.