WEBVTT

1
00:06:34.978 --> 00:06:39.689
Oh, okay, good afternoon. People.

2
00:06:39.689 --> 00:06:46.019
Parallel class 30th.

3
00:06:46.019 --> 00:06:52.559
2021.

4
00:06:52.559 --> 00:06:59.819
So, the usual a question Connor waiting.

5
00:07:00.838 --> 00:07:07.949
To get in. Okay, you can hear me. Can you here?

6
00:07:10.769 --> 00:07:15.869
Getting an echo.

7
00:07:17.668 --> 00:07:20.879
So.

8
00:07:22.619 --> 00:07:26.218
Question is.

9
00:07:28.649 --> 00:07:32.699
Is curious, can you also see my screen or not?

10
00:07:34.048 --> 00:07:38.579
Screen is still loading. Okay.

11
00:07:38.579 --> 00:07:44.428
Well, I'll describe what's happening today is our.

12
00:07:45.509 --> 00:07:59.999
Product presentations points about the paper format if you're going to if you're project sharing prior work, then tell me what's new what you're sharing and what's new.

13
00:07:59.999 --> 00:08:12.718
And I got some inspiration inspiration for you sharing your term projects, underhanded see, contest, obfuscate, ITC contests and so on and examples of failures that finally turned into successes.

14
00:08:12.718 --> 00:08:21.059
Some of how I productive program stuff. I've also got some blurbs on the site about excess scale computing.

15
00:08:21.059 --> 00:08:26.908
More stuff, if you have time how you're welcome to look at about C. plus plus.

16
00:08:26.908 --> 00:08:35.458
It's going into parallelism and so on. So if you have time, you might want to have fun for that. I have a link to the Super computing conference.

17
00:08:35.458 --> 00:08:45.899
Um, which is relevant to parallel computing and linkage act on. He's the person who maintains the top 500 list, for example, and got a couple of.

18
00:08:45.899 --> 00:08:49.499
He gives a talk and emerging technologies and so on.

19
00:08:50.693 --> 00:09:05.303
The next thing is the core survey in this course I tried to be a curator, I tried to select the best stuff relating to parallel computing and present it to you. I've created this. Course, in fact, I asked to be given permission to create the course.

20
00:09:05.724 --> 00:09:18.833
And so if you like the course if you thought it was worth your time intuition dollars then say so, on the course survey. I mean, most courses a very small percentage of the students fill up the survey if a larger percentage fill it out and thank you.

21
00:09:19.139 --> 00:09:28.589
And after the semester, I'm open to questions and discussions about any legal ethical topic after you graduate.

22
00:09:28.589 --> 00:09:32.399
So, does it have to be about parallel of computing.

23
00:09:32.399 --> 00:09:44.639
So, now it's time for your presentations and 1st, Ben and Jack and then Blaine Conner Dan, and I just did it in alphabetical order.

24
00:09:44.639 --> 00:09:53.639
So, Ben and Jack, and maybe you might have better luck than me about sharing your screen. So, give it a try. Now.

25
00:09:53.639 --> 00:10:02.278
And see, what happens. All right. Hello? Just want to Mike check real quick.

26
00:10:02.278 --> 00:10:07.649
You hear me, thank you.

27
00:10:08.668 --> 00:10:15.808
You guys also hear me I just want to make sure as well. Yeah, I can hear you. Great low volume, but yes.

28
00:10:15.808 --> 00:10:20.698
Okay, and not see your screen now. Oh, doing.

29
00:10:20.698 --> 00:10:24.328
Good I can see your screen. You got better luck than me.

30
00:10:24.328 --> 00:10:27.418
All right, so.

31
00:10:27.418 --> 00:10:33.058
For our project here, we're taking a look at micro structures. This was a.

32
00:10:33.058 --> 00:10:36.359
Combined project with my capstone.

33
00:10:36.359 --> 00:10:46.558
Project so, for some background inspiration yeah, this was a capstone project that I was a part of basically we were tasked with.

34
00:10:46.558 --> 00:10:50.278
Creating a web app slash web interface for.

35
00:10:50.278 --> 00:10:55.019
Primarily for engineering, 1600 students, so that they could use this.

36
00:10:55.019 --> 00:11:01.619
Learn some material science concepts in a way that they wouldn't otherwise be able to with the textbook and.

37
00:11:01.619 --> 00:11:04.918
The main thing here was that.

38
00:11:04.918 --> 00:11:08.639
A sponsor for the project who was a materials professor.

39
00:11:08.639 --> 00:11:18.119
Gave us an algorithm to generate to run a certain type of simulation and what students would be able to learn here or what they'd be able to see here is.

40
00:11:18.119 --> 00:11:21.359
Like, a dynamic picture changing.

41
00:11:21.359 --> 00:11:24.719
So, there's a time element here, something that they wouldn't get.

42
00:11:24.719 --> 00:11:29.399
In a textbook, um, and so some things to note, it's.

43
00:11:29.399 --> 00:11:34.048
I wrote that algorithm is very intensive. I should rather say that the.

44
00:11:34.048 --> 00:11:38.879
Implementation on our live servers very CPU intensive.

45
00:11:38.879 --> 00:11:44.129
Um, and multiple clients can easily overload this, um.

46
00:11:44.129 --> 00:11:52.198
I posted the link to it if you guys are on campus, or in our network, you can actually go to this link and see our project.

47
00:11:52.198 --> 00:11:57.958
I guess just, for the sake of time, I have it up here. So basically.

48
00:11:57.958 --> 00:12:01.198
This image right here is.

49
00:12:01.198 --> 00:12:07.139
A structure, this is what you would see under a microscope if looking at a.

50
00:12:07.139 --> 00:12:10.408
Material or a composition of materials.

51
00:12:10.408 --> 00:12:15.089
Changing transitioning from a liquid phase to a solid phase. So.

52
00:12:15.089 --> 00:12:23.999
Light blue being the liquid phase, and the darker blue, being the solid phase solid blobs. So you can play around.

53
00:12:25.109 --> 00:12:28.288
With some variables here and.

54
00:12:28.288 --> 00:12:32.188
Change a see how it relates to a phase diagram. This is another.

55
00:12:32.188 --> 00:12:37.979
Materials concept that they're stressing to teach, but this is the main focus point of our project.

56
00:12:37.979 --> 00:12:41.639
So, yeah, you can play around with these.

57
00:12:41.639 --> 00:12:45.418
See, this picture change correspondingly.

58
00:12:45.418 --> 00:12:49.859
So, let's get back to this.

59
00:12:49.859 --> 00:12:54.058
Yeah, I just kind of covered some of this, but yeah, microstructure is a.

60
00:12:54.058 --> 00:12:59.399
Representation of a solid structure under a microscope so you can see each 1 of these little.

61
00:12:59.399 --> 00:13:02.548
Locks if you will are called grains.

62
00:13:03.568 --> 00:13:10.168
And, yeah, so we took this algorithm and we tried to make it parallel.

63
00:13:10.168 --> 00:13:13.828
I will, I guess, give it over to ban you can talk about.

64
00:13:13.828 --> 00:13:18.058
Some of the specifics technical details.

65
00:13:18.058 --> 00:13:25.589
Sure, yeah, so the main tool we decide to use was 2 Pi.

66
00:13:25.589 --> 00:13:37.528
The microstructure program already used dump pie, which has some limited CPU parallelization, but it wasn't too expensive. It wasn't a huge speed up. So.

67
00:13:37.528 --> 00:13:46.229
We thought that would be even better. So coupon uses and could a platform.

68
00:13:46.229 --> 00:13:50.188
Um, you know, which is, you know, stuff that we've done.

69
00:13:50.188 --> 00:13:54.509
So, we took the your program.

70
00:13:54.509 --> 00:13:58.168
Um, and modified it to take advantage.

71
00:13:58.168 --> 00:14:01.379
Coupon the platform.

72
00:14:01.379 --> 00:14:04.379
So, I want to go to the next slide.

73
00:14:05.698 --> 00:14:08.969
So, here's just a little bit of the code. This is like.

74
00:14:08.969 --> 00:14:12.958
Some of the meat of a lot of where we modify it and.

75
00:14:12.958 --> 00:14:17.818
Chain stuff to use the library.

76
00:14:17.818 --> 00:14:22.288
Live a visual here we can go to the next 1.

77
00:14:25.198 --> 00:14:31.229
So, on smaller simulations, there wasn't a whole lot of time improvement.

78
00:14:31.229 --> 00:14:34.859
But where parallelization.

79
00:14:34.859 --> 00:14:39.208
And optimization really shines as the larger scale.

80
00:14:39.208 --> 00:14:42.328
Um, applications, so we found that.

81
00:14:42.328 --> 00:14:46.558
When we increase the 2 sizes of, you know.

82
00:14:46.558 --> 00:14:52.589
A 1000 by 5,000 by 4,000 pixel wise simulations improvements got.

83
00:14:52.589 --> 00:14:55.739
Much larger than it did with.

84
00:14:55.739 --> 00:14:59.489
Much faster than it was with numb pie and with.

85
00:14:59.489 --> 00:15:06.658
Uh, yeah, so the original.

86
00:15:06.658 --> 00:15:15.448
Simulation ran about 256 by 256 and at that level, it wasn't a whole lot of change, but at the level of 4,000 by 4,000, it got.

87
00:15:15.448 --> 00:15:22.198
Be much greater change, so, got a table on the next slide to show some of that.

88
00:15:22.198 --> 00:15:26.428
Um, so, yeah, point, you know.

89
00:15:26.428 --> 00:15:32.668
Point 01 seconds and point 03 seconds that the big of a deal, but at a size of 5,000 by 5,000.

90
00:15:32.668 --> 00:15:36.149
You know, simulated particles it was much faster.

91
00:15:37.438 --> 00:15:41.818
There's also I just want to know, I think it's important to point out that.

92
00:15:41.818 --> 00:15:48.538
We would never want to generate an image of 4,000 by 4,000, but this was just a way to sort of simulate.

93
00:15:48.538 --> 00:15:53.489
That the best stress test year would be to have multiple concurrent users on it, but.

