The case FOR deterministic results

3 January 2013

In my last post, I made the case against having a deterministic semantics. I’ve gotten a fair amount of feedback saying that, for a Web API, introducing nondeterminism is a very risky idea. Certainly the arguments are strong. Therefore, I want to take a moment and make the case for determinism.

Why determinism?

All things being equal, it’s clear that deterministic execution semantics are preferable. They’re easier to debug and they avoid the question of browser incompatibilities.

One interesting observation (most recently pointed out by Roc in this comment) is that while the intention of a nondeterministic ordering is to free up the implementor, what sometimes happens is that all people wind up relying on the behavior of one implementation, and then the others follow suit.

That said, there seem to be numerous examples of nondeterministic portions of JavaScript: the iteration order for properties, for example, or the order of callbacks to the comparator in Array.sort. But then there are plenty of examples of implementations being constrained by arbitrary behavior inherited from legacy interpreters. In any case, I am not an expert in these kind of nitty gritty cross-browser compatibility details.

If we did opt for nondeterministic semantics, it might be plausible to use a cheap PRNG like Xorshift in the sequential fallback so as to make it more likely that unwanted dependencies on execution ordering would be seen during testing (though, of course, this adds overhead too!).

What about performance?

It is true that nondeterministic semantics give maximum efficiency, but the magnitude of these performance gains is not entirely clear. In my previous post, for example, I stressed the behavior around bailouts. It is true that guaranteeing deterministic semantics will result in wasted work and in general make bailouts less efficient. However, it is also true that bailouts are the exceptional case: if things are working properly, they should only occur at the beginning of execution. Once sufficient type information has been gathered, parallel execution without bailouts should be the norm—unless of course it turns out the code is not parallelizable, either because it is impure or because it uses some unsupported language features, in which case we will simply use the sequential fallback from the start and not even attempt parallelism.

So, at least in the case of functions like map(), efficiency in the steady state should not be negatively impacted by deterministic semantics, presuming that the kernel function is pure.

But what about reduce, scan, and scatter?

As I wrote in the previous post, the current semantics of ParallelArray are inconsistent. They give deterministic results for map() but not for reduce(), scan(), or scatter(). The core problem here is that the standard sequential ordering for reduce() (i.e., left-to-right) is inherently sequential—and the most efficient ordering will depend on the precise implementation strategy (how many worker threads are involved, etc). But, at the cost of some efficiency, we can choose a deterministic ordering that still permits parallel execution.

For reduce, a good ordering might be to evaluate in a tree-like fashion. So we would first reduce indices 0 and 1, then 2 and 3, 4 and 5, and so on, resulting in an array with length N/2. We can then repeat the reduction until the result has length 1. If at any step we get an array with an odd length, we can reduce the final element in with the final pair. A similar ordering can be used for scan, though the need to preserve intermediate results creates complications.

The scatter operation, at least when a conflict function is provided, is much more difficult to parallelize. I think that the only way it is possible is to make each thread walk the entire list of targets but only process writes to a specific subset of the array. If no conflict function is provided, or if the conflict function is one that is known to be associative and commutative (such as integer addition—though not floating point, sadly), then parallelization of scatter is also relatively straightforward.

So what’s the right thing to do?

At this point, the answer is probably “measure” or perhaps “wait and see”. I am somewhat concerned though about basing too many decisions on the performance of our current implementation, both because it has not been heavily optimized and because it is only one model of execution (parallel worker threads). But we’ve got to go on something.

If the API were designed for me personally to use, I would want it to have nondeterministic semantics. This gives maximum flexibility to the implementation without opening the door to data races. However, I am certainly appreciative of the concerns regarding debugability—performance is not everything!