1
00:00:06,320 --> 00:00:11,499
[Music]

2
00:00:17,359 --> 00:00:21,600
good afternoon australians time from

3
00:00:19,439 --> 00:00:23,199
sydney welcome back from lunch i hope

4
00:00:21,600 --> 00:00:24,880
everybody's had a good lunch a good

5
00:00:23,199 --> 00:00:26,720
break a bit of a walk

6
00:00:24,880 --> 00:00:27,840
chase the dog whatever you needed to do

7
00:00:26,720 --> 00:00:29,679
i had cheese and crackers for the

8
00:00:27,840 --> 00:00:31,679
wallace and grommet friends out there

9
00:00:29,679 --> 00:00:34,079
and we welcome back this afternoon for

10
00:00:31,679 --> 00:00:36,000
our first session after lunch uh with

11
00:00:34,079 --> 00:00:38,399
david and he's joining us all the way

12
00:00:36,000 --> 00:00:41,200
from the sunny u.s in the center

13
00:00:38,399 --> 00:00:43,200
san francisco bay area

14
00:00:41,200 --> 00:00:44,480
so i'll introduce david now before i

15
00:00:43,200 --> 00:00:46,640
hand him over

16
00:00:44,480 --> 00:00:48,480
so ordinarily we think about things like

17
00:00:46,640 --> 00:00:51,199
converting numbers between the way we

18
00:00:48,480 --> 00:00:53,199
see them and the way computers use them

19
00:00:51,199 --> 00:00:55,760
as they solve a problem

20
00:00:53,199 --> 00:00:58,079
and for many purposes they are

21
00:00:55,760 --> 00:00:59,760
as data volumes grow and people's

22
00:00:58,079 --> 00:01:02,000
patience doesn't

23
00:00:59,760 --> 00:01:04,239
we see sometimes find that there is

24
00:01:02,000 --> 00:01:06,640
still room for improvement

25
00:01:04,239 --> 00:01:09,360
and proving them is fairly

26
00:01:06,640 --> 00:01:11,600
novice low-level coders can

27
00:01:09,360 --> 00:01:13,600
when our time together is done

28
00:01:11,600 --> 00:01:15,759
david will have taught us how that we

29
00:01:13,600 --> 00:01:17,040
can use some of the recent advantages in

30
00:01:15,759 --> 00:01:18,560
a few areas

31
00:01:17,040 --> 00:01:21,200
a few of these things that he's brought

32
00:01:18,560 --> 00:01:22,560
together a postgres sql

33
00:01:21,200 --> 00:01:24,799
some of the things to consider when

34
00:01:22,560 --> 00:01:27,200
attempting such improvements and looking

35
00:01:24,799 --> 00:01:29,119
at the world from a new perspective

36
00:01:27,200 --> 00:01:31,680
thanks david and please do ask questions

37
00:01:29,119 --> 00:01:33,600
in the chat during the session and david

38
00:01:31,680 --> 00:01:34,799
will be hanging around in the chat room

39
00:01:33,600 --> 00:01:36,159
afterwards if you've got another

40
00:01:34,799 --> 00:01:38,320
question you didn't quite get finished

41
00:01:36,159 --> 00:01:41,280
before our timer runs out so over to you

42
00:01:38,320 --> 00:01:43,759
david thank you thanks so much

43
00:01:41,280 --> 00:01:46,320
um greetings from yesterday and from the

44
00:01:43,759 --> 00:01:49,040
unseated territory of the olani people

45
00:01:46,320 --> 00:01:50,320
or castro valley as i usually refer to

46
00:01:49,040 --> 00:01:52,320
it

47
00:01:50,320 --> 00:01:54,399
thanks to all of you for coming i

48
00:01:52,320 --> 00:01:56,640
hope to earn your attention and your

49
00:01:54,399 --> 00:01:57,920
wakefulness in the coveted post lunch

50
00:01:56,640 --> 00:02:01,119
spot

51
00:01:57,920 --> 00:02:04,000
i'd like to shout out ivan uh my

52
00:02:01,119 --> 00:02:06,560
employer who pays me to work on open

53
00:02:04,000 --> 00:02:09,360
source software and perhaps yours if if

54
00:02:06,560 --> 00:02:11,520
you're interested we can talk later

55
00:02:09,360 --> 00:02:13,360
about that if you are

56
00:02:11,520 --> 00:02:16,400
you can leave your impressions now or

57
00:02:13,360 --> 00:02:20,239
later in the venulis channel called uh

58
00:02:16,400 --> 00:02:23,599
post talk chat womanjeka theater hope i

59
00:02:20,239 --> 00:02:24,640
got that at least not too mangled

60
00:02:23,599 --> 00:02:28,000
um

61
00:02:24,640 --> 00:02:30,319
so let's get started

62
00:02:28,000 --> 00:02:32,239
apparently i'm across from my retro

63
00:02:30,319 --> 00:02:33,680
computing talk

64
00:02:32,239 --> 00:02:37,040
i hope

65
00:02:33,680 --> 00:02:39,840
somebody recognizes this this this is

66
00:02:37,040 --> 00:02:41,840
now what we call retro computing

67
00:02:39,840 --> 00:02:44,720
or as my

68
00:02:41,840 --> 00:02:47,280
13 year old self used to call it

69
00:02:44,720 --> 00:02:47,280
computing

70
00:02:47,680 --> 00:02:54,800
we had resources available

71
00:02:50,720 --> 00:02:57,680
they included a concise instruction set

72
00:02:54,800 --> 00:02:58,879