94
00:15:53.489 --> 00:15:57.178
Generating an image of 16 times the size would.

95
00:15:57.178 --> 00:16:01.078
There will be the equivalent of 16 users, so.

96
00:16:03.149 --> 00:16:07.259
Yeah, so that was a pretty good success.

97
00:16:07.259 --> 00:16:12.538
Up next we've got a little demo to show the difference. Yeah.

98
00:16:12.538 --> 00:16:18.269
So, to set this up, I had a.

99
00:16:21.178 --> 00:16:27.928
Um, I had my Python environment activated here and then 1st, I just want to show you guys the.

100
00:16:27.928 --> 00:16:33.599
Plain numb pie version. This is 1000 by 24.

101
00:16:33.599 --> 00:16:39.538
Obviously pretty large for a computer screen, but again, just want to simulate concurrency.

102
00:16:40.798 --> 00:16:44.548
So, if I move this around each new.

103
00:16:44.548 --> 00:16:49.528
Liquid fraction value that pops up. You'll see a time.

104
00:16:49.528 --> 00:16:52.619
Start to complete for generating the new picture.

105
00:16:52.619 --> 00:16:59.969
It's just sitting right around 2 tenths of a 2nd point 2, 2, 2 1.

106
00:16:59.969 --> 00:17:04.469
So, this isn't very, it's a very useful.

107
00:17:04.469 --> 00:17:10.048
For someone trying to learn through dynamic visualizations is very choppy.

108
00:17:10.048 --> 00:17:16.229
Do you guys are able to see it over Webex but it is pretty choppy and.

109
00:17:16.229 --> 00:17:22.138
Here's the time to prove it, so.

110
00:17:22.138 --> 00:17:25.288
I cleared again and then if I were to run our.

111
00:17:25.288 --> 00:17:29.189
Who Pi version.

112
00:17:29.189 --> 00:17:35.009
Same image size same exact simulation. Really? But now.

113
00:17:35.009 --> 00:17:39.538
Bring these sliders they're are seeing.

114
00:17:39.538 --> 00:17:43.558
About 20 times speed improvement.

115
00:17:43.558 --> 00:17:48.419
There's much better, much lower latency from.

116
00:17:48.419 --> 00:17:52.348
Me sliding this to getting a new image, though. I can see the dynamics.

117
00:17:52.348 --> 00:17:55.828
Very well.

118
00:17:55.828 --> 00:18:01.348
So, yeah, and I think this is an improvement that I wanted to.

119
00:18:02.429 --> 00:18:07.318
Stress for the following semester, my capstone project since, uh.

120
00:18:07.318 --> 00:18:15.929
Right now the live version is only running on a I think it's an 8,700 K and a VM. So there's only a couple of CPU cores.

121
00:18:15.929 --> 00:18:22.769
I guess my CPU implementation was a pretty best case scenario. So I have a pretty.

122
00:18:22.769 --> 00:18:25.888
Fast single core.

123
00:18:25.888 --> 00:18:31.499
But, yeah, I guess we, we can take some questions down. There are any.

124
00:18:31.499 --> 00:18:40.288
That's just about what we have. Well, thank you. Interesting. It's sort of like a morphological dilation operation. It looks like.

125
00:18:40.288 --> 00:18:49.409
Growing the circles yeah, so I wasn't I wasn't quite sure of the exact math that goes into this, but it definitely.

126
00:18:49.409 --> 00:18:52.439
Is a pixel by pixel calculation.

127
00:18:52.439 --> 00:18:56.669
And there's a lot of vector operations. Here's.

128
00:18:56.669 --> 00:19:00.778
It'll accept the random seed, which is basically we'll see the center point.

129
00:19:00.778 --> 00:19:06.749
Of these grains, so I could change the seed and we would get a completely different image, but yet.

130
00:19:06.749 --> 00:19:11.249
Basically, you have a center point for grains.

131
00:19:11.249 --> 00:19:16.288
And you have a growth rate, also random within certain bands and.

132
00:19:16.288 --> 00:19:21.568
Each pixel is calculated hey, which grain will get to me. 1st.

133
00:19:21.568 --> 00:19:25.138
From there, we can create an image.

134
00:19:25.138 --> 00:19:28.739
Okay, I see what you're doing now. Anyone else have any questions.

135
00:19:28.739 --> 00:19:33.898
Did you say you have job opportunities available for people on this.

136
00:19:34.949 --> 00:19:44.729
I didn't, I didn't say that I'm sorry I was seeing if this project were to continue into the following semester. I think this is a.

137
00:19:44.729 --> 00:19:52.019
Important route of implementation. I got it for a classroom environment. Yeah. Everyone's going to be on this.

138
00:19:52.019 --> 00:19:58.409
Thank you other questions. Okay.

139
00:20:00.148 --> 00:20:03.598
Blame you.

140
00:20:03.598 --> 00:20:11.578
I just saw a question pop up. Oh, sorry. Yeah. It's always tough to see it when I'm sharing. So you can just read it. What Python class.

141
00:20:13.259 --> 00:20:18.148
Um, we're using a to draw the window here.

142
00:20:18.148 --> 00:20:25.108
And then from a non Pi array, or I guess in this case, a, we're using pillow.

143
00:20:25.108 --> 00:20:28.739
Pillows image from array method.

144
00:20:28.739 --> 00:20:33.358
To create the image, does that answer your question?

145
00:20:33.358 --> 00:20:39.239
Okay, other questions.

146
00:20:39.239 --> 00:20:43.648
Great. Okay.

147
00:20:43.648 --> 00:20:46.769
Lane, what do you have to tell us.

148
00:20:46.769 --> 00:20:49.769
Can you guys hear me? Yes.

149
00:20:49.769 --> 00:20:52.919
Okay, great. I'll try sharing my screen now.

150
00:20:55.378 --> 00:21:00.689
Working perfectly. Okay. Great. Great. Good to hear.

151
00:21:00.689 --> 00:21:04.499
Uh, yeah, so let me get started then.

152
00:21:04.499 --> 00:21:08.038
So, uh, I did my projects on.

153
00:21:08.038 --> 00:21:14.519
Um, evaluating moves for the board game corridor, using parallel computing so that I could make a.

154
00:21:14.519 --> 00:21:17.759
Very basic that you could play against in the game.

155
00:21:19.108 --> 00:21:25.259
So just an overview for the game, because a lot of this might not make sense if you don't have a visualization of what the game looks like.

156
00:21:25.259 --> 00:21:34.138
Um, it's an eye by 9 boards. There's between each 1 of these boards you can put a 2 by 2 wall piece, which can either be facing up or sideways. So there's.

157
00:21:34.523 --> 00:21:48.413
8, by 8 different places, you can put it in 2 directions each 1 and each player gets 10. well, also there's 20 wells in total. So each round, you can either place a wall or move your player and you're not allowed to entirely cut off the other side. You can't make it impossible for the player.

158
00:21:48.443 --> 00:21:54.983
You just can make it as circuitous as you want and the 1st player to the other side wins alternating rounds of, like.

159
00:21:55.288 --> 00:21:59.429
Being able to either play softball or move or place to move and trying to make it to the other side.

160
00:22:00.509 --> 00:22:09.239
So the way I broke this problem down was, um, into a couple of different, like, main files 1 of them that was just handling, like, um.

161
00:22:09.239 --> 00:22:15.509
The board itself, and the structure that I use for, like, indicating what the, like, moves are.

162
00:22:15.509 --> 00:22:20.669
A path finding algorithms so that you could figure out if it, if the other sites been cut off or not.

163
00:22:20.669 --> 00:22:31.618
Um, game dot H, which handles the valid player moves to figure out like, if I'm here, what can I actually do in terms of being able to place wells or being able to move.

164
00:22:31.618 --> 00:22:38.489
And then finally the actual parallel component of this games research that goes through all of the different.

165
00:22:38.489 --> 00:22:42.838
Possibilities down to a certain depth and tries to.

166
00:22:42.838 --> 00:22:47.669
Make a prediction on what an opponent would do based on how those moves the play apps.

167
00:22:47.669 --> 00:22:59.578
So, the main algorithm here is what's known as many maps. It's a game research algorithm. So here is an example of a game tree with tic TAC toe, just because it's an incredibly simple game.

168
00:22:59.578 --> 00:23:03.088
Where you could imagine, like, given some starting board state.

169
00:23:03.088 --> 00:23:14.848
A certain player could only make, like, a certain number of moves and then in contrast, in response to those, the next player could make a certain number of moves. And in response to that 1, you could make another move and.

170
00:23:14.848 --> 00:23:19.019
Assuming that the game doesn't go on sufficiently long.

171
00:23:19.019 --> 00:23:28.648
You could go through for a simple enough game, all of the possibilities, and just tell what would let you win based on what all possible responses you could do for the other player.

172
00:23:28.648 --> 00:23:34.229
However, uh, that is problematic with games with large search speech and search spaces because, um.

173
00:23:34.229 --> 00:23:37.949
Unlike tic TAC toe, which only has like, this 3 by 3 grid.

174
00:23:39.144 --> 00:23:51.894
There is no possible way to go through all of the different, like, wall possibilities because there's, um, 133 different move possibilities at each Player's position. So it grows exponentially way too fast for it to be controlled.

175
00:23:53.009 --> 00:23:56.249
So, the idea here is that, um.

176
00:23:56.249 --> 00:23:59.459
Using parallel computing break the sub tree down.

177
00:23:59.459 --> 00:24:04.618
Into smaller subseries that can be solved on their own and then have the solutions brought back together.

178
00:24:04.618 --> 00:24:10.019
Because using those parallel computing resources, you can explore more of the train, the same amount of time.

179
00:24:10.019 --> 00:24:15.628
I did this mainly with an algorithm called alphabet recruiting, which lets you.

180
00:24:15.628 --> 00:24:21.328
Trim off or prune branches from the tree that look no rational player would do like.

181
00:24:21.328 --> 00:24:27.388
If all the notes on a certain cyber to lead to you losing or lead to the opponent player.

