[CS61C FA20] Lecture 27.4 - Caches IV: Actual CPUs — Transcript

Full transcript

1. 0:00and welcome back this is it we're at the
2. 0:03end this is the last lecture of the
3. 0:05caches series i love this
4. 0:08let's let's show some pictures of actual
5. 0:10cpus
6. 0:11let's see what it looks like so let's
7. 0:14take a look the early power pc
8. 0:17so this is a long time ago this is a
9. 0:19couple generations ago
10. 0:21we mentioned before that we have
11. 0:23separate
12. 0:24l1 caches for instructions of data
13. 0:27partly because they come from different
14. 0:29places of memory so why corrupt them so
15. 0:31how big are they 32 kbytes of
16. 0:35instructions is data okay and data for
17. 0:37each instructional data
18. 0:39i've got an external l2 cache interface
19. 0:42and the integrated controller and cache
20. 0:44tags so look at this
21. 0:46this l2 cache tags just the tags for l2
22. 0:51are here
23. 0:52okay here's my data cache this is my l1
24. 0:55data cache this is my l1 instruction
25. 0:58cache that's the big one
26. 1:00and what i've stored is i don't have l2
27. 1:02on chip but i have the
28. 1:04tags are on chip so i can do the tag
29. 1:06comparison really fast
30. 1:07but that's it by the way here are the
31. 1:10tags for data there's the tags for my l1
32. 1:12this is my tags for my l1 and these are
33. 1:14my tags
34. 1:15for my l1 instruction so each of these
35. 1:18related to that guy
36. 1:19this is the l2 cache tags um
37. 1:22i've got different units here so we can
38. 1:23talk about this i've got uh here's maybe
39. 1:26here's a memory management unit to kind
40. 1:27of coordinate all this we don't know
41. 1:28about the tlb for now
42. 1:30i've got a load store unit an energy
43. 1:32unit and a floating point unit
44. 1:33and that lets me in the best case we
45. 1:36talked about this a little bit of
46. 1:37pipelining
47. 1:38super scalar means i can run parallel
48. 1:40jobs so i can actually have
49. 1:41three i could have a load store
50. 1:44situation i can
51. 1:44integer i'm doing some aou kind of thing
52. 1:46i'm doing a floating point thing
53. 1:48all at the same time that was pretty
54. 1:50cool
55. 1:51so here is six execution units allow for
56. 1:54two integer
57. 1:56and double precision and one double
58. 1:58precision at exactly the same time in
59. 1:59terms of my execution units
60. 2:01that's it this is this is a single core
61. 2:02this wasn't a multi-core machine back in
62. 2:04the day
63. 2:06pentium m this is again early here we go
64. 2:09same kind of thing here's uh so a
65. 2:11completely different company that was
66. 2:13motorola this is intel
67. 2:14here's my here's my 32 kb of
68. 2:17instructional and data here
69. 2:19and anything else i want to say oh look
70. 2:21this is
71. 2:23here is my l2 cache
72. 2:26so not just well here's the tags for l2
73. 2:29i put
74. 2:29the ca i put this is on the chip l2 is
75. 2:32on the chip this is
76. 2:33l2 there's my l2 there okay here's my l1
77. 2:37and there's my l1 i love this so that's
78. 2:39pretty cool and now
79. 2:41watch what happens as i get to the next
80. 2:42level
81. 2:45intel core i7 it's hard to get more
82. 2:48recent than this
83. 2:49i've got six cores here's six cores
84. 2:52okay in that core
85. 2:56i'm not showing you what
86. 2:59the cache looks like but rarely
87. 3:05i'm going to show you something
88. 3:06interesting each core
89. 3:08has an l2 on it
90. 3:12and probably i actually don't know where
91. 3:15it is i think it's
92. 3:16usually what happens is you look for the
93. 3:17things that look all the same see how
94. 3:19this looks like it's all the same
95. 3:20housing
96. 3:21i don't have it in this document so i'm
97. 3:24not i'm just going to guess
98. 3:25that i'm looking at here as an area and
99. 3:28here there's a big area i'm not sure