we'll see what i mean by concise

73
00:02:57,680 --> 00:03:01,760
presently

74
00:02:58,879 --> 00:03:04,480
there were some registers i don't recall

75
00:03:01,760 --> 00:03:06,560
exactly how many offhand

76
00:03:04,480 --> 00:03:09,120
it's been a while

77
00:03:06,560 --> 00:03:14,239
but not terribly many

78
00:03:09,120 --> 00:03:17,840
there was a whole 64 kb of memory

79
00:03:14,239 --> 00:03:19,360
and if everything went badly wrong there

80
00:03:17,840 --> 00:03:22,640
was

81
00:03:19,360 --> 00:03:22,640
the floppy disk

82
00:03:23,040 --> 00:03:28,879
so concise

83
00:03:25,200 --> 00:03:30,959
well the 6502 was originally built to be

84
00:03:28,879 --> 00:03:33,760
a

85
00:03:30,959 --> 00:03:36,720
machine controller i i believe

86
00:03:33,760 --> 00:03:38,239
and it turns out that concise did not

87
00:03:36,720 --> 00:03:41,680
include

88
00:03:38,239 --> 00:03:44,159
this instruction

89
00:03:41,680 --> 00:03:46,720
i think it's called mull on modern

90
00:03:44,159 --> 00:03:49,440
machines

91
00:03:46,720 --> 00:03:51,920
uh well

92
00:03:49,440 --> 00:03:55,040
in order to simulate this instruction

93
00:03:51,920 --> 00:03:58,159
people tried the usual approaches first

94
00:03:55,040 --> 00:04:02,159
iterating and adding

95
00:03:58,159 --> 00:04:04,879
but this turned out to be far too slow

96
00:04:02,159 --> 00:04:08,159
so eventually they went with

97
00:04:04,879 --> 00:04:10,159
pre-computing or currying if you want to

98
00:04:08,159 --> 00:04:13,439
use some

99
00:04:10,159 --> 00:04:15,519
i guess slightly gatekeeping terminology

100
00:04:13,439 --> 00:04:16,320
but i guess that's what people do so

101
00:04:15,519 --> 00:04:19,040
they

102
00:04:16,320 --> 00:04:21,120
used an algebraic identity

103
00:04:19,040 --> 00:04:23,600
this is not a class in algebra please

104
00:04:21,120 --> 00:04:25,840
don't stare at this slide and try to try

105
00:04:23,600 --> 00:04:29,199
to read it right now i'll put it up

106
00:04:25,840 --> 00:04:32,240
later anyhow the upshot is that we can

107
00:04:29,199 --> 00:04:35,199
pre-compute some lookup tables about two

108
00:04:32,240 --> 00:04:37,199
kilobytes worth in order to do those

109
00:04:35,199 --> 00:04:38,720
uh multiplies

110
00:04:37,199 --> 00:04:42,080
um

111
00:04:38,720 --> 00:04:44,400
we are using these to avoid

112
00:04:42,080 --> 00:04:46,320
executing some instructions because the

113
00:04:44,400 --> 00:04:48,400
most efficient instruction is the one

114
00:04:46,320 --> 00:04:50,639
you don't execute

115
00:04:48,400 --> 00:04:52,479
and in this case we're doing some

116
00:04:50,639 --> 00:04:54,479
instructions that aren't actually

117
00:04:52,479 --> 00:04:57,199
contained in the

118
00:04:54,479 --> 00:05:00,240
isa

119
00:04:57,199 --> 00:05:04,720
okay so this was kind of a background

120
00:05:00,240 --> 00:05:06,720
thing that i i dimly remembered

121
00:05:04,720 --> 00:05:07,680
years past

122
00:05:06,720 --> 00:05:10,000
um

123
00:05:07,680 --> 00:05:13,520
right about this time there was this

124
00:05:10,000 --> 00:05:16,080
band called akadaka with a new singer

125
00:05:13,520 --> 00:05:18,160
and it becomes famous in the northern

126
00:05:16,080 --> 00:05:20,080
hemisphere

127
00:05:18,160 --> 00:05:22,560
um

128
00:05:20,080 --> 00:05:22,560
also

129
00:05:23,039 --> 00:05:26,720
a

130
00:05:25,039 --> 00:05:28,880
database system

131
00:05:26,720 --> 00:05:31,199
done by some of the people

132
00:05:28,880 --> 00:05:33,039
who had done ingress

133
00:05:31,199 --> 00:05:35,039
was released

134
00:05:33,039 --> 00:05:36,000
it's called postgres

135
00:05:35,039 --> 00:05:39,919
then

136
00:05:36,000 --> 00:05:44,240
a decade later it gets its sql layer and

137
00:05:39,919 --> 00:05:44,240
with it the worst name in the business

138
00:05:46,960 --> 00:05:50,560
data volumes

139
00:05:48,960 --> 00:05:53,039
keep growing

140
00:05:50,560 --> 00:05:55,360
over this time

141
00:05:53,039 --> 00:05:55,360
and

142
00:05:55,440 --> 00:05:59,120
people

143
00:05:56,560 --> 00:06:01,120
get impatient or at any rate start to

144
00:05:59,120 --> 00:06:02,720
express their impatience a little bit

145
00:06:01,120 --> 00:06:06,639
more

146
00:06:02,720 --> 00:06:09,680
and they say things like

147
00:06:06,639 --> 00:06:12,160
beta export is

148
00:06:09,680 --> 00:06:12,160
slow

149
00:06:12,800 --> 00:06:18,000
well it's slow is one of those very

150
00:06:14,560 --> 00:06:21,280
interesting uh challenges in computing

151
00:06:18,000 --> 00:06:23,360
you just never know where to lead