182
00:24:27.388 --> 00:24:33.028
Having a relative advantage, where, you know that they wouldn't choose that move, because it would let you be better off.

183
00:24:33.028 --> 00:24:37.618
Then you can sort of disregard a whole part of the tree and not have to keep searching in it.

184
00:24:37.618 --> 00:24:44.638
And since it's an algorithm, that's quasi independent from each other. There's a couple of variables that get passed around. It can be paralyzed.

185
00:24:44.638 --> 00:24:48.088
So the main parallelization here was using open and P.

186
00:24:48.088 --> 00:24:55.949
Um, so each, um, open MP thread gots a set of like, its own boards from the starting node. Uh.

187
00:24:55.949 --> 00:25:05.128
So, like, the board is in a current state, and you're current player, and there's a list of options you can do. And each of the open entry threads was handling part of this for loop.

188
00:25:05.128 --> 00:25:11.068
For deciding, I'm like, going through the whole game tree down to a certain depth. Um.

189
00:25:11.068 --> 00:25:16.798
For what they would be able to do from there so that that initial process is split apart.

190
00:25:16.798 --> 00:25:21.598
So, what that would look like, is a parallel section where each thread gets its own copy of the immediate board.

191
00:25:21.598 --> 00:25:30.298
And each 1 is helping to solve through this for loop of all the 1st level moves it could make. And then in its own time calling the century solver.

192
00:25:31.409 --> 00:25:39.534
So there were some complications with this where alpha beta printing also returns a variable that you have to, like, or 2 variables that you have to keep track of that's global to all of them.

193
00:25:39.804 --> 00:25:47.304
Because if you don't keep the variables global, then the centuries can't learn information from the other ones and evaluates like, what would be the best move faster.

194
00:25:47.638 --> 00:25:54.479
But open and works better when we have the least amount of information that must be passed around, like, atomics and criticals because they tend to.

195
00:25:54.479 --> 00:26:06.118
Have a lot of overhead and hold stuff up. So, um, after, like, I initially wrote this up, uh, just to speed it up. I did too many things. I rewrote the algorithm to have for you to fewer critical blocks by, um.

196
00:26:06.118 --> 00:26:10.979
Rewriting it so that there was only an upper bound instead of a lower end upper bound.

197
00:26:10.979 --> 00:26:17.699
And that ended up causing me decent speed increase and, um, adding what's known as a transposition table, which is just a record of, like.

198
00:26:17.699 --> 00:26:26.098
At this board location, what ended up being the result so that multiple trees could talk to each other through that board and.

199
00:26:26.098 --> 00:26:34.409
Um, if 1 of them has already evaluated this possibility, like, there's more than 1 way to get to the same configuration of a board. Like, you could move somewhere and then move back.

200
00:26:34.409 --> 00:26:38.578
But the, you don't have to reevaluate that a 2nd time. If you've already cash the results.

201
00:26:39.719 --> 00:26:46.588
So here are my basic results, uh, for the optimizations I just talked about in my original code for running, like.

202
00:26:46.588 --> 00:26:51.929
With Max parameters each time I added 1 of the optimizations I was matched to.

203
00:26:51.929 --> 00:26:55.078
Save a couple of seconds off the total thing.

204
00:26:55.078 --> 00:27:00.689
And then just in comparison to it, being single threaded after adding each number of thread in open MP.

205
00:27:00.689 --> 00:27:08.489
Starts to give a pretty decent time saving, but, uh, as you get into extremely high numbers of threads, I'm assuming the overhead starts to overtake.

206
00:27:08.489 --> 00:27:22.798
The advantage of having multiple things solving trees so I can also give a demonstration of this. I need a little output and hopefully this will cooperate with me and display but otherwise I have the video.

207
00:27:22.798 --> 00:27:28.558
I'll just have to switch to a different thing. So what we have here is.

208
00:27:28.558 --> 00:27:32.578
On that board, each of these is representing in empty space.

209
00:27:32.578 --> 00:27:35.669
As is the 2nd player he is the primary player.

210
00:27:35.669 --> 00:27:42.388
And as they move around, just me, being the secondary player, playing against the AI.

211
00:27:42.388 --> 00:27:47.848
You can place down different pieces and different walls that block off different sections.

212
00:27:47.848 --> 00:27:53.038
And actually, it doesn't seem that this isn't rendering, right?

213
00:27:54.419 --> 00:28:02.038
Give me a 2nd. Okay this might be a little bit better.

214
00:28:09.868 --> 00:28:13.288
So, your player, the computer player.

215
00:28:14.398 --> 00:28:17.398
You're trying to get to the other side and you can put up walls or move.

216
00:28:17.398 --> 00:28:21.298
So, initially I put down a lot of stuff then, but it only bought me a turn.

217
00:28:21.298 --> 00:28:25.798
And then the AI just sort of slowly boxes me in.

218
00:28:28.858 --> 00:28:32.159
And then when I try to escape, puts down another wall and another wall.

219
00:28:37.078 --> 00:28:44.278
And puts down its own wall so that I can't stop it from going somewhere until eventually it gets to where it's 1 space away from the end. And it's 1.

220
00:28:52.979 --> 00:29:06.689
So, just to get back to where we were, uh, so, uh, some other possible improvements I thought of, although wasn't able to implement in the time for this thing, we're trying to make the open empty structures more efficient. Like, I could try to whittle down the critical section a bit more.

221
00:29:06.689 --> 00:29:10.528
I could catch more information between rounds because when I have it.

222
00:29:10.528 --> 00:29:19.618
Set up to play against an, a, I, I don't cash the information between each time. You try again it restarts the whole thing and throws out the cash each time.

223
00:29:19.618 --> 00:29:27.868
Because I didn't want, um, a lower quality previous estimate to mess up with its new estimates since it's in a new position.

224
00:29:28.223 --> 00:29:37.104
But, it would be much more efficient if I were to figure out a way to make that work. And there are some entirely different algorithms that I had no time to implement for this type of thing. But are absolutely amazing.

225
00:29:37.374 --> 00:29:48.923
Like, um, 1, alternative to Minnie Max and alpha beta for name is the Monte Carlo tree search where you just go down a certain number of moves in the game tree. And then after that, since you couldn't possibly explore them all.

226
00:29:49.229 --> 00:30:03.808
What it does is, it would just randomly go through the rest of the roofs and until it hits the end of the game and then give a probability value at each note of how likely you are to win based, on random movements following the depth. You can explore.

227
00:30:03.808 --> 00:30:07.828
Which is what Google use when they train the alpha going.

228
00:30:07.828 --> 00:30:12.719
It's pretty interesting. So, um, any questions.

229
00:30:14.159 --> 00:30:22.558
Yeah, thank you. Interesting. What the idea is extend to a different game or a bigger game.

230
00:30:22.558 --> 00:30:26.669
Oh, yeah, yeah, totally medium X and alpha beta printing.

231
00:30:26.669 --> 00:30:31.919
Are, uh, pretty generic thing they work for any game, which is why they're usually talked about in, like.

232
00:30:31.919 --> 00:30:35.729
To player game, programming, type stuff.

233
00:30:35.729 --> 00:30:42.898
I had originally got the inspiration from this, because a couple of years back, I had taken interaction to artificial intelligence.

234
00:30:42.898 --> 00:30:49.769
The CS course, and 1 of the things we had done was implement these types of things but for Pac, man.

235
00:30:49.769 --> 00:31:01.769
The PacMan so much more simple game where the movements are much more predictive. So I was wondering to try it out using, like, both parallel computing and, um, a totally different game. So I could write it all from scratch and see what I could do.

236
00:31:04.528 --> 00:31:12.509
Okay, thank you. Anything else. What type of do you use? I didn't use the GPU. I used.

237
00:31:12.509 --> 00:31:18.358
Um, purely the CPU because it got really messy to try to make.

238
00:31:18.358 --> 00:31:26.878
So, like, at each node, there's a lot of different decisions that have to happen in terms of branching and I just couldn't find a way to get around to the.

239
00:31:26.878 --> 00:31:33.568
Like, the, the work problem of if 1 of the things is going in a different direction, you have issues.

240
00:31:33.568 --> 00:31:37.979
And I read a couple of papers for some people managed to get around that.

241
00:31:37.979 --> 00:31:45.598
But nowadays, if people are going to use the GPU, most people would write up the Monte Carlo research because you can use that. They're on that to solve the.

242
00:31:45.598 --> 00:31:48.959
Instead of randomness, continuing on, you can use the.

243
00:31:48.959 --> 00:31:57.179
Averages to train and are on it and then use the neural net to go beyond what you can explore in a treaty.

244
00:31:57.179 --> 00:32:01.739
Okay, I got it. Thank you. Great.

245
00:32:01.739 --> 00:32:10.259
On or you're up next let me let me pull everything.

246
00:32:10.259 --> 00:32:25.169
Okay, let me know when you can see my screen.

247
00:32:26.219 --> 00:32:29.878
Now, job, search, web, scraping through. Indeed dot com.

248
00:32:30.203 --> 00:32:44.993
Wonderful All righty. So hey, everyone my project was focused on developing a personalized job search application to be applied through. Indeed. Com, using web scraping.

249
00:32:45.358 --> 00:32:54.538
So, at this point, we've all had to deal with the pain of going through a job search, whether that be for internships and Co Ops or for.

250
00:32:54.538 --> 00:33:05.759
Full time employment, following graduation and I don't know about you, but I really hate the endless scrolling through different job boards on sites like LinkedIn. Indeed. And glass door.

251
00:33:06.324 --> 00:33:15.953
There are so many jobs posted on these sites every day, which don't get me wrong is great when looking for a position, and it opens up doors to a lot of potential opportunities.

252
00:33:16.223 --> 00:33:21.624
But it seems like finding the right fitting position is like, finding a needle in a haystack.

253
00:33:21.898 --> 00:33:25.288
Like, you can see in the top rate screenshot.

254
00:33:25.288 --> 00:33:39.173
These job boards seek to overcome this large volume of postings by providing different search features and qualifications only return the postings that match the users search specifications and they offer quite a few great search criteria,

255
00:33:39.173 --> 00:33:41.693
such as location experience and salary.

