[CS61C FA20] Lecture 24.2 - Caches I: Library Analogy — Transcript

Full transcript

1. 0:00and welcome back and our way to
2. 0:02understand what caches are sometimes it
3. 0:04helps to have a really good analogy
4. 0:06this library analogy is really very
5. 0:08strong so let's take a look at what that
6. 0:10means and understand how that connects
7. 0:11to what caches are
8. 0:13here's the big picture you've seen this
9. 0:15before this is the high level
10. 0:16abstraction layer of
11. 0:18why this is the new school machine
12. 0:20structures class not just the old school
13. 0:21machine structures class and today we're
14. 0:23looking at those two pieces
15. 0:24caches and main memory and how they
16. 0:26interact
17. 0:28here is the picture of the components of
18. 0:29the computer just to remind you on the
19. 0:31left is the processor
20. 0:32control and data path controls the brain
21. 0:34data path is the braun the wires
22. 0:36you have some interaction with memory
23. 0:37and memory has interaction with i
24. 0:39i and o but the idea is how would this
25. 0:42change
26. 0:42if i bring in what caches are so keep
27. 0:44this in mind as i start to teach you
28. 0:45what caches are
29. 0:46in the next slide or two so here's the
30. 0:50analogy with libraries and let's think
31. 0:51of make sure that we're familiar with
32. 0:53how this works
33. 0:54so you've got to find a book in a big
34. 0:55library you know you're a cal student
35. 0:57you're trying to find a book
36. 0:58in the stacks so what's your process
37. 1:01well you've got to go and go find the
38. 1:04catalog card card catalog and it's
39. 1:05really large and so you're
40. 1:07searching search and searching search
41. 1:08and searching and the larger the library
42. 1:09the slower that is because there's more
43. 1:10things to search through
44. 1:12there's also the round trip time to
45. 1:14actually walk to the stacks pick up the
46. 1:15book
47. 1:16and actually retrieve the desired book
48. 1:17so there's kind of two stages one is to
49. 1:19look at where it is
50. 1:20and then the second is to figure out
51. 1:22actually go get it and bring it back so
52. 1:24that's getting it bring it back and then
53. 1:25firing out where
54. 1:26it is the larger you make libraries the
55. 1:28worse
56. 1:29the worse that is imagine the card
57. 1:30catalog is you know floors of card
58. 1:32catalogs
59. 1:33imagine how painful that might be back
60. 1:34in the days before search by the way
61. 1:36before you know having computers and
62. 1:38libraries this is the model just so we
63. 1:40appreciate how painful it used to be
64. 1:42okay boomer yeah you got it all right
65. 1:44here we go
66. 1:45[Music]
67. 1:46so electronic uh memories
68. 1:49i e virtue you know virtual libraries in
69. 1:51that sense uh have the same issue
70. 1:53and plus the technology as you go to
71. 1:57different technologies to store the bits
72. 1:59as you go farther away from the cpu the
73. 2:02fastest one is the register it's
74. 2:03literally the fastest one of the
75. 2:04registers and you go farther away from
76. 2:06that they get larger and they get much
77. 2:07slower and so that the full technology
78. 2:09changes there and that's different than
79. 2:11libraries larger it's like
80. 2:12a longer distance another floor this is
81. 2:14like what if another floor required you
82. 2:16to walk through quicksand or something
83. 2:17that's it's a
84. 2:17so it's even worse when you get to um
85. 2:20computer memories in terms of
86. 2:21a large library well just have to walk
87. 2:22longer you know new york public library
88. 2:24maybe one of the largest ones in the
89. 2:25world whatever
90. 2:26it's just walking farther here it's
91. 2:27actually walking farther and the time is
92. 2:29slower per step
93. 2:30it's even worse all right what we want
94. 2:33and the ideal case by the way we want a
95. 2:35really large memory we want as large as
96. 2:37we can
97. 2:37but as fast as the smallest element
98. 2:40that's the perfect thing
99. 2:41boy i want it fastest registers but as
100. 2:42big as as as infinity
101. 2:45so that's really what we're looking for
102. 2:47this is more back setup again
103. 2:50back in the 80s um
104. 2:53these are two crafts of the graphs
105. 2:54between the speed of a processor
106. 2:56and the speed of memory uh how many
107. 2:58clock cycles
108. 3:00were to do to do something and so what
109. 3:02you're seeing is the cpu performance has
110. 3:04increased
111. 3:05incredibly um but in terms of the time
112. 3:09to get to memory memory is increased
113. 3:11much slower than that and that's
114. 3:12that increase of seven percent per year
115. 3:14versus fifty-five percent appear this
116. 3:15graph ends in 2010 but it and it
117. 3:17flattened out you know
118. 3:18cpu performance flattened out um after
119. 3:20about two thousand four thousand five
120. 3:22we call this the sea change we'll
121. 3:24actually talk about that in this class
122. 3:26um but still that gap was really big and
123. 3:28they so the gap is closing but it's not
124. 3:29closing very fast
125. 3:30so that's an issue so if you think about
126. 3:32in 1980 um
127. 3:34basically one instruction is the same as
128. 3:36dram access so if i happen to
129. 3:38need to go to memory that's not a big
130. 3:40deal but in 2020
131. 3:42to go to memory is a thousand
132. 3:44instructions so
133. 3:45the cpu if i do nothing else if i'm not
134. 3:47smart about how i use my time
135. 3:49the cpu is sitting idle for a thousand
136. 3:52cycles
137. 3:52a thousand full cycles which might not
138. 3:54seem a big deal to you because it's at a
139. 3:56gigahertz level so it's like what's the
140. 3:57big deal but in terms of
141. 3:59what if i do this a lot what if i'm
142. 4:00doing this a lot that ends up
143. 4:03being a bit have a big performance hit
144. 4:04on our software especially as i'm trying
145. 4:06to
146. 4:06compute on larger and larger quantities
147. 4:09of data
148. 4:09i can't afford to have that that hit
149. 4:11that memory hit to do that so we're
150. 4:13doing
151. 4:14all these whole set of lectures is to
152. 4:16deal with this slide here
153. 4:17how we try to deal with that disparity
154. 4:20between
155. 4:20cpu speed and dram speed that's it
156. 4:24so in the next couple of lectures that's
157. 4:25the last slide on this particular
158. 4:27issue we're going to see whether an idea
159. 4:29called caches can resolve
160. 4:31and solve this problem okay the idea is
161. 4:33in the big picture
162. 4:34can i be as fast as i can as fast as a
163. 4:37register but
164. 4:38at the speed sorry as fast registered
165. 4:40but at the size of
166. 4:42unimaginable size i imagine the size of
167. 4:44my hard drive even bigger than that
168. 4:46could i work with that okay we'll see
169. 4:48the next lecture