152
00:06:21,280 --> 00:06:26,160
it turned out that

153
00:06:23,360 --> 00:06:27,840
sending integers to standard out

154
00:06:26,160 --> 00:06:31,680
was taking

155
00:06:27,840 --> 00:06:35,120
significant time in these exports

156
00:06:31,680 --> 00:06:36,080
the code looked kind of familiar

157
00:06:35,120 --> 00:06:38,080
um

158
00:06:36,080 --> 00:06:39,280
it had a loop

159
00:06:38,080 --> 00:06:42,240
and

160
00:06:39,280 --> 00:06:45,440
what turns out to be

161
00:06:42,240 --> 00:06:47,199
kind of an expensive operation

162
00:06:45,440 --> 00:06:50,960
at this level

163
00:06:47,199 --> 00:06:52,240
um which was uh which was that little

164
00:06:50,960 --> 00:06:55,199
divide

165
00:06:52,240 --> 00:06:56,400
that you see right after the word value

166
00:06:55,199 --> 00:06:58,240
there

167
00:06:56,400 --> 00:06:59,599
um

168
00:06:58,240 --> 00:07:00,479
and so

169
00:06:59,599 --> 00:07:02,639
i

170
00:07:00,479 --> 00:07:05,120
i got to work

171
00:07:02,639 --> 00:07:07,599
um this kind of looked like something

172
00:07:05,120 --> 00:07:09,599
that i'd seen before and i'd hoped that

173
00:07:07,599 --> 00:07:12,479
i could sort of use

174
00:07:09,599 --> 00:07:14,560
sort of mixture as before which is a

175
00:07:12,479 --> 00:07:17,280
lookup table

176
00:07:14,560 --> 00:07:21,280
um

177
00:07:17,280 --> 00:07:24,080
i applied this to figuring out um how

178
00:07:21,280 --> 00:07:25,759
many decimal digits were in

179
00:07:24,080 --> 00:07:27,840
the number

180
00:07:25,759 --> 00:07:30,400
so that we could avoid some of those

181
00:07:27,840 --> 00:07:33,360
expensive operations have some digit

182
00:07:30,400 --> 00:07:33,360
pairs to hand

183
00:07:34,400 --> 00:07:39,680
the situation

184
00:07:35,919 --> 00:07:39,680
is a little bit different to be fair

185
00:07:41,840 --> 00:07:46,560
the apple 2e had just one core that we

186
00:07:44,879 --> 00:07:48,720
really cared about

187
00:07:46,560 --> 00:07:52,160
i mean there were other processors off

188
00:07:48,720 --> 00:07:53,840
to the side that you could use and and

189
00:07:52,160 --> 00:07:55,360
people did

190
00:07:53,840 --> 00:07:57,199
i just uh

191
00:07:55,360 --> 00:08:00,800
i just wasn't quite that

192
00:07:57,199 --> 00:08:00,800
that advance at that time

193
00:08:01,599 --> 00:08:05,840
and

194
00:08:03,280 --> 00:08:05,840
given the

195
00:08:05,919 --> 00:08:10,960
dearth of cores

196
00:08:08,560 --> 00:08:14,879
it didn't have a branch predictor that

197
00:08:10,960 --> 00:08:14,879
we needed to worry too much about

198
00:08:16,160 --> 00:08:20,960
and so we didn't have to

199
00:08:18,080 --> 00:08:22,639
we didn't have to tiptoe around it

200
00:08:20,960 --> 00:08:25,680
but

201
00:08:22,639 --> 00:08:27,360
we do now and this is one of the things

202
00:08:25,680 --> 00:08:30,639
that

203
00:08:27,360 --> 00:08:30,639
we think about doing

204
00:08:31,440 --> 00:08:34,800
is avoiding that branch predictor or at

205
00:08:33,680 --> 00:08:36,159
any rate

206
00:08:34,800 --> 00:08:37,440
assuaging it

207
00:08:36,159 --> 00:08:40,800
so

208
00:08:37,440 --> 00:08:41,919
with those lookup tables we can

209
00:08:40,800 --> 00:08:45,039
do

210
00:08:41,919 --> 00:08:45,039
branchless computing

211
00:08:45,600 --> 00:08:49,519
or at any rate

212
00:08:47,040 --> 00:08:51,519
limit the kind and limit the no amount

213
00:08:49,519 --> 00:08:53,360
of branching

214
00:08:51,519 --> 00:08:55,760
that we do

215
00:08:53,360 --> 00:08:58,000
so

216
00:08:55,760 --> 00:08:58,000
um

217
00:08:58,800 --> 00:09:03,040
i think there's more that we can do

218
00:09:03,200 --> 00:09:09,040
we should be able to um

219
00:09:06,080 --> 00:09:10,640
to make it faster and for this i would

220
00:09:09,040 --> 00:09:12,160
like to thank

221
00:09:10,640 --> 00:09:15,040
professor

222
00:09:12,160 --> 00:09:18,080
daniel lumiere who's out of

223
00:09:15,040 --> 00:09:18,080
eastern canada

224
00:09:18,320 --> 00:09:22,399
his

225
00:09:19,760 --> 00:09:25,120
his blog is always fascinating on this

226
00:09:22,399 --> 00:09:25,120
kind of score

227
00:09:26,399 --> 00:09:31,279
and

228
00:09:27,600 --> 00:09:31,279
it's uh it it's

229
00:09:32,640 --> 00:09:36,800
some of the future work that i'm her

230
00:09:34,880 --> 00:09:39,519
work that i hope to get

231
00:09:36,800 --> 00:09:41,600
into the postgresql code

232
00:09:39,519 --> 00:09:44,160
is based on things that he's

233
00:09:41,600 --> 00:09:45,519
collaborated with other people

234
00:09:44,160 --> 00:09:46,720