256
00:33:41.993 --> 00:33:51.054
But the problem is that these job boards aren't smart enough to understand the context surrounding the keywords and the qualifications that the user inputs and.

257
00:33:51.358 --> 00:33:55.709
Not all physicians in the same industry, especially in software engineering.

258
00:33:55.709 --> 00:34:01.019
Use the same technologies or techniques. So, 1 position may be a better fit than another.

259
00:34:01.019 --> 00:34:08.338
Uh, even if they both mentioned the same key word so take, for example, the screenshot on the bottom here. Um.

260
00:34:08.338 --> 00:34:16.139
This is the requirement section of 1 of the top positions that comes up on. Indeed. When you search for a software engineer Python.

261
00:34:16.139 --> 00:34:24.478
Um, you can see highlighted in blue that it does mention Python and it meets the search criteria as a result of that.

262
00:34:24.478 --> 00:34:37.284
You're the only problem for me personally, is that I despise doing any sort of web development, and I want to discard any searches that focus on web development and looking at the rest of these requirements in this list.

263
00:34:37.284 --> 00:34:47.423
And specifically the technologies that I've mentioned, uh, looking for extreme proficiency, and such as no react and restful web services.

264
00:34:47.639 --> 00:34:55.134
I can tell you that this position and focuses on web development, but again indeed doesn't know that. Indeed doesn't recognize.

265
00:34:55.134 --> 00:35:03.143
The Python is only mentioned as just other technologies and that it was likely included as just another keyword in the posting.

266
00:35:03.449 --> 00:35:11.429
The same could work the other way around. I wouldn't want to exclude a job posting just because it mentioned Java script somewhere in the listing.

267
00:35:11.429 --> 00:35:17.009
But when web technologies become a large portion of the listed technologies in the description.

268
00:35:17.009 --> 00:35:28.103
It's fair to assume that I'm no longer interested in the position and I have now spend the time reading through the description, and looking at the company just to discard it. And while it may not have taken me very long.

269
00:35:28.103 --> 00:35:34.074
This can happen dozens of times in just a short browsing session. And this time does quickly add up.

270
00:35:34.378 --> 00:35:39.958
So, how do we get around this and only focus on the postings that might actually be a good personal fit.

271
00:35:39.958 --> 00:35:49.228
Well, so my 1st idea was to use web scraping if I could scrape and collect the relevant information and keywords from these job postings.

272
00:35:49.228 --> 00:36:03.623
I could then use that collected information to process and filter the different jobs, using the technical and contextual knowledge that I have above what each job search functionality has for instance I could only show myself the job postings.

273
00:36:03.623 --> 00:36:17.903
That contain less than say, 20%, web, related technologies. And are the positions that meet, say, 75% of the technologies that I want to work with, and that I meet the qualifications for. So, to understand how I needed to go about doing this.

274
00:36:19.318 --> 00:36:26.159
i1st need to do a little research to gain some general knowledge about the scraping and the associated technologies. So.

275
00:36:26.159 --> 00:36:35.909
What is web scraping? Well, a formal definition of web scraping is the process of collecting structured data web data in an automated fashion.

276
00:36:35.909 --> 00:36:43.259
More generally, it's essentially just collecting the publicly available data from the web, or from a website.

277
00:36:43.259 --> 00:36:50.668
And there are a few different ways to do it most general of, which is essentially just straight website.

278
00:36:50.668 --> 00:36:55.409
You take in the website and you scrape or filter out the information that you're interested in.

279
00:36:55.409 --> 00:37:02.128
Say the images are just the headings, um, if you've copied and pasted something from a Web site, you've done.

280
00:37:02.128 --> 00:37:15.778
Some form of scraping, you took your knowledge of the content of the site and the context in which it was presented and you pulled or scraped the relevant definition or other texts that you were interested in. But this process.

281
00:37:15.778 --> 00:37:24.748
Uh, is very manual and gets incredibly slow and it's very time consuming so a web scraping fame framework can help to automate this.

282
00:37:24.748 --> 00:37:33.418
To apply this more specifically to the parallel computing topic of the class. We can look into distributed web crawling web, scraping.

283
00:37:33.418 --> 00:37:38.489
Um, yes, there is a difference between crawling and scraping.

284
00:37:38.489 --> 00:37:49.829
With crawling pretty much, just being indexing the sites. Um, this is pretty much what search engines do is they use massive computing clusters to crawl the connected parts of the Internet.

285
00:37:49.829 --> 00:37:56.938
And this is where the term deep web comes from, the deep web is essentially just the parts of the Internet.

286
00:37:56.938 --> 00:38:01.528
Uh, whereas scraping is more focused on the collection of data from these websites.

287
00:38:01.528 --> 00:38:07.679
So, distributed scraping models, essentially employ many computers to scrape information from websites.

288
00:38:07.679 --> 00:38:20.760
The scraping can be very repetitive and very tedious. Especially when you start looking at similarly formatted websites, or you crawl out to connected sites, for instance, following the links on a web page.

289
00:38:20.760 --> 00:38:30.119
Um, and then trying to scrape that new link that you found, and then repetitively recursively doing that through all of the link sites. Um.

290
00:38:30.385 --> 00:38:33.295
This lends itself super well to parallelization.

291
00:38:33.775 --> 00:38:45.864
So while a distributed model may be slightly overkill for the job search application that I'm specifically implementing as I'm not going to be scraping thousands or hundreds of thousands of websites.

292
00:38:46.139 --> 00:38:51.925
It is still good to look into and it was good to research and tie that back into the class.

293
00:38:52.315 --> 00:39:04.135
So, after having gotten a little bit of general knowledge about scraping and seen a few tutorials about how to get it working, it was time to start looking specifically into what technologies to use to do it.

294
00:39:04.614 --> 00:39:08.335
So, I quickly decided that Python was going to be my go to here.

295
00:39:08.610 --> 00:39:16.170
And started researching different, I'm scraping frameworks for Python and scraping. Is what immediately came up.

296
00:39:16.735 --> 00:39:30.925
From the research I did, it's probably the most popular Python frameworks for web scraping. It is entirely open source, but it's consistently maintained and it's very widely used, which makes it a good and reliable or open source project to use.

297
00:39:31.704 --> 00:39:33.715
It's also very intuitive.

298
00:39:34.019 --> 00:39:42.659
It makes it very easy to get started with web, scraping and crawling, allowing a user to write a functional spider and probably 10 lines of code.

299
00:39:42.659 --> 00:39:46.500
But despite this ease of use, it can also be expanded.

300
00:39:46.500 --> 00:39:56.940
To be applied to very complex crawling jobs and there's also a distributed version called scraped, be distributed that allows for those large scale, computationally demanding jobs to be done.

301
00:39:56.940 --> 00:40:03.030
Granted, the scraper distributed platform is less widely used and less maintained.

302
00:40:04.135 --> 00:40:18.925
While scraping is great on its own. It can also be used in conjunction with web, which is something that I found out later in the project and That'll be important here in a few slides. But this is a big deal, because usually web scraping is done as an alternative to using a web.

303
00:40:19.739 --> 00:40:23.670
So the fact that you can use both really is very appealing.

304
00:40:23.670 --> 00:40:30.239
So, after settling on, I needed to do a little bit of Recon on the website that I'd be scraping.

305
00:40:30.355 --> 00:40:44.364
As any developer, or as any crawler that you develop will be specific to the outline of the format of the associated website and you can't really have a spider that collects information from all kinds of websites, or is a general kind of scraper.

306
00:40:45.264 --> 00:40:49.885
So, for the sake of creating a use case, I focus specifically on indeed dot com.

307
00:40:50.219 --> 00:40:54.150
Uh, through the Recon, I was able to find a few different things, so.

308
00:40:54.150 --> 00:41:07.590
For 1, while each job posting on, indeed includes a description and position outline. There's no set format for these. So the structure quality and amount of content in these descriptions varies from posting to posting.

309
00:41:07.590 --> 00:41:17.844
And the variability made it rather difficult to scrape completely, but I could at least searched for keywords within those descriptions, regardless of that format. Right?

310
00:41:18.715 --> 00:41:29.635
So upon further research, I also found that indeed, actually offers 2 different web. It offers a get job and a job search API and through reading through these and testing them out.

311
00:41:29.909 --> 00:41:38.159
I found that I could really easily use the job search API in conjunction with the scraping application to get some more structured data than I would.

312
00:41:38.159 --> 00:41:52.135
Otherwise be able to because of the variability and the job descriptions, and then I could use the get job API to return. The filter drops back to the user for viewing. There was also some variability on the application links.

313
00:41:52.164 --> 00:41:56.485
So, my idea was to follow the links to the actual company site.

314
00:41:57.324 --> 00:42:09.414
To perhaps scrape some information from them about the specific requisition ID, or really any other additional information that may be on that site. But not all of the company's site links, take you to the listing.

315
00:42:09.894 --> 00:42:12.925
Some sites will take you to a job posting on the company's site.

316
00:42:13.199 --> 00:42:24.539
But others will take you to an online application page through a site, like work day and then others don't even have a company of link at all and they just take you to a quick apply capability or indeed. Itself.

317
00:42:24.539 --> 00:42:32.880
So, the indeed API would that was going to be the most important for my application was the job search.

318
00:42:33.385 --> 00:42:47.065
This would essentially provide the functionality of performing the search itself and provide me with the list of jobs that I would then be interested in scraping. So there were a few main highlights of this API that made it very useful in developing the desired functionality.

319
00:42:48.025 --> 00:42:50.695
1st, you're able to specify the output data format.

320
00:42:50.940 --> 00:43:02.909
So, for me, this was really important because I needed the Jason data in order to make it efficient enough to collect the scraped information and then parse it for the processing back to the user.

321
00:43:02.909 --> 00:43:10.469
Also, so through many of the functionalities, the offers that are well connected to Java script.

322
00:43:10.469 --> 00:43:19.650
Such as the ability to pass a call back function directly to the API, it became clear that I probably should have done this in Java script. But.