100. 3:30this is we can take a look at this
101. 3:31but this has l1
102. 3:34and l2 per core we talked about we saw
103. 3:37that number from when i did under
104. 3:38the perimeter of that and here's what's
105. 3:40fun
106. 3:42i also have on chip this is on chip i
107. 3:44also have
108. 3:46a shared l3 across all cores
109. 3:50so all cores share this l3 which is
110. 3:53pretty neat
111. 3:55in conclusion boy bringing it all
112. 3:57together
113. 3:59we have talked about caching we spent
114. 4:01four lectures talking about caching
115. 4:03in detail i mean this is remarkable and
116. 4:05caching shows up
117. 4:06as i mentioned before in one of my first
118. 4:08lectures caching is a big deal
119. 4:09caching's a big deal in computer science
120. 4:11it's one of the big ideas that we've
121. 4:12contributed to the space
122. 4:13i don't know if caching it like has an
123. 4:16example
124. 4:17abstraction exists in engineering you
125. 4:19know there's abstraction you hide the
126. 4:20details of how you
127. 4:21you know have a shuttle flight or
128. 4:22something how you build a bridge
129. 4:24abstraction is there
130. 4:25i don't know if caching occurs in a
131. 4:28non-computer engineering computer
132. 4:30science context if there's an equivalent
133. 4:32thing
134. 4:32like that i certainly know that i have
135. 4:34you know
136. 4:35any kind of technology where i can make
137. 4:37a simple copy of something that's there
138. 4:38but if i don't make
139. 4:39if it's a physical element i don't know
140. 4:41if the cat if the idea of a cache
141. 4:43and a memory hierarchy exists outside of
142. 4:45a computer engineering computer science
143. 4:46kind of idea
144. 4:47but we certainly have software caches
145. 4:49and caches except everywhere
146. 4:50file system caches web page caches game
147. 4:53databases table bases
148. 4:55phone most recent calls all those things
149. 4:57are caches
150. 4:58software memorization software my
151. 5:00goodness software memorization is a big
152. 5:02idea is a cache in some sense
153. 5:04as i'm walking through fibonacci uh why
154. 5:07don't you just write it down as you're
155. 5:08seeing it
156. 5:09and so you're kind of caching it and now
157. 5:10if i have i seen it recently yes i have
158. 5:12i forget about the recently part but i
159. 5:13haven't seen it before yes well grab it
160. 5:14from my
161. 5:15cache and my memoization storage same
162. 5:18idea so that same idea is with caches
163. 5:21and again the big idea is if something's
164. 5:23expensive we want to do repeatedly do it
165. 5:24once and catch the result
166. 5:26so if i have this computation as being
167. 5:27really heavy i don't care what it is
168. 5:29do it once and remember what it is
169. 5:31that's what it is that's the kind of big
170. 5:32idea in kind of computing
171. 5:33but it's also don't go to sacramento
172. 5:35don't compute it and spend an hour
173. 5:36computing it if i'm going to ask for it
174. 5:38twice computer once remember it now i've
175. 5:39got it for the second time especially if
176. 5:41it's a function and only and the output
177. 5:44is only a function of its inputs if it's
178. 5:45somehow a function of all these other
179. 5:46state elements i can't cache it because
180. 5:47it might be different
181. 5:48but if it's a pure function so i
182. 5:50remember that for the next time um
183. 5:52similarly if i have to go to sacramento
184. 5:54why go to sacramento all the time
185. 5:55bring it to some local space don't have
186. 5:57to go to the book i mean imagine the
187. 5:59world without caches that
188. 6:01really like dumb and dumber level okay
189. 6:04i'm gonna write a paper
190. 6:05i go to the find it look at the thing i
191. 6:07find the book i get it out i read one
192. 6:09line of it okay great put it back
193. 6:10hashtag my paper oh what else did they