to write

235
00:09:45,519 --> 00:09:49,360
so

236
00:09:46,720 --> 00:09:52,080
we're going to

237
00:09:49,360 --> 00:09:54,160
do more caching or currying or whatever

238
00:09:52,080 --> 00:09:55,760
you want to call the the

239
00:09:54,160 --> 00:09:58,880
process of

240
00:09:55,760 --> 00:10:00,640
computing first holding on to some of

241
00:09:58,880 --> 00:10:02,720
that computation

242
00:10:00,640 --> 00:10:05,839
and then using it later

243
00:10:02,720 --> 00:10:05,839
many many times

244
00:10:06,240 --> 00:10:11,360
so

245
00:10:08,240 --> 00:10:11,360
one of the things that

246
00:10:13,200 --> 00:10:17,360
we can make faster

247
00:10:15,279 --> 00:10:18,640
is

248
00:10:17,360 --> 00:10:21,440
when we have

249
00:10:18,640 --> 00:10:22,880
a binary

250
00:10:21,440 --> 00:10:25,839
integer let's

251
00:10:22,880 --> 00:10:29,680
for simplicity let's say it's

252
00:10:25,839 --> 00:10:29,680
it's an integer and it's unsigned

253
00:10:30,560 --> 00:10:36,000
so

254
00:10:33,200 --> 00:10:38,240
for a 32-bit integer

255
00:10:36,000 --> 00:10:38,240
um

256
00:10:40,160 --> 00:10:42,399
the

257
00:10:43,360 --> 00:10:49,440
binary representation will have its

258
00:10:47,440 --> 00:10:50,480
most significant bit

259
00:10:49,440 --> 00:10:53,279
in

260
00:10:50,480 --> 00:10:54,959
one of 32 places

261
00:10:53,279 --> 00:10:57,519
or one of

262
00:10:54,959 --> 00:11:00,000
64 places depending on the size of the

263
00:10:57,519 --> 00:11:00,000
integer

264
00:11:00,720 --> 00:11:07,839
each of those places

265
00:11:03,600 --> 00:11:09,519
represents a range of numbers that has

266
00:11:07,839 --> 00:11:13,200
at most

267
00:11:09,519 --> 00:11:16,399
two different numbers of decimal digits

268
00:11:13,200 --> 00:11:18,480
in its representation

269
00:11:16,399 --> 00:11:22,079
so for example

270
00:11:18,480 --> 00:11:23,600
a number with its left most one position

271
00:11:22,079 --> 00:11:26,800
four

272
00:11:23,600 --> 00:11:29,120
represents digits from eight which is

273
00:11:26,800 --> 00:11:29,920
one decimal digit

274
00:11:29,120 --> 00:11:33,519
to

275
00:11:29,920 --> 00:11:33,519
15 which is two

276
00:11:38,240 --> 00:11:40,880
so

277
00:11:39,200 --> 00:11:44,079
we can use

278
00:11:40,880 --> 00:11:47,200
a count leading zeros instruction

279
00:11:44,079 --> 00:11:49,279
to find out where that significant one

280
00:11:47,200 --> 00:11:50,800
is

281
00:11:49,279 --> 00:11:51,680
then we use

282
00:11:50,800 --> 00:11:54,240
uh

283
00:11:51,680 --> 00:11:56,959
some cleverness which we need to

284
00:11:54,240 --> 00:12:00,560
document carefully in the code so people

285
00:11:56,959 --> 00:12:00,560
understand what we're doing

286
00:12:01,200 --> 00:12:05,839
we add a magic number from this

287
00:12:04,079 --> 00:12:10,079
short

288
00:12:05,839 --> 00:12:15,279
table that we've we've constructed

289
00:12:10,079 --> 00:12:16,399
that magic number is guaranteed to be

290
00:12:15,279 --> 00:12:18,880
um

291
00:12:16,399 --> 00:12:21,040
larger when added to the number we're

292
00:12:18,880 --> 00:12:22,079
looking at than the original integer

293
00:12:21,040 --> 00:12:26,240
size

294
00:12:22,079 --> 00:12:29,040
so we have this say a 32-bit integer

295
00:12:26,240 --> 00:12:30,079
and then we add this magic number to it

296
00:12:29,040 --> 00:12:33,200
and

297
00:12:30,079 --> 00:12:37,279
what hap what happens is that

298
00:12:33,200 --> 00:12:38,399
um it would it overflows into the next

299
00:12:37,279 --> 00:12:41,760
larger size

300
00:12:38,399 --> 00:12:45,120
so it's no longer inside the lower

301
00:12:41,760 --> 00:12:47,279
32 bits let's say

302
00:12:45,120 --> 00:12:50,079
and when we look at the ones beyond the

303
00:12:47,279 --> 00:12:52,720
width of the original number we get

304
00:12:50,079 --> 00:12:53,680
the number of decimal digits

305
00:12:52,720 --> 00:12:56,000
in the

306
00:12:53,680 --> 00:12:58,880
original number

307
00:12:56,000 --> 00:13:01,440
um so we can do this in

308
00:12:58,880 --> 00:13:02,639
a lot fewer instructions

309
00:13:01,440 --> 00:13:06,320
i think

310
00:13:02,639 --> 00:13:08,560
we can get it down to well i know for 32

311
00:13:06,320 --> 00:13:10,000
bits we can get it down to

312
00:13:08,560 --> 00:13:12,240
one

313
00:13:10,000 --> 00:13:14,560
clz

314
00:13:12,240 --> 00:13:16,880
one table lookup

315
00:13:14,560 --> 00:13:18,839
and i hope to keep that

316
00:13:16,880 --> 00:13:21,120
from

317
00:13:18,839 --> 00:13:23,839
overrunning um

318