323
00:43:19.650 --> 00:43:23.250
I wasn't I wasn't about to deal with that so.

324
00:43:23.250 --> 00:43:30.030
The API also allows any of the parameters that are passed in for the search parameters.

325
00:43:30.030 --> 00:43:35.849
Um, available in the as well as a start index parameter.

326
00:43:35.849 --> 00:43:41.429
And this start index became hugely important to my application.

327
00:43:41.429 --> 00:43:51.900
Because it allowed me to just update an existing search and I did give me the ability to have persistence between searches. So, using this index.

328
00:43:51.900 --> 00:43:57.300
I only had to track the ending ID of the last return job for any previous search.

329
00:43:57.300 --> 00:44:06.179
And then I could pass that ID back to future searches to give the functionality of just updating the scraping as opposed to having to.

330
00:44:06.179 --> 00:44:09.929
Search through all of the jobs and scrape through all of them again.

331
00:44:09.929 --> 00:44:17.909
This not only makes the future collections much quicker, but it also only returns that newly posted jobs to the users.

332
00:44:17.909 --> 00:44:24.239
So, with all of that information, I was able to come up with the final design and then develop my application.

333
00:44:24.239 --> 00:44:31.409
So, my final design would focus on taking input from the user in the form of personal technical skills.

334
00:44:31.409 --> 00:44:46.380
Skills they wanted to work in and skills that they wanted to avoid as well as any of the other parameters such as years of experience, location and salary. These input parameters would then be used to formulate a search that would be passed through to the job search.

335
00:44:46.380 --> 00:44:54.840
Using the return data specifically the link of the job posting, I would then kick off multiple scraping jobs.

336
00:44:54.840 --> 00:45:03.030
Scraped valuable information from all of the search results as well as following any links that were contained in the descriptions or on that page.

337
00:45:03.030 --> 00:45:12.750
And in order to escape from all of them as well, with these return sets of data, I would then apply user preferences of applied skills and skills that they wanted to avoid.

338
00:45:13.135 --> 00:45:21.594
In order to filter out some of the jobs, and then return the remaining ones to the user and then call the get job. If the user selects to view the specific position.

339
00:45:22.284 --> 00:45:32.635
I then store the ending job ID index and a temporary file, which is searched for when a new searches kicked off. And if the new search matches the parameters of the previous 1.

340
00:45:32.940 --> 00:45:38.159
Only, the new jobs are returned, so, as to avoid processing the same positions time and time again.

341
00:45:38.159 --> 00:45:51.030
I'd love to show you guys as an example of this in action, but my code is currently a mass. I kind of messed it up earlier today, as I was working through some bugs with the scrappy return data.

342
00:45:51.030 --> 00:46:02.155
Um, the problem is that I'm getting a lot of garbage data and relevant data that I really don't want, and I'm working to clean it up so that I only get the information I want and don't have to. It doesn't get messy.

343
00:46:02.905 --> 00:46:14.514
So, please just see the flow chart to the right for visualization of how it should work. I didn't actually get blocked from any of the websites as a result of running the code, which was a little surprising to me.

344
00:46:14.820 --> 00:46:22.469
Although I did get curious at 1 point and intentionally got myself blocked from indeed for creeping, which was interesting to see.

345
00:46:22.469 --> 00:46:27.539
I think that the reason that I'm not getting blocked is because.

346
00:46:27.539 --> 00:46:37.019
I'm making a lot of API calls um, and that's the that's most of my interaction with the actual indeed dot com website.

347
00:46:37.019 --> 00:46:46.170
And then there are no, uh, company site specifically that I'm doing it enough work on are visiting enough times that it raises any flags.

348
00:46:46.170 --> 00:46:50.550
So that's pretty much the overview. Or are there any questions I can answer?

349
00:46:50.550 --> 00:46:59.250
Yeah, thank you. So this would work for other stuff besides jobs, I guess like apartments or something maybe.

350
00:46:59.250 --> 00:47:05.219
Yeah, without a doubt, it could definitely be easily applied to any sort of, uh, um.

351
00:47:05.219 --> 00:47:12.900
You know, posting site like that, you would just have to change up some of the scraping stuff again to match the.

352
00:47:12.900 --> 00:47:22.949
To match the new style of website and new outline of the website and then see if they have any, for instance, I think that Zillow has an API that you can use in conjunction with everything.

353
00:47:22.949 --> 00:47:30.300
Cool. Okay great. No other questions. Let us see, Dan.

354
00:47:30.300 --> 00:47:34.530
Would you like to tell us what you're up? Do? Hello?

355
00:47:34.530 --> 00:47:43.500
I don't know if you can hear me can hear you. All right awesome. Okay. Screen share.

356
00:47:45.000 --> 00:47:51.090
I can see you.

357
00:47:51.090 --> 00:47:54.690
So.

358
00:47:54.690 --> 00:48:02.699
I did my present, or I do my project in conjunction with my graphics.

359
00:48:02.699 --> 00:48:06.780
Project, so I chose to do it on.

360
00:48:07.284 --> 00:48:15.505
Trying to accurately render Opal so the stone, if you're all familiar with it, so kind of also related to material science in a way.

361
00:48:16.255 --> 00:48:22.675
So, taking what we learned about Ray tracing, and then applying it to or applying parallel computing.

362
00:48:23.039 --> 00:48:26.429
To hopefully get a nice speed up.

363
00:48:26.429 --> 00:48:31.349
So kind of the implementation for retracing. So.

364
00:48:31.349 --> 00:48:42.025
I used our kind of a backbone event. We used open GL, and I put in C. plus plus. So we kind of had a backbone program that I started off with.

365
00:48:42.025 --> 00:48:47.905
So, I had to look into a lot of the ways, the drawing loops and all that were set up to.

366
00:48:48.570 --> 00:48:56.699
Uh, actually do the speed up, but the general rate tracing algorithms. So there's a animation loop that.

367
00:48:56.699 --> 00:49:07.199
Originally went 100 pixels at a time, and it did a refinement. So it would divide up the size of the pixels. Originally the size. The pixels were not actually.

368
00:49:07.199 --> 00:49:13.260
Screen pixel sizes, but larger so I have changed that so that it would actually just be.

369
00:49:13.260 --> 00:49:23.489
Pixel on the screen, so it will call that animation loop and using the camera projection matrix.

370
00:49:23.489 --> 00:49:32.789
Shoot array from the camera coordinates through that respective pixel on the screen and recursively.

371
00:49:34.465 --> 00:49:48.835
Trace the pass of the Ray as it bounces through the scene. So, depending on the material properties that it collides with, here's a nice kind of representation. So this is a, you can think of it as a glass cube. So all the.

372
00:49:49.170 --> 00:49:52.530
The gray rays, the initial.

373
00:49:52.530 --> 00:50:01.559
Camera Ray, and here's a rotated view from the back where their faces open. So, as it intersects, you have the normal.

374
00:50:01.559 --> 00:50:10.079
Being this short red factor, blue rays are all refracted raised. So bending through the class.

375
00:50:10.079 --> 00:50:23.760
All red rays are reflected rays, so the inside will also have a reflected portion of the rate. And the Max bounce limit here is 5. so, after 5 intersections with the boundary of the cube, you'll see that.

376
00:50:23.760 --> 00:50:30.090
Just pass it straight through the wall and then green rays are to test if the.

377
00:50:30.090 --> 00:50:37.500
Position is in shadow, so innately this program is very dependent on recursion, which.

378
00:50:37.500 --> 00:50:42.599
It's of course, not easily translatable to a graphics card. Even though some people.

379
00:50:42.599 --> 00:50:48.090
Nvidia have implemented it on graphics cards, which I find amazing.

380
00:50:48.090 --> 00:50:53.250
So, I decided to do this on the CPU. So.

381
00:50:53.250 --> 00:51:02.130
I split up the trying loop to trace a whole line so it scans an entire line and once the line is done.

382
00:51:02.130 --> 00:51:12.659
Then it has dropped the screen kind of a big problem with also paralyzing tracing algorithm. Like this is that most.

383
00:51:12.659 --> 00:51:17.219
Basically, all the time, you cannot.

384
00:51:17.219 --> 00:51:28.230
Right. To the screen buffer in parallel it has to be a sequential Operation just the way pixels work. So kind of a way that this was.

385
00:51:28.230 --> 00:51:32.369
Oh, did it stop sharing? That's weird.

386
00:51:35.010 --> 00:51:38.369
I don't know if it's still sharing now.

387
00:51:38.369 --> 00:51:43.289
Sorry, that might have been something that may I do.

388
00:51:43.289 --> 00:51:48.570
Oh, is that all good visible again? Okay. Okay cool.

389
00:51:48.570 --> 00:52:00.659
Yeah, so drawing to the screen cannot really be done in parallel. It's always a sequential operation. So kind of to get around this.

390
00:52:00.659 --> 00:52:06.389
The tracing gets that pixel color, it pushes it back to a vector.

391
00:52:06.389 --> 00:52:10.110
So that's the main thing we have to be careful of. So, having a.

392
00:52:10.110 --> 00:52:13.619
The main thing was just adding a critical I don't.

393
00:52:13.619 --> 00:52:17.460
Think you can do pushbacks anatomically.

394
00:52:17.460 --> 00:52:23.670
I believe from what I've research, so kind of inefficient, but I'll get to that later.

395
00:52:23.670 --> 00:52:27.030
So that was the main thing for that. So, as a.

396
00:52:27.030 --> 00:52:30.630
As the vector is filled with the entire line.

397
00:52:30.630 --> 00:52:37.980
Then pixels are then taken from that and written to the screens functionally. So the group, the brunt of the.

398
00:52:37.980 --> 00:52:42.539
Computation are done and pushed it to the vector and then the copy is just a.

399
00:52:42.539 --> 00:52:55.014
Quick operation, so the 2nd part was a thing called photon mapping, which is kind of obscure so retracing itself.

400
00:52:55.135 --> 00:53:00.144
It's kind of weird how it's done from the camera, but that's done to not waste time.

