/
benchmarks.Rmd
274 lines (204 loc) · 8.67 KB
/
benchmarks.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
--
title: "Benchmarking Iteration"
output: rmarkdown::html_vignette
self_contained: yes
vignette: >
%\VignetteEngine{knitr::rmarkdown}
%\VignetteIndexEntry{Benchmarking Iteration}
%\VignetteEncoding{UTF-8}
---
This document illustrates the differences in performance between `iterors` and previous packages `iterators`/`itertools`/`itertools2`. Basic gains come from the `nextOr` method avoiding the need for `tryCatch`, as well as the fact that iterors can be used simply as functions, skipping S3 dispatch. Additionally, over the process of curating the `iterors` package, several iterator implementations have been optimized for performance.
Because this is a more computationally intensive vignette, it is built manually and included in the released package statically. The `Rmd` source is in the Git repo under the `scratch` directory; knit it manually and copy it into the vignettes directory to publish. When this document was last built, it was under these conditions:
* iterors version 1.0, git revision 257dc955 ()
* itertools2 version 0.1.1
* itertools version 0.1.3
* OS: Linux 5.4.0-148-generic
* CPU: Intel(R) Core(TM) i5 CPU M 520 @ 2.40GHz.
* R version 4.3.0 (2023-04-21), svn rev 84292
# Basic iterators (self-contained/memory-backed)
## Consuming iterators
Since the following tests will use different versions of `icount`, `iter_consume`, and `as.list`,
first check how close those functions are.
```r
as.list.bench <- microbenchmark::microbenchmark(
iter_consume = iter_consume(iterators::icount(3000)),
iterors.consume = iterors::consume(iterors::icount(3000)),
iterators.as_list = as.list(iterators::icount(3000)),
iterors.as_list = as.list(iterors::icount(3000)),
times = 10)
plotit(as.list.bench)
```
![plot of chunk unnamed-chunk-2](static_figure/unnamed-chunk-2-1.png)
Note that we are using our own function `iter_consume`, as `itertools::consume` has poor performance due to invoking `try` on every item.
## Simple iteration over a vector
```r
vector.bench <- microbenchmark::microbenchmark(
iterators = iter_consume(iterators::iter(1:1000)),
iterors = iterors::consume(iterors::iteror(1:1000)),
times = 10)
plotit(vector.bench)
```
![plot of chunk unnamed-chunk-3](static_figure/unnamed-chunk-3-1.png)
## Chunking a vector
```r
chunk.bench <- microbenchmark::microbenchmark(
itertools = iter_consume(itertools::isplitVector(1:6000, chunkSize=17)),
iterors = iterors::consume(iterors::iteror(1:6000, chunkSize=17)),
times = 10)
plotit(chunk.bench)
```
![plot of chunk unnamed-chunk-4](static_figure/unnamed-chunk-4-1.png)
## Extracting rows or columns
```r
arr <- array(1:1000000, c(1000,1000))
slice.bench <- microbenchmark::microbenchmark(
iterators.rows = iter_consume(iterators::iter(arr, by="row")),
iterators.cols = iter_consume(iterators::iter(arr, by="column")),
itertools.array_rows = iter_consume(itertools::iarray(arr, MARGIN = 1)),
itertools.array_cols = iter_consume(itertools::iarray(arr,MARGIN = 2)),
iterors.rows = iterors::consume(iterors::iteror(arr, by=1)),
iterors.cols = iterors::consume(iterors::iteror(arr, by=2)),
times = 10)
plotit(slice.bench)
```
![plot of chunk unnamed-chunk-5](static_figure/unnamed-chunk-5-1.png)
## Split matrix into chunks
```r
arr <- array(1:1000000, c(1000,1000))
array_chunks.bench <- microbenchmark::microbenchmark(
itertools.rows = iter_consume(itertools::isplitRows(arr, chunkSize=7)),
itertools.cols = iter_consume(itertools::isplitCols(arr, chunkSize=7)),
iterors.rows = iterors::consume(iterors::iteror(arr, by=1, chunkSize=7)),
iterors.cols = iterors::consume(iterors::iteror(arr, by=2, chunkSize=7)),
times = 10)
plotit(array_chunks.bench)
```
![plot of chunk unnamed-chunk-6](static_figure/unnamed-chunk-6-1.png)
## Enumerating N-D array indices:
```r
icountn.bench <- microbenchmark::microbenchmark(
iterators = iter_consume(iterators::icountn(c(4,5,6,7))),
iterors = iterors::consume(iterors::icountn(c(4,5,6,7))),
times =10)
plotit(icountn.bench)
```
![plot of chunk unnamed-chunk-7](static_figure/unnamed-chunk-7-1.png)
## Cartesian product of vectors
```r
igrid.bench <- microbenchmark::microbenchmark(
# itertools2 = iter_consume(itertools2::iproduct(1:10, letters, month.name)),
itertools = iter_consume(itertools::product(1:10, letters, month.name)),
itertools2 = iter_consume(itertools2::iproduct(1:10, letters, month.name)),
iterors = iterors::consume(iterors::igrid(1:10, letters, month.name)),
times = 10)
plotit(igrid.bench)
```
![plot of chunk unnamed-chunk-8](static_figure/unnamed-chunk-8-1.png)
## Random number generation
When `independent=TRUE` there is a slight amount of extra work in saving and restoring the
random number generator settings. The difference becomes smaller the larger "n" you pick.