00:13:21,120 --> 00:13:26,240
our cache lines or or at any rate from

319
00:13:23,839 --> 00:13:27,680
overrunning them too much

320
00:13:26,240 --> 00:13:28,800
and one add

321
00:13:27,680 --> 00:13:31,200
and one

322
00:13:28,800 --> 00:13:31,200
shift

323
00:13:31,279 --> 00:13:35,440
um

324
00:13:32,320 --> 00:13:36,399
and what we're left with

325
00:13:35,440 --> 00:13:37,360
is

326
00:13:36,399 --> 00:13:39,360
this

327
00:13:37,360 --> 00:13:40,560
little guy

328
00:13:39,360 --> 00:13:42,720
which

329
00:13:40,560 --> 00:13:45,199
um that's the

330
00:13:42,720 --> 00:13:48,160
that's what the new code looks like and

331
00:13:45,199 --> 00:13:49,519
uh yes it's a little bit confusing and

332
00:13:48,160 --> 00:13:50,880
this is where

333
00:13:49,519 --> 00:13:53,360
uh

334
00:13:50,880 --> 00:13:55,600
we're a facility and write writing

335
00:13:53,360 --> 00:13:58,560
english and then getting good feedback

336
00:13:55,600 --> 00:14:01,600
on how that's coming across

337
00:13:58,560 --> 00:14:02,399
will really help to to do this this is

338
00:14:01,600 --> 00:14:04,399
uh

339
00:14:02,399 --> 00:14:06,720
this is not the kind of thing that you

340
00:14:04,399 --> 00:14:10,079
want to have sitting there on its own

341
00:14:06,720 --> 00:14:12,399
the way i have it here on the screen

342
00:14:10,079 --> 00:14:12,399
um

343
00:14:12,639 --> 00:14:17,199
so

344
00:14:14,079 --> 00:14:19,199
can we get it faster

345
00:14:17,199 --> 00:14:22,320
i don't know for sure

346
00:14:19,199 --> 00:14:25,040
but i suspect that it's possible

347
00:14:22,320 --> 00:14:27,360
and then there's other things that are

348
00:14:25,040 --> 00:14:30,480
uh a part of

349
00:14:27,360 --> 00:14:30,480
printing decimal

350
00:14:30,880 --> 00:14:35,360
uh printing decimal representations to

351
00:14:33,440 --> 00:14:37,199
standard output that i haven't covered

352
00:14:35,360 --> 00:14:38,560
here

353
00:14:37,199 --> 00:14:39,839
and i would

354
00:14:38,560 --> 00:14:41,760
love to

355
00:14:39,839 --> 00:14:44,800
work with some other people to figure

356
00:14:41,760 --> 00:14:47,519
out what uh where where best to to shave

357
00:14:44,800 --> 00:14:51,519
that back and to make the

358
00:14:47,519 --> 00:14:54,800
the user experience um

359
00:14:51,519 --> 00:14:56,000
helpful for for uh for all our users out

360
00:14:54,800 --> 00:14:58,720
in the world

361
00:14:56,000 --> 00:14:59,839
i would love your input

362
00:14:58,720 --> 00:15:02,560
uh

363
00:14:59,839 --> 00:15:05,760
with um

364
00:15:02,560 --> 00:15:08,480
as i was doing this i was thinking that

365
00:15:05,760 --> 00:15:10,560
you know postgres isn't all that special

366
00:15:08,480 --> 00:15:12,639
i mean i've been working on it for some

367
00:15:10,560 --> 00:15:14,959
time but

368
00:15:12,639 --> 00:15:17,360
maybe this shouldn't be bespoke code for

369
00:15:14,959 --> 00:15:20,000
that and i would love to figure out ways

370
00:15:17,360 --> 00:15:23,199
to get this into

371
00:15:20,000 --> 00:15:24,639
glib c or other

372
00:15:23,199 --> 00:15:26,880
lib cs

373
00:15:24,639 --> 00:15:28,720
that are upstream

374
00:15:26,880 --> 00:15:31,120
of where i am

375
00:15:28,720 --> 00:15:35,120
and

376
00:15:31,120 --> 00:15:37,440
oh my goodness i had time this for uh

377
00:15:35,120 --> 00:15:40,079
more interaction and questions and stuff

378
00:15:37,440 --> 00:15:44,800
so i guess i'll

379
00:15:40,079 --> 00:15:44,800
i guess i'll i guess we'll do those now

380
00:15:50,480 --> 00:15:54,079
hey thanks for the great presentation i

381
00:15:52,480 --> 00:15:56,240
haven't had any questions come through

382
00:15:54,079 --> 00:15:57,759
in the chat line so if you are thinking

383
00:15:56,240 --> 00:16:00,399
of a question please bring them through

384
00:15:57,759 --> 00:16:01,839
um i was reminiscing the apple 2e at the

385
00:16:00,399 --> 00:16:03,920
front and

386
00:16:01,839 --> 00:16:06,720
i will confess that i actually covered

387
00:16:03,920 --> 00:16:08,240
it in apple ii gs when they first came

388
00:16:06,720 --> 00:16:09,360
out but we couldn't afford one of those

389
00:16:08,240 --> 00:16:12,240
when i was in

390
00:16:09,360 --> 00:16:15,440
senior school so i ended up on a x86

391
00:16:12,240 --> 00:16:17,440
secondhand xx86 compatible hyundai

392
00:16:15,440 --> 00:16:19,360
with a lovely orange screen and then we

393
00:16:17,440 --> 00:16:22,230
got the paper white but yeah

394
00:16:19,360 --> 00:16:23,680
some good nostalgia for me and

395
00:16:22,230 --> 00:16:25,680
[Music]