401
00:53:00.659 --> 00:53:07.440
Tracing rays from the light source. A lot of them won't actually go into the camera so.

402
00:53:07.675 --> 00:53:17.125
From a lot of research, we realize that you can do it the other way. So tracing from the camera but of course, the results aren't as accurate as they could be.

403
00:53:18.715 --> 00:53:26.215
1 thing that both of these methods don't really account for is global elimination. So, like, when light.

404
00:53:26.519 --> 00:53:29.550
Comes out of, or goes through.

405
00:53:29.550 --> 00:53:35.369
Like, a glass sphere, or a magnifying glass, you'll get a kind of.

406
00:53:35.369 --> 00:53:40.769
Condensed point of bright light, so if you could imagine, that's.

407
00:53:40.769 --> 00:53:51.539
If you're reading from a texture that there's no reason that a spot should be that bright. If you're just looking at the texture color. So, at this method aims to do is.

408
00:53:51.539 --> 00:53:59.010
Kind of imitate those bright spots that you'd get from, like, bending and.

409
00:54:00.119 --> 00:54:08.940
Accumulating in certain points on your textures so it's done in the other direction. So from your light source, a bunch of photons.

410
00:54:08.940 --> 00:54:18.210
Defined by the user are shot across the screen and depending on the material properties, usually diffuse materials will.

411
00:54:18.210 --> 00:54:26.730
Accumulate these photons and then during the Ray tracing every rate intersection with a surface.

412
00:54:26.730 --> 00:54:41.429
Will collect a certain number of photons, which is a limits also by the user. So an area we'll expand to accumulate, let's say 50 photons and then, depending on the size of the area.

413
00:54:41.429 --> 00:54:46.860
And how many photons were collected you can infer a brightness. So.

414
00:54:46.860 --> 00:54:57.360
To make this more efficient Katie tree was used so spatial data structure, which you can. This is the 3 dimensional version, and the 2 dimensional version where.

415
00:54:57.360 --> 00:55:00.869
Uh, you're splitting your space in half.

416
00:55:00.869 --> 00:55:04.139
Not necessarily in for half of the.

417
00:55:04.139 --> 00:55:09.599
Access, but, uh, half of the dimension of where.

418
00:55:09.599 --> 00:55:17.219
Points are even on either side. So here we see there's 7 on the left of this golden line and 7 on them. Right? And any.

419
00:55:17.219 --> 00:55:22.500
Level down, you can see the same thing. So, 3 above the screen 3 below.

420
00:55:22.500 --> 00:55:31.110
And so forth, so, in this example, a lot, mainly all of the photons are on this point. So paralyzing this.

421
00:55:31.110 --> 00:55:35.190
It's also a recursive thing, because those photons, of course, have to bounce and.

422
00:55:35.190 --> 00:55:39.780
Do whatever interacting with the materials that I need to.

423
00:55:40.045 --> 00:55:52.554
But it's a lot simpler since you don't have to collect any of the photons with the race. They just stick and are placed into the KD tree. So, again, the Katy tree is built with factors.

424
00:55:52.554 --> 00:55:56.664
So each leaf has a vector associated with it. So being careful to.

425
00:55:58.320 --> 00:56:01.469
But those pushbacks in a critical.

426
00:56:01.469 --> 00:56:06.329
Here's some of the results I had so that would be nice to show.

427
00:56:06.329 --> 00:56:12.719
There was a paper from 2003 that I was basing this algorithm off of which.

428
00:56:12.719 --> 00:56:16.800
You know, 17 years ago, the results weren't too good, but I think.

429
00:56:16.800 --> 00:56:21.510
These definitely built on their methods.

430
00:56:21.510 --> 00:56:29.550
So some of the timing results, so I have a 3.6 gigahertz 16 threads processor.

431
00:56:29.550 --> 00:56:32.789
Um, all of these were done with my processor.

432
00:56:32.789 --> 00:56:37.320
I also don't have the video and I was using make.

433
00:56:37.320 --> 00:56:40.469
On Windows, so that really.

434
00:56:40.469 --> 00:56:48.150
Our deal and get open and Peter work. So I didn't want to bother with doing PGI doing with the graphic card but.

435
00:56:48.150 --> 00:57:00.989
It worked decently, so for the Ray tracing, it was this image, the scene. So it's a plane with a texture applied and a glass ball with like a red apply.

436
00:57:00.989 --> 00:57:05.909
And I slowly increase the screen size or the.

437
00:57:05.909 --> 00:57:18.300
Friend or scene size sequentially we see both linear increase relative to the screen size. So the area's increasing by 4 times and we see that kind of 4 times.

438
00:57:18.300 --> 00:57:27.480
Increase with the render time and in parallel same thing, it's only roughly about a 40% speed up with the rate tracing.

439
00:57:27.480 --> 00:57:31.619
And I'll get into the details of kind of why.

440
00:57:31.619 --> 00:57:43.494
It's not back also, so photon mapping was way better. So simpler operations and increasing the number of photons shot into the scene.

441
00:57:43.525 --> 00:57:46.675
So the, the scene for this 1 was the same as this.

442
00:57:48.719 --> 00:57:54.150
So, we go from half a 1Million increasing to 10Million also.

443
00:57:54.150 --> 00:58:00.570
Linear time increase, but this time we get a nice 6 time speed up. So.

444
00:58:00.570 --> 00:58:06.750
You know, pressing, he was the bind waiting quite a long time.

445
00:58:06.750 --> 00:58:11.820
If it's done sequentially, so.

446
00:58:11.820 --> 00:58:16.050
The main problem I saw was that since I was using this kind of.

447
00:58:16.050 --> 00:58:20.670
Pre written drawing loop I didn't have much.

448
00:58:20.670 --> 00:58:23.880
I felt like you could have been done way more efficiently.

449
00:58:23.880 --> 00:58:33.000
Rather than writing to effector, and then copying over, and actually, instead of drawing pixels, it was kind of done a different way where.

450
00:58:33.000 --> 00:58:46.380
Flat were drawn to the screen is geometry so it wasn't directly to the screen. I feel like that was taking away from a lot. Also the complexity of the drawing.

451
00:58:46.380 --> 00:58:57.960
Or the trace function was not to optimize for parallelization. So, you know, increasing that size of the circle, I feel like that could have also been paralyzed.

452
00:58:57.960 --> 00:59:03.179
Maybe even on the GPU since that's not the recursive at all.

453
00:59:03.179 --> 00:59:07.800
And then just the logic in general for drawing was a.

454
00:59:09.059 --> 00:59:13.230
Not optimal copying over data a bunch of times and.

455
00:59:13.230 --> 00:59:21.630
Maybe even doing it and I, I feel like since 1 scan line was only 200 pixels, 16 cores.

456
00:59:21.630 --> 00:59:25.800
Uh, you're not really optimizing the time use of those course.

457
00:59:25.800 --> 00:59:33.360
So, it may be doing it as the entire screen, but you don't get a progressive rendering of the screen then you might.

458
00:59:33.360 --> 00:59:43.019
Get some time benefits from not having to restart those threats over and over and this kind of proof of, uh.

459
00:59:43.019 --> 00:59:56.130
Where I was doing this, so here's the animation where the original 1 that was doing 100 pixels at a time here. We have our pregnant 4, and it calls the re, trace function.

460
00:59:56.130 --> 01:00:04.619
So, here we draw pixel and those will coordinate and gets our colors putting the push back in a critical to ensure.

461
01:00:04.619 --> 01:00:07.619
You can't write.

462
01:00:07.619 --> 01:00:20.670
Effectors aren't thread safe, sadly, so doing that and then this really big trace Ray function with all these recursive. Sorry? For the comment code. Yeah. So all this recursion.

463
01:00:20.670 --> 01:00:30.000
And similarly, with photon mapping, we have our recursive photon thing. Each of the pushbacks since the KD tree.

464
01:00:30.000 --> 01:00:44.639
Are in their own critical block, not starting a critical of course, would be way faster if it could be done atomic instead it doesn't have to set up the whole environment for the and the loop for shooting all the photons done in parallel for.

465
01:00:44.639 --> 01:00:49.380
I think that's it. There's any questions.

466
01:00:49.380 --> 01:00:59.010
Thank you so your promo speed up it's about equal to the number of core is actually not the number of hyper threads with the number. Of course. Yeah.

467
01:00:59.010 --> 01:01:07.800
Which is good. I was I was pleasantly surprised that it was of almost exact thing. I was linear. I'm sure if I did, like.

468
01:01:07.800 --> 01:01:15.090
With the photon mapping would be anything more obvious, but if I did like a drastic we large number, like, 100.

469
01:01:15.090 --> 01:01:21.179
You know, a 1Billion maybe photons. Maybe I'd start to see kind of a drop off, but.

470
01:01:21.179 --> 01:01:25.769
I was even surprised that even with that amount of recursion, I was still doing.

471
01:01:25.769 --> 01:01:29.789
Relatively well.

472
01:01:29.789 --> 01:01:34.079
What is Cole, I guess embarrassing me parallel or something? Yes.

473
01:01:34.079 --> 01:01:39.510
I'm still I'm, I'm curious how even with the recursion.

474
01:01:39.510 --> 01:01:44.340
They made it so fast with the I mean, now the, all the.

475
01:01:44.340 --> 01:01:47.940
Getting better at doing all this stuff, but yeah.

476
01:01:47.940 --> 01:01:55.289
It's, it's really cool to me. Yeah. I don't know how the invidia rate racers engines are implemented, but.

477
01:01:55.289 --> 01:01:58.679
Yeah, somehow they it, I guess yeah.

478
01:01:58.679 --> 01:02:03.269
Okay, anything else.

479
01:02:03.269 --> 01:02:07.199
Finally, Randy, you're on the stage now.

480
01:02:08.670 --> 01:02:19.019
Yeah, so, um, let me share this screen. Okay you screen.

481
01:02:19.019 --> 01:02:27.744
In here you can see you. Yeah, sure. So, um, yeah, so my project is real time processing was could an open CV.

482
01:02:28.195 --> 01:02:35.094
Um, so the 1st data image was the algorithm I choose is a gas in blurry so it's a mechanism delta's algorithm.