194. 6:12say i have to go get the book again are
195. 6:14you kidding me that's like going to
196. 6:15sacramento all the time
197. 6:16put it on your table that's the cash so
198. 6:18that library analogy introduced
199. 6:20you know first or second lecture same
200. 6:22idea here
201. 6:23go grab it and have it in a local space
202. 6:24so it's handy you can use it that's the
203. 6:26idea
204. 6:27a ton of knobs holy moly how many knobs
205. 6:31you have
206. 6:31how big is your cash overall what's your
207. 6:34technology behind that
208. 6:36what's your block size how wide is your
209. 6:38cash what's your right pilot what's your
210. 6:39aspect ratio
211. 6:40once you have your cash is it is it why
212. 6:42you know fat and thin
213. 6:44is it short and fat or one big block
214. 6:47size or is it really small block size
215. 6:49and now it's really tall and skinny
216. 6:51what's my right policy right through
217. 6:52versus right back we talked about that
218. 6:55what's my associativity what's my knob
219. 6:57for my
220. 6:58n my m is it one or is it n
221. 7:01what's my block replacement policy we
222. 7:03talked about a couple of policies
223. 7:05do i have a second level cash do i have
224. 7:07a third level oh my gosh i mean
225. 7:09but all that for a cash designer is just
226. 7:12more
227. 7:13more freedom more flex more pens of an
228. 7:16artist
229. 7:16to draw whatever i want to be able to
230. 7:18create what i want to optimize for
231. 7:19whatever usage case i have
232. 7:21and again there are software usage cases
233. 7:22there hardware uses caches all these
234. 7:24things are out there
235. 7:24for a cache designer which again is not
236. 7:26always a hardware designer could be a
237. 7:27software designer thinking about what
238. 7:28parameters for the cache that you want
239. 7:31and again what's your metric you're
240. 7:32trying to think about you have some
241. 7:34metric as you're having all these knobs
242. 7:35don't just
243. 7:36randomly okay i'll randomly set it i'm
244. 7:37good thanks so much where's the paycheck
245. 7:39no you
246. 7:39you go and you kind of try to run it on
247. 7:41traces you run it on practice things you
248. 7:42see
249. 7:43if it's performing and you adjust you
250. 7:44tweak maybe i'll change the block
251. 7:45replacement policy
252. 7:46maybe i'll make it more associative all
253. 7:48those things are parameters that you can
254. 7:49play all the things listed here
255. 7:50are things you can play with as you're
256. 7:52trying to adjust it for that ever work
257. 7:53whatever tracers you could be running
258. 7:54those are simulated traces may not be
259. 7:56perfect but it'll be a
260. 7:58show me the example kind of how i'm
261. 7:59going to be hitting this cache in the
262. 8:00future
263. 8:01thinking about that and use that
264. 8:02performance model to adjust between all
265. 8:04those choices
266. 8:05and also factor in your budget factor in
267. 8:07your technology your budget
268. 8:08what's there all of a sudden flash comes
269. 8:10around how am i that factor
270. 8:12that was really fun watching the
271. 8:13computer engineering world as flash
272. 8:15became really
273. 8:15interesting where that fit in by the way
274. 8:18your hard drive has a cache
275. 8:20go look at your hard drive specs there's
276. 8:21a cache on your hard drive
277. 8:23because going to disk spinning disk is
278. 8:25really slow so there'll be a cache on
279. 8:27your
280. 8:27hard drive caches are everywhere so
281. 8:30think about what technology you have the
282. 8:31space
283. 8:32based on and think of your cost for your
284. 8:34particular workload
285. 8:35that's it thank you so much for coming
286. 8:37to these cast lectures i really had a
287. 8:38good time teaching them to you
288. 8:39and we'll see at the next module take
289. 8:41care folks