This is a library for performing speculative execution, based off of this paper. It is written for Rust 0.8 and compiles with rustpkg.

Usage

There are two functions provided in the speculate library: spec and specfold.

Simple, two-task speculation

The spec function looks like this:

fn spec<A: Eq + Send + Clone, B>(producer: ~fn() -> A,
                                 predictor: ~fn() -> A,
                                 consumer:  ~fn(A) -> B) -> B

Both the producer and consumer functions are assumed to take a long time to run, while the predictor should be quick. The producer is run in its own task, then the predictor comes up with a guess, passes it to the consumer, which is run. Once the producer finishes, its result is compared to that of the predictor. If they match, then the result produced by the consumer is returned. If they do not match, then the consumer is re-run with the actual result from the producer.

Here is an adapted illustration from a Haskell implementation of the same paper. p is the producer, g is the predictor (guess), c is consumer, and a is the actual value returned by the producer.

The best-case timeline looks like:

foreground: [----- p -----]
foreground:               [-]    (check g == a)
spark:         [----- c g -----]
overall:    [--- spec p g c ---]

The worst-case timeline looks like:

foreground: [----- p -----]
foreground:               [-]              (check g == a)
foreground:                 [---- c a ----]
spark:         [----- c g -----]
overall:    [-------- spec p g c ---------]

Iterative speculation

The specfold function launches a configurable number of tasks to work in parallel. It looks like this:

fn specfold<A: Eq + Clone + Send>(iters: uint,
                                  loop_body: &fn() -> ~fn(int, A) -> A,
                                  predictor: &fn() -> ~fn(int) -> A)

In this case, iters tasks are spawned which each execute an iteration of loop_body. Each task executes loop_body()(idx, predictor()(idx)) in parallel. After all tasks have been launched, the main task sequentially checks each of the predictions and re-runs the loop body if a prediction was incorrect. A future version may attempt to do this in parallel.

The reason predictor and loop_body are functions which return other functions is because I couldn't get Rust's compiler to leave me alone otherwise.

CSS parser

A modified version of rust-cssparser is included and is used as a more real-world test of the library. The original version mixes tokenization with parsing, which is fine in the single-threaded case, but doesn't work as well here. The version included does only tokenization, which is useful when trying to parallelize. The spec_css library implements a speculative lexer using specfold.

Benchmarking the lexer

If you put CSS files in a folder called sample-data at the project root and run the executable produced by the testing library, it will, for each file, run the lexer sequentially and in parallel and write to stdout a CSV file. The CSV file has columns name, seq, par, size, where seq and par are the time taken (in microseconds) to tokenize the file sequentially and in parallel, respectively, and size is the size of the file in bytes.

The number of tasks is controlled by providing a command-line argument like so (for 6 tasks):

./build/x86_64-unknown-linux-gnu/testing/testing 6

The default number of tasks is 4.

Sample benchmark results

Here are the results from tokenizing the CSS files from the Alexa top 11 (I felt like including Amazon) on my Core 2 Quad Q6600 @ 2.40GHz with 4 tasks. This is the speedup of the parallel version compared to the sequential, sorted by file size (in KiB):

name	speedup	size
linkedin00.css	1.837	418.3
youtube01.css	0.986	293.04
facebook04.css	1.865	244.61
linkedin05.css	1.229	220.54
linkedin02.css	1.848	199.26
twitter03.css	1.84	161.32
yahoo04.css	1.848	154.4
youtube02.css	1.662	133.97
linkedin01.css	1.884	122.45
bootstrap.css	1.585	117.08
wikipedia06.css	1.31	116.46
outlook12.css	0.774	116.42
amazon14.css	1.866	108.49
outlook14.css	1.642	96.09
bootstrap.min.css	1.888	95.06
facebook07.css	1.419	77.38
qq00.css	1.757	68.29
wikipedia02.css	0.867	66.18
facebook01.css	1.947	52.32
facebook03.css	1.588	48.96
google08.css	1.32	45.02
google00.css	1.241	42
facebook09.css	1.626	40.71
yahoo05.css	1.376	40.11
wikipedia01.css	0.857	39.18
outlook01.css	1.387	35.95
facebook00.css	1.843	33.78
outlook10.css	1.413	30.11
linkedin04.css	1.282	25.45
facebook08.css	1.288	25.32
wikipedia03.css	0.968	25.08
amazon13.css	1.199	22.34
amazon15.css	1.508	21.88
linkedin03.css	1.925	21.73
wikipedia07.css	1.589	20.56
outlook06.css	0.991	18.62
bootstrap-theme.css	1.748	16.42
google03.css	1.356	16.33
outlook05.css	1.876	16.31
bootstrap-theme.min.css	1.89	14.64
amazon00.css	1.154	13.98
outlook15.css	1.254	13.94
amazon11.css	1.573	8.53
qq04.css	0.896	8.4
google07.css	0.803	7.64
wikipedia00.css	1.611	7.21
twitter00.css	1.546	5.76
outlook07.css	1.14	5.66
outlook18.css	1.489	5
outlook13.css	1.141	4.92
youtube00.css	1.184	4.76
amazon16.css	1.492	4.41
outlook17.css	1.552	3.7
baidu00.css	1.534	3.25
amazon09.css	1.106	3.19
google02.css	1.44	3.05
google01.css	1.041	2.76
qq05.css	1.453	2.76
google06.css	1.069	2.53
outlook02.css	1.439	2.41
amazon03.css	1.1	2.28
amazon01.css	1.227	2.09
yahoo03.css	1.003	1.96
yahoo06.css	1.001	1.81
google05.css	0.986	1.73
outlook08.css	0.897	1.34
amazon12.css	0.987	1.32
outlook00.css	0.904	1.26
baidu01.css	1.092	0.97
amazon06.css	0.571	0.95
yahoo00.css	0.715	0.95
yahoo02.css	0.827	0.74
outlook16.css	0.975	0.73
facebook02.css	0.758	0.68
facebook05.css	0.608	0.57
amazon10.css	0.647	0.5
amazon04.css	0.767	0.49
outlook11.css	0.668	0.45
facebook06.css	0.564	0.42
google04.css	0.534	0.41
outlook09.css	0.526	0.36
amazon02.css	0.504	0.32
qq01.css	0.61	0.31
amazon08.css	0.461	0.27
amazon05.css	0.373	0.22
qq02.css	0.365	0.18
wikipedia04.css	0.2	0.18
yahoo01.css	0.39	0.18
twitter01.css	0.163	0.11
amazon07.css	0.223	0.11
twitter02.css	0.124	0.07
wikipedia05.css	0.13	0.07
outlook03.css	0.131	0.06
qq03.css	0.108	0.03
outlook04.css	0.074	0.02

haxney / speculate