483
01:02:35.094 --> 01:02:44.965
Is yourself a 55 filter, uh, which follows a gassing distribution uh, and you use that filter to slice through image a pixel that picks up.

484
01:02:45.539 --> 01:02:52.650
And each pixel's value is up at based on the filter value in the position, and for us round and pixels.

485
01:02:53.784 --> 01:03:06.445
And so, the tourism platform, it shows is a CUDA and opens in Python. So imagine are in a proposal I was going to use PICA, that, which is a direct could for Python.

486
01:03:06.684 --> 01:03:17.184
So, in Pi, could I let you write a, could essay function as a string and, um, we'll throw it into a cecum pillar and a Python model.

487
01:03:17.730 --> 01:03:23.849
Hence, the program will automatically cause a module, a, as a imported function into a Python program.

488
01:03:24.144 --> 01:03:24.565
However,

489
01:03:24.594 --> 01:03:26.574
I set up to work,

490
01:03:26.605 --> 01:03:26.965
uh,

491
01:03:27.175 --> 01:03:27.355
to,

492
01:03:27.414 --> 01:03:27.505
uh,

493
01:03:27.534 --> 01:03:28.974
Petco to work with some,

494
01:03:28.974 --> 01:03:29.215
um,

495
01:03:29.244 --> 01:03:29.755
simple,

496
01:03:30.025 --> 01:03:30.445
um,

497
01:03:30.474 --> 01:03:32.905
Matrix calculation and I thought it's working and,

498
01:03:33.175 --> 01:03:33.355
uh,

499
01:03:33.355 --> 01:03:34.195
when I moved to the image,

500
01:03:34.195 --> 01:03:37.255
versus I think I stopped working because awesome competitor issue with,

501
01:03:37.525 --> 01:03:37.945
um,

502
01:03:37.974 --> 01:03:40.554
Microsoft studio and Windows.

503
01:03:41.010 --> 01:03:55.170
So, I have to go to another tool, which is a library card number. So it's a Python package for hyper computing and it has a lot of beauty functions for, and also could have function.

504
01:03:55.644 --> 01:04:06.534
And the coding style for number is exactly the same as we're learning lectures of the. But the only difference is is, um, syntax Python.

505
01:04:06.715 --> 01:04:17.155
So also let you, um, to allocate memory and, uh, assign how many, like like the size of each, each block and upgrade manually in, in number.

506
01:04:17.610 --> 01:04:31.380
So, the 1st thing I do is a offline image processing. So there's a demo of my board programs on the left is our new image on down to a left image and on a right it's a version.

507
01:04:31.380 --> 01:04:44.130
And so our insights was a 2000 by 1006 are tried to run on CPU for lights take me about, like, 20 minutes to process. So I have to.

508
01:04:44.130 --> 01:04:55.829
Skill by 20%, so which means adding more image and see 0 time for processing says a smart image is about how seconds.

509
01:04:56.244 --> 01:05:09.534
And when I moved to, um, couldn't respond in time, is about, like, 300 minutes. 2nd, and this is including, uh, copying data choosy and coming back to the, uh, host.

510
01:05:11.335 --> 01:05:11.664
And,

511
01:05:11.695 --> 01:05:11.994
uh,

512
01:05:12.264 --> 01:05:14.664
the device time I measure is 260 may 2nd,

513
01:05:14.664 --> 01:05:15.144
but,

514
01:05:15.445 --> 01:05:15.625
uh,

515
01:05:15.625 --> 01:05:16.195
in the,

516
01:05:16.224 --> 01:05:16.434
uh,

517
01:05:16.434 --> 01:05:18.505
in the video processing example,

518
01:05:18.505 --> 01:05:23.844
I found his process might also include some utilization time of the,

519
01:05:23.844 --> 01:05:24.264
and actually,

520
01:05:24.264 --> 01:05:27.954
as actual accommodation time could be much more shorter.

521
01:05:30.389 --> 01:05:43.019
Yeah, so for the data processing, I couldn't have some issue as far as that. So, uh, firstly, uh, they tend to not allow me to use a lot of grades for a calculation where I assigned, um, more and, um.

522
01:05:43.465 --> 01:05:56.454
64 creates a progress start to crash and so I have to scale the video by also about 20%. So if I said video too large, it's a video only runs for a few frames before a crash.

523
01:05:56.485 --> 01:06:02.454
So, it's really where to account a run for, like, 10 frames and parts and arrow for.

524
01:06:04.795 --> 01:06:19.135
And so I said, for this reason, um, so some, some on the Internet proposed, that might be because I'm using the same GPU for both display and the calculations. So it has some limitation on display timeouts a limit.

525
01:06:19.590 --> 01:06:31.590
So, that might be a prompt. So, uh, 1 solution I'm trying to, uh, to solve this problem is to the IRA into a, a 1 Sierra because.

526
01:06:31.590 --> 01:06:44.159
I think most of the example of doing major calculation is by flirting the matrix into a wanting area and pass it to kudos. So, I think I'm trying to use it's a message to solve that issue.

527
01:06:44.159 --> 01:06:48.329
So, now, I think I can show a demo off my.

528
01:06:48.329 --> 01:06:52.920
Video processing, let me.

529
01:06:53.940 --> 01:07:06.144
So, we know is a renasant because I had to scale it by only 20%. So, on the last 1 is the original, uh, video output from a webcam and on the right is the product version.

530
01:07:06.144 --> 01:07:09.744
So it's working but it's just have to be a really small.

531
01:07:12.960 --> 01:07:27.599
Yes, I think that's it. So here, uh, also, uh, um, processing time. So I can see each frame is processing a time is only about 1 to 2 minutes seconds. So it's a pretty fast.

532
01:07:27.599 --> 01:07:31.679
Uh, West yeah, yeah, I think that's it.

533
01:07:31.679 --> 01:07:34.769
Any question well, thank you.

534
01:07:34.769 --> 01:07:37.829
So, what type of were you using.

535
01:07:37.829 --> 01:07:42.449
I have a G force 30. okay.

536
01:07:43.469 --> 01:07:50.400
So, how would you extend this into the future? Maybe.

537
01:07:51.780 --> 01:07:56.489
Um, so I think on.

538
01:07:56.489 --> 01:08:11.369
So, limited program is limited to like, any processing but, uh, the 1st job I need to do is to fix the limitation of see a great I can use all my. So.

539
01:08:13.769 --> 01:08:16.859
Okay, beautiful. Yeah.

540
01:08:17.909 --> 01:08:22.020
So, it's a faster could do a larger image than.

541
01:08:22.020 --> 01:08:30.539
Maybe, yeah, maybe plug 2 onto your computer 1 for process 1 for display or something. Yeah. Could be.

542
01:08:30.539 --> 01:08:34.920
Okay.

543
01:08:34.920 --> 01:08:42.060
So this is basically the end of the course if anyone has any questions.

544
01:08:42.060 --> 01:08:51.869
Any issues you'd like to raise if you want to take a break while you're finishing your projects, you can read about the underhanded C code contest. Perhaps.

545
01:08:51.869 --> 01:09:00.359
That I've got a link to the international called contest see all the ways that you should not.

546
01:09:00.359 --> 01:09:05.189
Be programming your term projects, but you can have some fun there.

547
01:09:07.409 --> 01:09:17.850
Other than that, if there's no questions again, I'm available for questions and any legal ethical topic, even after you leave RPI.

548
01:09:17.850 --> 01:09:21.090
And I'll look forward to seeing your.

549
01:09:21.090 --> 01:09:32.760
Projects I would like to get them on the due date because the registrar wants grades from me something like 36 hours after the post due date for the projects or something because.

550
01:09:32.760 --> 01:09:44.189
Some people think they're graduating and to the senior course. So we're just starting the messy special cases taken after human being to process. And that takes time.

551
01:09:46.050 --> 01:09:50.520
Other than that, I have a good summer.

552
01:09:50.520 --> 01:09:53.729
So, quick question, professor.

553
01:09:53.729 --> 01:10:01.949
Is there anything we need to submit for the presentation? There's I see the.

554
01:10:01.949 --> 01:10:07.229
The submission slot on grade slip, but it does that allow us to attach anything.

555
01:10:07.229 --> 01:10:10.979
No, I was just going to give you points for the presentation on that.

556
01:10:10.979 --> 01:10:14.220
Okay, okay, that's what I figured but thank you.

557
01:10:14.220 --> 01:10:23.789
Oh, 1 thing I thought if anyone would like to have a shout out on my course website I'd be glad to do that.

558
01:10:23.789 --> 01:10:27.029
You know, this is optional, but if you'd like to.

559
01:10:27.029 --> 01:10:36.779
Need to post a link to your resume or something, or a link to your project, or whatever. It's totally voluntary on your part, but I'll be glad to do it. So.

560
01:10:37.920 --> 01:10:41.489
You can email me at any time and I'll link stuff in.

561
01:10:43.619 --> 01:10:49.859
Other than that, I actually also a question so.

562
01:10:49.859 --> 01:10:54.060
For since I did a conjunction project, I had to write a.

563
01:10:54.060 --> 01:11:02.039
Report for that. Is it? All right? I included a pretty detailed section on the parallel implement. That's good enough. Right? Is that.

564
01:11:02.039 --> 01:11:11.369
Okay, I'll just submit that. Yes cool. Thank you. No, we just put that paragraph and because sometimes, okay somebody submitted an identical project to do courses.

565
01:11:11.369 --> 01:11:14.880
That's fair. I got it. Cool.

566
01:11:14.880 --> 01:11:26.520
No, there are any requirements and expectations you have for this report. I wasn't able to find any specific specs on the website.

567
01:11:26.520 --> 01:11:33.869
Well, 1st, physical format. I've got something I put on today's blog that I copied out of the syllabus.

568
01:11:33.869 --> 01:11:41.939
Um, it'd be nice to make it look like a professional conference on paper actually. So I suggest the I triple E conference format.

569
01:11:41.939 --> 01:11:53.460
Would be 1, if you prefer ACM or something else that's okay. They I believe, for example, they've got outlines for both tech and Microsoft word.