396
00:16:23,680 --> 00:16:27,199
i lost my date to introduction to

397
00:16:25,680 --> 00:16:28,639
database books at some time i long

398
00:16:27,199 --> 00:16:31,680
across the line but yeah a little bit of

399
00:16:28,639 --> 00:16:34,160
sql still in my hair there somewhere

400
00:16:31,680 --> 00:16:36,880
yeah well it turns out that

401
00:16:34,160 --> 00:16:38,720
all these sql things are i mean they

402
00:16:36,880 --> 00:16:39,519
they kind of hide the

403
00:16:38,720 --> 00:16:42,000
the

404
00:16:39,519 --> 00:16:43,920
underlying sort of

405
00:16:42,000 --> 00:16:46,880
regular computer science that you think

406
00:16:43,920 --> 00:16:47,759
of the data structures the algorithms

407
00:16:46,880 --> 00:16:50,800
the

408
00:16:47,759 --> 00:16:50,800
you know parsing

409
00:16:51,120 --> 00:16:56,160
execution engines compilers

410
00:16:54,480 --> 00:16:58,160
they they kind of

411
00:16:56,160 --> 00:16:59,680
they kind of

412
00:16:58,160 --> 00:17:02,240
help you pretend that those things

413
00:16:59,680 --> 00:17:04,079
aren't still there

414
00:17:02,240 --> 00:17:07,120
but they're there and and if you want to

415
00:17:04,079 --> 00:17:08,319
dig there they're pretty easy to find

416
00:17:07,120 --> 00:17:10,000
wonderful we've actually had two

417
00:17:08,319 --> 00:17:12,720
questions come through now so the first

418
00:17:10,000 --> 00:17:15,520
one is maybe the problem with

419
00:17:12,720 --> 00:17:17,679
sd out or standard out would be

420
00:17:15,520 --> 00:17:19,600
maybe there's a problem with std out

421
00:17:17,679 --> 00:17:20,559
would a socket be better

422
00:17:19,600 --> 00:17:22,319
uh

423
00:17:20,559 --> 00:17:24,400
so

424
00:17:22,319 --> 00:17:25,520
probably

425
00:17:24,400 --> 00:17:28,400
um

426
00:17:25,520 --> 00:17:31,280
that that but um

427
00:17:28,400 --> 00:17:35,679
since i found an improvement in

428
00:17:31,280 --> 00:17:35,679
actually changing the code that that

429
00:17:38,080 --> 00:17:42,320
that does the computation

430
00:17:40,880 --> 00:17:44,880
that that actually

431
00:17:42,320 --> 00:17:48,799
does the instructions i suspect that

432
00:17:44,880 --> 00:17:51,120
it's not all standard out and and so

433
00:17:48,799 --> 00:17:53,679
um sockets might not

434
00:17:51,120 --> 00:17:54,880
uh fix this

435
00:17:53,679 --> 00:17:58,160
because we're

436
00:17:54,880 --> 00:18:00,400
we're generally doing an awful lot of

437
00:17:58,160 --> 00:18:04,720
these things they're they're they tend

438
00:18:00,400 --> 00:18:04,720
to be sort of a dump of a database or

439
00:18:05,840 --> 00:18:08,000
the

440
00:18:09,120 --> 00:18:12,320
setting up a new replica somewhere and

441
00:18:11,520 --> 00:18:15,440
so

442
00:18:12,320 --> 00:18:16,240
huge volumes of data are going through

443
00:18:15,440 --> 00:18:18,240
um

444
00:18:16,240 --> 00:18:20,720
so i don't know that it would

445
00:18:18,240 --> 00:18:23,919
i i thought that sockets would help you

446
00:18:20,720 --> 00:18:26,559
more with a sort of latency issue

447
00:18:23,919 --> 00:18:29,120
um maybe i'm maybe i'm mistaken about

448
00:18:26,559 --> 00:18:30,640
that this is more of a throughput issue

449
00:18:29,120 --> 00:18:33,280
um

450
00:18:30,640 --> 00:18:34,000
let's see benchmarks um i

451
00:18:33,280 --> 00:18:36,240
do

452
00:18:34,000 --> 00:18:38,000
i vaguely recall i got about five

453
00:18:36,240 --> 00:18:39,280
percent out of this which was kind of

454
00:18:38,000 --> 00:18:40,240
amazing

455
00:18:39,280 --> 00:18:43,120
um

456
00:18:40,240 --> 00:18:44,880
the let's see there

457
00:18:43,120 --> 00:18:48,240
might have trouble getting it upstreamed

458
00:18:44,880 --> 00:18:51,440
in g lipsy well uh i have not tried it

459
00:18:48,240 --> 00:18:51,440
yet so i don't know

460
00:18:52,480 --> 00:18:57,919
so i'll i guess i'll find out um but i

461
00:18:55,919 --> 00:19:00,799
i'd uh i'd like to get something that's

462
00:18:57,919 --> 00:19:02,799
a little bit um

463
00:19:00,799 --> 00:19:06,240
more sort of fully

464
00:19:02,799 --> 00:19:06,240
fully baked before i do that

465
00:19:06,640 --> 00:19:12,400
because i yeah i don't know that this is

466
00:19:09,520 --> 00:19:15,120
specialty code it's just more efficient

467
00:19:12,400 --> 00:19:18,400
and i also want to make sure that it is

468
00:19:15,120 --> 00:19:21,120
more efficient on things that are not

469
00:19:18,400 --> 00:19:22,320
intel 64-bit

470
00:19:21,120 --> 00:19:24,400
because of course that would be

471
00:19:22,320 --> 00:19:26,799
important in in the upstream

472
00:19:24,400 --> 00:19:30,480
of c's

473
00:19:26,799 --> 00:19:31,760
a table of magic numbers and more about

474
00:19:30,480 --> 00:19:36,160
where their

475
00:19:31,760 --> 00:19:36,160
where their values derive um

476
00:19:36,240 --> 00:19:39,840
let me see if i can

477
00:19:40,799 --> 00:19:46,799
hold that up here they're basically the