```r
sampling.bench <- microbenchmark::microbenchmark(
iterators = iter_consume(iterators::irunif(100, count=100)),
iterors.independent = as.list(iterors::irunif(100, count=100, independent=TRUE)),
iterors.dependent = as.list(iterors::irunif(100, count=100, independent=FALSE)),
times = 100)
plotit(sampling.bench)
```
![plot of chunk unnamed-chunk-9](static_figure/unnamed-chunk-9-1.png)
What are the costs of speed here? here?
```
library(profvis)
```
# Higher order iterators / iterator functions
## Chaining iterators end-to-end
```r
chain.bench <- microbenchmark::microbenchmark(
itertools = iter_consume(
do.call(itertools::chain,
lapply(1:50, iterators::icount))),
itertools2 = iter_consume(
do.call(itertools2::ichain,
lapply(1:50, iterators::icount))),
iterors =
iterors::consume(do.call(iterors::i_chain, lapply(1:50, iterors::icount))),
iterors.collapse =
iterors::consume(iterors::i_concat(iterors::i_apply(1:50, iterors::icount))),
times=10)
plotit(chain.bench)
```
![plot of chunk unnamed-chunk-10](static_figure/unnamed-chunk-10-1.png)
## Keeping or dropping items according to a criterion
```r
i_keep.bench <- microbenchmark::microbenchmark(
iterators = as.list(itertools::ifilter(\(x) floor(sqrt(x))^2 == x, iterators::icount(5000))),
iterors = as.list(iterors::i_keep(iterors::icount(5000), \(x) floor(sqrt(x))^2 == x)),
times = 10)
plotit(i_keep.bench)
```
![plot of chunk unnamed-chunk-11](static_figure/unnamed-chunk-11-1.png)
## Select only unique values
`iterors::i_unique` uses a hash table for better performance.
```r
i_unique.bench <- microbenchmark::microbenchmark(
iterators = {
it <- itertools2::iunique(iterators::isample(1:100, 1))
as.numeric(itertools2::take(it, 100))
},
iterors = {
it <- iterors::i_unique(iterors::isample(1:100, 1, independent=FALSE))
iterors::take(it, 100, "numeric")
},
times=10)
plotit(i_unique.bench)
```
![plot of chunk unnamed-chunk-12](static_figure/unnamed-chunk-12-1.png)
## i_recycle
Note that `itertools2::icycle` only works for self-contained vector-based iterators, while the other two use a general purpose buffer.
```r
icycle.bench <- microbenchmark::microbenchmark(
itertools2 = as.list(itertools2::icycle(iterators::icount(20), times = 20)),
itertools = iter_consume(itertools::recycle(iterators::icount(20), times = 20)),
iterors = iterors::consume(iterors::i_recycle(iterors::icount(20), times = 20)),
times = 20)
plotit(icycle.bench)
```
![plot of chunk unnamed-chunk-13](static_figure/unnamed-chunk-13-1.png)
## i_tee
Note that `itertools2::tee` only works for vector-based iterators it
knows how to clone, while `iterors::i_tee` uses a general purpose
buffering mechanism.
```r
i_tee.bench <- microbenchmark::microbenchmark(
itertools2={
iter_consume(
do.call(itertools::chain,
itertools2::itee(1:20, n=20)))
},
iterors={
iterors::consume(
iterors::i_concat(
iterors::i_tee(iterors::icount(20), n=20)))
},
times=20)
plotit(i_tee.bench)
```
![plot of chunk unnamed-chunk-14](static_figure/unnamed-chunk-14-1.png)
## Sliding window iterators
`iterors` uses specialized routines for n=2 and n=3. For n=4 and above, a
general purpose queue is used.
```r
marks=c(alist(
itertools2.w2={iter_consume(itertools2::ipairwise(1:200))}),
if (exists("itripletwise", asNamespace("itertools2"), inherits=FALSE)) {
alist(itertools2.w3={iter_consume(itertools2::itripletwise(1:200))})
},
alist(
iterors.w2={iterors::consume(iterors::i_window(1:200, 2))},
iterors.w3={iterors::consume(iterors::i_window(1:200, 3))},
iterors.w4={iterors::consume(iterors::i_window(1:200, 4))},
iterors.w50={iterors::consume(iterors::i_window(1:200, 50))}))
i_window.bench <- microbenchmark::microbenchmark(list=marks, times=10)
plotit(i_window.bench)
```
![plot of chunk unnamed-chunk-15](static_figure/unnamed-chunk-15-1.png)