1 unstable release

Uses old Rust 2015

0.1.0 Nov 5, 2017

#1037 in Concurrency

MPL-2.0 license

22KB
256 lines

polyester

some parallel iterator adapters for rust

Documentation (coming soon!) | (manually-updated docs for master)

This work-in-progress library features an extenstion trait, Polyester, that extends Iterator and provides some parallel operations that allow you to spread consumption of the iterator across multiple threads.

To use this crate in your own project, add the following to your Cargo.toml:

[dependencies]
polyester = "0.1.0"

...and the following to your crate root:

extern crate polyester;

use polyester::Polyester;

but wait, isn't that what rayon does?

Not quite. rayon is built on creating adaptors that fit its own ParallelIterator trait, which can only be created from fixed-length collections. polyester wants to work from arbitrary Iterators and provide arbitrary Iterators back (or consume them and provide some appropriate output). Basically, it's purely an extension trait for anything that already implements Iterator. This means that there are distinct design goals for polyester that give it different performance characteristics from rayon.

architecture overview

The internal design of this library changes pretty rapidly as i consider the ramifications of various design decisions, but you're welcome to take a look and offer suggestions to improve performance or suggest additional adaptors. For the moment, this is the basic idea, as of this writing (2017-10-30):

The major problem with wanting to spread an iterator's values across threads is that the iterator itself becomes a bottleneck. There is only one source of iteration state, and it requires a &mut self borrow to get the next item. The way that polyester currently deals with this is placing the iterator into a background thread so it can be loaded into per-thread queues, which the worker threads pick up at the same time.

There's a major drawback to this approach, though: If the worker threads are not expected to be performing a lot of per-item work, this will always be slower than just doing it sequentially. Therefore, the current version of polyester is only recommended if you need to perform intensive (or erratically intensive) work per item, or cannot afford to work sequentially or collect the iterator beforehand (to hand off to rayon instead). Note that if you have an expensive sequential operation in the iterator before you hand it off to polyester, that will impact how fast the cache-filler can generate items.

Anyway, once this "hopper" is prepared, handles to it are given to a number of worker threads, so that they can run user-supplied closures on the items. Each thread has its own queue to load items from, and if its own queue is empty it will begin walking forward through other thread's queues to attempt to load more items before waiting. Each queue has an associated SignalEvent (from synchronoise) which the cache-loader worker will signal periodically while filling items or once the iterator has been exhausted.

From here, each adaptor has its own processing:

par_fold

par_fold performs a basic folding operation with its items: Start from a seed accumulator, and iteratively add items to it until you're done. However, since there are multiple folds happening at the same time, there needs to be an additional step where all the "intermediate" accumulators are brought together. This is done by offering each worker thread a channel to hand their finished accumulator back to the calling thread, which performs the "outer fold" to merge all the accumulators together.

par_map

par_map has a simpler goal with a slightly more complicated implementation: Since it wants to turn the parallel iterator back into a sequential one, it performs the given map in each thread before handing each item back to another channel, whose Receiver is used in the corresponding ParMap Iterator implementation.

Note that due the nature of the hopper creation and processing, order is not guaranteed for the items that come out of either adapter. (In fact, for par_fold each thread is basically getting every Nth item, and the outer_fold will need to reconstruct the ordering if it wants it.)

Dependencies

~305KB