478
00:19:43,760 --> 00:19:50,080
differences between 2 to the

479
00:19:46,799 --> 00:19:51,440
n and 10 to the m

480
00:19:50,080 --> 00:19:54,640
um

481
00:19:51,440 --> 00:19:57,440
and they they're the the property is

482
00:19:54,640 --> 00:19:58,799
that for the largest

483
00:19:57,440 --> 00:20:00,880
um

484
00:19:58,799 --> 00:20:02,080
uh power of 10

485
00:20:00,880 --> 00:20:03,039
they will

486
00:20:02,080 --> 00:20:05,200
um

487
00:20:03,039 --> 00:20:06,640
they will add up to

488
00:20:05,200 --> 00:20:09,760
um

489
00:20:06,640 --> 00:20:12,640
one number and for uh the nine the all

490
00:20:09,760 --> 00:20:16,480
nines number below it they will add up

491
00:20:12,640 --> 00:20:18,720
to they they will they will make the the

492
00:20:16,480 --> 00:20:21,360
higher

493
00:20:18,720 --> 00:20:22,559
part of the the sum

494
00:20:21,360 --> 00:20:24,640
um

495
00:20:22,559 --> 00:20:25,760
a different number so that's the that's

496
00:20:24,640 --> 00:20:28,080
where those

497
00:20:25,760 --> 00:20:29,280
that that's where the magic is

498
00:20:28,080 --> 00:20:32,320
um

499
00:20:29,280 --> 00:20:34,400
i don't i'm not pulling it up right here

500
00:20:32,320 --> 00:20:35,919
but i can send it

501
00:20:34,400 --> 00:20:37,840
in

502
00:20:35,919 --> 00:20:39,679
the the

503
00:20:37,840 --> 00:20:42,559
pr that i had

504
00:20:39,679 --> 00:20:44,880
um originally sent for postgres 14 which

505
00:20:42,559 --> 00:20:47,200
didn't make it because i still needed to

506
00:20:44,880 --> 00:20:50,400
do some work

507
00:20:47,200 --> 00:20:53,760
uh was there something in my current

508
00:20:50,400 --> 00:20:58,240
domain that prompted this research

509
00:20:53,760 --> 00:21:02,400
um well i i wanted to to see uh to to

510
00:20:58,240 --> 00:21:05,200
improve my uh my c skills and i from the

511
00:21:02,400 --> 00:21:05,200
bottom would help

512
00:21:06,159 --> 00:21:09,679
i i realized that

513
00:21:11,919 --> 00:21:16,880
that

514
00:21:13,039 --> 00:21:19,360
making exports and maybe someday imports

515
00:21:16,880 --> 00:21:21,440
a little faster would have the

516
00:21:19,360 --> 00:21:22,799
largest impact on the largest number of

517
00:21:21,440 --> 00:21:24,080
people

518
00:21:22,799 --> 00:21:25,919
the uh

519
00:21:24,080 --> 00:21:29,039
size of the array this is one of those

520
00:21:25,919 --> 00:21:33,280
things that i think may be improvable

521
00:21:29,039 --> 00:21:36,559
um because right now i have uh 32

522
00:21:33,280 --> 00:21:38,400
entries in there for 32-bit

523
00:21:36,559 --> 00:21:41,039
um

524
00:21:38,400 --> 00:21:43,440
numbers for calculating those those uh

525
00:21:41,039 --> 00:21:46,080
those law of integer log tens i guess

526
00:21:43,440 --> 00:21:46,080
you could call them

527
00:21:46,799 --> 00:21:51,760
and so i think it i should be able to

528
00:21:50,480 --> 00:21:55,120
get

529
00:21:51,760 --> 00:21:55,120
whatever it is uh

530
00:21:55,919 --> 00:22:01,360
whatever the size is in bits divided by

531
00:21:59,200 --> 00:22:04,400
four i think because that's

532
00:22:01,360 --> 00:22:05,840
that's the that's the largest number of

533
00:22:04,400 --> 00:22:09,120
uh

534
00:22:05,840 --> 00:22:10,400
base 10 digits that are sorry each

535
00:22:09,120 --> 00:22:12,559
um

536
00:22:10,400 --> 00:22:13,520
each set of four bits can have at most

537
00:22:12,559 --> 00:22:16,159
one

538
00:22:13,520 --> 00:22:18,240
transition or what well is guaranteed to

539
00:22:16,159 --> 00:22:20,480
have one transition between one number

540
00:22:18,240 --> 00:22:21,440
of decimal digits and another

541
00:22:20,480 --> 00:22:24,840
so

542
00:22:21,440 --> 00:22:27,840
um so i think it should be possible to

543
00:22:24,840 --> 00:22:28,960
get something that uses

544
00:22:27,840 --> 00:22:31,200
um

545
00:22:28,960 --> 00:22:34,640
n by four

546
00:22:31,200 --> 00:22:37,039
uh bits or somewhere pretty close to it

547
00:22:34,640 --> 00:22:39,840
sorry n by four entries or something

548
00:22:37,039 --> 00:22:42,240
pretty close to it for uh an end it

549
00:22:39,840 --> 00:22:43,440
integer

550
00:22:42,240 --> 00:22:45,440
um

551
00:22:43,440 --> 00:22:46,720
thanks david i'm happy to read the

552
00:22:45,440 --> 00:22:48,240
questions out if you like i'll give you

553
00:22:46,720 --> 00:22:50,000
a few more seconds to think about them

554
00:22:48,240 --> 00:22:51,760
you will use the politicians trick read

555
00:22:50,000 --> 00:22:54,080
the question and then repeat it back see

556
00:22:51,760 --> 00:22:56,559
more thinking time

557
00:22:54,080 --> 00:22:59,440
apologies if i pronounce anything wrong

558
00:22:56,559 --> 00:23:02,000
how big is the array instructions can

559
00:22:59,440 --> 00:23:04,640
provide div and mod in one

560
00:23:02,000 --> 00:23:06,000
saving multiply by 10

561
00:23:04,640 --> 00:23:09,360
uh

562
00:23:06,000 --> 00:23:10,880