570
01:11:53.460 --> 01:11:57.119
And, you know, for what the format should be.

571
01:11:58.560 --> 01:12:01.590
Something 4 to 8 pages and.

572
01:12:01.590 --> 01:12:05.399
And also upload.

573
01:12:05.399 --> 01:12:10.020
Apart from the paper, you want to upload something.

574
01:12:10.020 --> 01:12:16.170
More details of what how your program works and maybe even the source code sample images, stuff like that.

575
01:12:16.170 --> 01:12:22.140
If it will fit on degrade scope fine if it's more convenient, put it somewhere else and send me a.

576
01:12:24.090 --> 01:12:30.449
Awesome Thank you. Make it something. It'd be proud to show to a potential employer perhaps.

577
01:12:31.859 --> 01:12:37.739
And also from this course, you've got lots of key words to put on your resumes.

578
01:12:37.739 --> 01:12:43.319
You know, so when corner scrapes here or something well, he'll get.

579
01:12:43.319 --> 01:12:47.520
Yeah, scraping for jobs, not resumes they get selected, so.

580
01:12:49.170 --> 01:12:56.430
Other than that, I encourage you to browse around things like the Super computing conference website.

581
01:12:56.430 --> 01:13:00.600
And maybe Jack on camera, or brows look for his latest talk.

582
01:13:00.600 --> 01:13:05.340
He, he's a leader and Super computing so it's fun to look at.

583
01:13:05.340 --> 01:13:15.239
Is talks on, you know, what the state of the art it is, and what's happening and so on and Super computing again.

584
01:13:15.239 --> 01:13:18.810
Browsing around and video for the last NVIDIA.

585
01:13:18.810 --> 01:13:23.789
Conference is just last, like, a month or 2 ago and browse around.

586
01:13:23.789 --> 01:13:27.359
Some of the leading talks there.

587
01:13:27.359 --> 01:13:36.869
It can be fun to see what they think is important. And latest thing in the news in video, what they're trying to buy are my guess that hasn't been fully approved yet.

588
01:13:36.869 --> 01:13:43.020
They are trying to eat into intel's customer base because.

589
01:13:43.020 --> 01:13:48.479
They're trying to go for they're trying to see how they can make the GP use, do more general purpose computing.

590
01:13:48.479 --> 01:13:54.119
So, I guess they're figuring that the Intel architecture.

591
01:13:54.119 --> 01:13:59.010
Maybe, it's sort of gotten as complex as it can and.

592
01:13:59.010 --> 01:14:02.729
You know, the next generation into general computing might be.

593
01:14:02.729 --> 01:14:07.560
Or using and video sell, they're trying to put stuff in them to.

594
01:14:07.560 --> 01:14:14.789
Make them use, reduce that also the software latest additions to C. plus plus just the standards.

595
01:14:14.789 --> 01:14:28.560
To try and have you put a few directors in saying what can be parallelized and if your algorithm is parallelizable, that's always the hard part. So they're trying to make it easier on the programmer and they're trying to make the hardware more flexible. So.

596
01:14:28.560 --> 01:14:35.579
If I'm to see what happens, anyone have any opinions on that.

597
01:14:37.319 --> 01:14:44.399
I'd like to see the competition and in 10 years in video might not be around anymore. I've seen companies.

598
01:14:44.399 --> 01:14:48.510
Go up up up up up and then click down very fast faster. Then they went up.

599
01:14:48.510 --> 01:14:51.600
Hope it doesn't happen on video, but if it.

600
01:14:51.600 --> 01:14:57.989
Overconfident America and it can happen. So you have no Intel, but they're actually pretty good.

601
01:14:57.989 --> 01:15:10.050
Sitting around, so what can happen is invidia been there from the start and when he finally leaves and the question is, can they survive that?

602
01:15:10.050 --> 01:15:14.729
That's always a critical part because, you know, companies, what 25 years old give retain.

603
01:15:16.140 --> 01:15:19.170
What's your opinion on that? Do you think they'll survive if he does?

604
01:15:19.170 --> 01:15:23.039
It's hard to predict the future depends on who replaces him.

605
01:15:23.039 --> 01:15:26.939
And I've also seen companies.

606
01:15:26.939 --> 01:15:38.550
Where a grad student has come to the digital equipment Corp and they, every university computer science department, every leading 1 use deck computers, they're open architecture.

607
01:15:38.550 --> 01:15:50.399
Company was shepherded by a guy called Ken Olson, he built the company up at 1 point. It was number 2, the number 2 computer company in the United States, and they held there. They rented a notion line or the Q2 or something for their annual.

608
01:15:50.399 --> 01:15:57.930
Conference in Boston, and then also he after doing everything, right? For decades he did everything wrong all at once.

609
01:15:57.930 --> 01:16:02.189
And killed the company, it crashed early fast.

610
01:16:02.189 --> 01:16:09.569
So, that has happened when i1st is at our is a company called sun, micro computers. Um.

611
01:16:09.569 --> 01:16:13.109
Came out of the Stanford project open standards for workstations.

612
01:16:13.109 --> 01:16:22.170
And I bought a 50,000 dollar workstation in the eighty's that had like a 1 MIP synergy, only Micro.

613
01:16:22.170 --> 01:16:31.260
Motorola, black and white display. Like, I think it was a couple of megabytes of main memory company went up and up and up and then.

614
01:16:32.340 --> 01:16:40.470
Started screwing up, I guess, and then Oracle bought them. That's where Jabber came from. Of course. So companies.

615
01:16:40.470 --> 01:16:47.640
You know, it's a lesson there you got to keep working at it to be continue to be successful.

616
01:16:47.640 --> 01:16:53.579
Actually met, who was 1 of the CO, founders of Sun Microsystems.

617
01:16:53.579 --> 01:17:02.789
It's a really cool guy. What's he like you said, he's awesome. My uncle actually started a start up with his son and Valley.

618
01:17:02.789 --> 01:17:14.069
He's not, he's not what you typically expect from 1 of the VC people. He's not super ignorant. They're stuck up. Really?

619
01:17:14.069 --> 01:17:19.380
Down to earth guy, what's he doing? Now? He just I actually don't know.

620
01:17:19.380 --> 01:17:32.640
There's a whole parallel universe of people that's a very active, but they're not visible on the public Internet into venture capitalists and so on, for example, I've known some people as they go into that universe and they're.

621
01:17:32.640 --> 01:17:38.069
Working hard productive people, but the public doesn't see what they're doing. I guess.

622
01:17:38.069 --> 01:17:41.159
Yeah, I think my understanding at this point is he's just.

623
01:17:41.159 --> 01:17:44.279
Dumping money into tech startups.

624
01:17:44.279 --> 01:17:48.989
Okay, where things to do.

625
01:17:48.989 --> 01:17:52.739
Yeah, I can think of lots of them.

626
01:17:54.149 --> 01:18:08.159
I met Bill Joy once who did various things he did parts of Unix back before Linux and when he is at Berkeley, and then he was 1 of the people involved in sign at the start also.

627
01:18:08.159 --> 01:18:13.739
And then, I don't know what he ended up doing. You made some people just buy a yacht and retire.

628
01:18:15.805 --> 01:18:30.564
And there are some people do repeated things. There was a profit Stanford called Jim Clark that he was involved in pounding graphic stuff. Then he's involved in funding Netscape and funding some medical records company.

629
01:18:30.840 --> 01:18:38.670
And then, well, then he ended up getting divorced a couple of times and buying a yacht that was 1 of the larger yachts in the world. So.

630
01:18:38.670 --> 01:18:42.390
To patterns.

631
01:18:42.390 --> 01:18:45.960
Different 1st weights to retire. Yes.

632
01:18:45.960 --> 01:18:49.770
Well, you could work hard on your yacht maybe.

633
01:18:49.770 --> 01:18:53.039
Is, and you get a virus comes in, you can't take the God anywhere.

634
01:18:53.039 --> 01:19:00.659
So, yeah, that's a theme. Now, people you got a past Internet, you could work and unusual places. It could be very productive.

635
01:19:02.729 --> 01:19:07.590
I would hope to be out in San Diego for 2 months and.

636
01:19:07.590 --> 01:19:12.449
Starting in 3 weeks actually, so I'm going to be working Internet.

637
01:19:14.520 --> 01:19:18.149
At the moment, I'm 10 miles from our.

638
01:19:18.149 --> 01:19:24.270
What are you what are people's plans here all you all graduating now or going to grad school or getting jobs? What.

639
01:19:26.010 --> 01:19:32.189
I'm graduating okay. Getting a job. I'm working for a contractor in DC.

640
01:19:32.189 --> 01:19:37.979
Okay, most of graduating be working as a software engineer in Boston.

641
01:19:37.979 --> 01:19:42.930
Good I'll probably go for a.

642
01:19:42.930 --> 01:19:46.079
Okay, and what topic.

643
01:19:46.079 --> 01:19:49.170
Aerospace engineering. Okay.

644
01:19:49.170 --> 01:19:53.640
I need super computing there. Yep, definitely.

645
01:19:53.640 --> 01:19:57.329
Continuing my masters. Okay.

646
01:20:00.090 --> 01:20:05.819
Anything else that most crowd? Anything I'm thinking about taking up yesterday degree.

647
01:20:05.819 --> 01:20:11.609
I'm sorry, you're doing what, Randy I'm going to pursue a pH. D degree.

648
01:20:11.609 --> 01:20:17.069
Okay, what in in the action and machine learning. Okay.

649
01:20:18.869 --> 01:20:25.319
Anything else? No well, then good. He said goodbye. Good luck.

650
01:20:25.319 --> 01:20:33.449
Lots of success and and you get really rich then send some money back to our maybe, but we don't need another art center. So.

651
01:20:33.449 --> 01:20:38.489
So, signing up.

652
01:20:38.489 --> 01:20:41.670
Okay, great Thank you. Yeah.

653
01:20:41.670 --> 01:20:44.939
For surveys, thank you.

654
01:20:44.939 --> 01:20:46.500
Okay, bye.