yeah i sorry i thought i was um i

563
00:23:09,360 --> 00:23:11,679
thought i was

564
00:23:10,880 --> 00:23:14,880
you know

565
00:23:11,679 --> 00:23:15,840
speaking to that so basically the

566
00:23:14,880 --> 00:23:18,080
the

567
00:23:15,840 --> 00:23:19,679
the array

568
00:23:18,080 --> 00:23:20,720
is right now

569
00:23:19,679 --> 00:23:23,440
the

570
00:23:20,720 --> 00:23:25,840
uh the um

571
00:23:23,440 --> 00:23:27,919
a number of elements which is the number

572
00:23:25,840 --> 00:23:32,480
of bits in the

573
00:23:27,919 --> 00:23:36,159
uh number so for 32-bit numbers it has

574
00:23:32,480 --> 00:23:38,480
32 entries each of which is an n64

575
00:23:36,159 --> 00:23:41,600
and this is all this is all done with

576
00:23:38,480 --> 00:23:44,480
basically plus count leading zeros and

577
00:23:41,600 --> 00:23:45,840
uh shift

578
00:23:44,480 --> 00:23:47,840
so

579
00:23:45,840 --> 00:23:50,880
maybe there's something more efficient

580
00:23:47,840 --> 00:23:52,559
that takes fewer clocks that's certainly

581
00:23:50,880 --> 00:23:54,559
possible

582
00:23:52,559 --> 00:23:56,640
excellent thank you the next the next

583
00:23:54,559 --> 00:23:58,480
question we've got is there a link to

584
00:23:56,640 --> 00:24:00,559
the limia post

585
00:23:58,480 --> 00:24:01,440
that motivated the improvement

586
00:24:00,559 --> 00:24:04,030
uh

587
00:24:01,440 --> 00:24:05,760
let me see if i can

588
00:24:04,030 --> 00:24:09,799
[Music]

589
00:24:05,760 --> 00:24:09,799
quickly find that

590
00:24:22,080 --> 00:24:26,000
here we go

591
00:24:23,200 --> 00:24:28,240
[Music]

592
00:24:26,000 --> 00:24:28,240
um

593
00:24:31,440 --> 00:24:34,960
so that was

594
00:24:38,400 --> 00:24:44,720
sorry um

595
00:24:40,480 --> 00:24:44,720
let's see if i dump this in

596
00:24:44,840 --> 00:24:50,960
the chat here with that help

597
00:24:49,279 --> 00:24:53,200
i think there may be

598
00:24:50,960 --> 00:24:55,520
there may be a follow yeah there there

599
00:24:53,200 --> 00:24:56,960
may be a follow-up one

600
00:24:55,520 --> 00:24:57,760
um

601
00:24:56,960 --> 00:25:00,320
that

602
00:24:57,760 --> 00:25:02,480
is more recent no i think this is yeah i

603
00:25:00,320 --> 00:25:06,480
think this is the most recent one

604
00:25:02,480 --> 00:25:07,679
um so yeah so so this uh

605
00:25:06,480 --> 00:25:08,720
so uh

606
00:25:07,679 --> 00:25:10,960
um

607
00:25:08,720 --> 00:25:13,440
rare in my experience of academics

608
00:25:10,960 --> 00:25:16,000
anyway uh

609
00:25:13,440 --> 00:25:19,120
professor lemieux

610
00:25:16,000 --> 00:25:19,120
tends to get

611
00:25:20,320 --> 00:25:26,559
really into things as they actually

612
00:25:22,960 --> 00:25:26,559
execute on actual machines

613
00:25:26,960 --> 00:25:30,159
and then test them and then you know

614
00:25:29,039 --> 00:25:31,919
publish

615
00:25:30,159 --> 00:25:33,360
things that even somebody like me can

616
00:25:31,919 --> 00:25:35,120
follow

617
00:25:33,360 --> 00:25:38,320
good

618
00:25:35,120 --> 00:25:38,320
there's hope for me yet then

619
00:25:38,640 --> 00:25:41,279
excellent thank you that's all the

620
00:25:40,080 --> 00:25:43,679
questions we've got that has come

621
00:25:41,279 --> 00:25:45,760
through so far so are you um happy to

622
00:25:43,679 --> 00:25:47,760
jump across to the chat channel post the

623
00:25:45,760 --> 00:25:49,440
the meeting room of course

624
00:25:47,760 --> 00:25:51,919
excellent so if people can

625
00:25:49,440 --> 00:25:53,279
jump over to the chatting channel post

626
00:25:51,919 --> 00:25:54,799
this conference

627
00:25:53,279 --> 00:25:56,880
we're scheduled to finish in about

628
00:25:54,799 --> 00:25:58,880
another 20 minutes maybe people can have

629
00:25:56,880 --> 00:26:00,799
a long afternoon tea or maybe there's

630
00:25:58,880 --> 00:26:02,559
that tricky question you can't quite

631
00:26:00,799 --> 00:26:04,720
word at the moment and you need a few

632
00:26:02,559 --> 00:26:06,559
more to think about it so thank you so

633
00:26:04,720 --> 00:26:08,400
much for joining us from the other side

634
00:26:06,559 --> 00:26:09,520
of the planet um your evening our

635
00:26:08,400 --> 00:26:10,559
afternoon

636
00:26:09,520 --> 00:26:12,240
and

637
00:26:10,559 --> 00:26:14,000
for those of you who are watching the

638
00:26:12,240 --> 00:26:15,520
recording we didn't have any surprise

639
00:26:14,000 --> 00:26:17,520
visitors from david family that were

640
00:26:15,520 --> 00:26:19,360
threatening to run through and surprise

641
00:26:17,520 --> 00:26:21,200
him in the middle of the broadcast so

642
00:26:19,360 --> 00:26:22,640
please do thank your family for taking

643
00:26:21,200 --> 00:26:24,320
time out from your family time this

644
00:26:22,640 --> 00:26:25,919
evening

645
00:26:24,320 --> 00:26:30,360
all right have a good evening

646
00:26:25,919 --> 00:26:30,360
see everybody back after afternoon team