[CS61C FA20] Lecture 21.2 - Pipelining I: Processor Performance Iron Law — Transcript

Full transcript

1. 0:01[Music]
2. 0:09hello
3. 0:10welcome back to 61c module on
4. 0:13the measurement and improvement of
5. 0:15performance
6. 0:16through pipelining
7. 0:19so we get some understanding that there
8. 0:22are different views of performance
9. 0:25and one of the main ways how we measure
10. 0:27the performance is the time that it
11. 0:29takes to execute a program
12. 0:31but there are generally very many
13. 0:34parameters that affect that execution
14. 0:36time
15. 0:37they are very difficult to detangle from
16. 0:40each other
17. 0:42so we need to break them down down
18. 0:44somehow
19. 0:46there is something that we call the iron
20. 0:49law of processor performance
21. 0:51that tells us how long does it take to
22. 0:54execute a program
23. 0:56based on some of the basic parameters of
24. 0:59the machine
25. 1:00and the program that is being executed
26. 1:03so let's take a look at that in this
27. 1:06diagram we have time
28. 1:09that is the takes to execute the program
29. 1:12simply expanded or this fraction get
30. 1:16multiplied and divided by the number of
31. 1:18instructions
32. 1:19and the number of cycles
33. 1:24so time to execute the program is equal
34. 1:27to the number of instructions in the
35. 1:29program
36. 1:30the number of cycles per instruction
37. 1:34and time to execute one cycle
38. 1:38so this is now a lot easier to break
39. 1:42down and understand the fundamental
40. 1:44causes or fundamental parameters that
41. 1:47affect each of these
42. 1:50different programs may have
43. 1:53different number of instructions one of
44. 1:56the fundamental architectural
45. 1:58parameters is the number of cycles that
46. 2:01it takes to complete an instruction
47. 2:03and finally often
48. 2:07processor speed is associated with the
49. 2:09time to
50. 2:10to execute one cycle or the frequency at
51. 2:13which processor runs
52. 2:15but none of these parameters can be
53. 2:18analyzed alone they all have to be put
54. 2:22together
55. 2:22to understand that how long does it take
56. 2:25to perform a task
57. 2:27so let's get into that let's analyze
58. 2:30each one
59. 2:31of these three basic components number
60. 2:34of instructions per program
61. 2:36number of cycles per instruction in time
62. 2:38to complete a cycle
63. 2:40to see how they what affects them
64. 2:43and how do they how can they get
65. 2:47put together in the end
66. 2:51so first one how many does it
67. 2:54how many instructions does it take to
68. 2:56execute a program
69. 2:57well it really depends on what kind of a
70. 2:59task are we
71. 3:00trying to complete here how is that
72. 3:04task coded up for example if we are
73. 3:07trying to perform image compression
74. 3:10there are um it is very different than
75. 3:13try a task of trying to play a game of
76. 3:16go um
77. 3:18each of these tasks can be implemented
78. 3:21by using different algorithms
79. 3:23and these algorithms can have vastly
80. 3:26different number of instructions that
81. 3:27they
82. 3:28need to be executed for example
83. 3:31things that we have learned in 61a
84. 3:34whether it depends as of all then
85. 3:37or of n squared matter a lot here
86. 3:42then the number of instructions for a
87. 3:45program
88. 3:46will greatly depend on the programming
89. 3:48language some
90. 3:49higher level programming language will
91. 3:51have languages will have a very
92. 3:53compact description of of a task
93. 3:56but may blow up to very many assembly
94. 3:59instructions it will be much more
95. 4:01efficient perhaps to code up this task
96. 4:03in assembly
97. 4:05but we have seen that already coding up
98. 4:08things in assembly is tedious
99. 4:10and not everybody is willing to do that
100. 4:13especially not for every task that you
101. 4:14would like to code up
102. 4:17that is tightly combined with the
103. 4:19compiler compiler performance and
104. 4:21compiler effort
105. 4:22some compilers are going to generate
106. 4:26assembly level code that has less
107. 4:29instructions
108. 4:30but some other compilers may
109. 4:33generate more instructions and finally
110. 4:37it really depends on the isa
111. 4:39some isas like most of the risks
112. 4:43will require more assembly level
113. 4:45instructions than
114. 4:46cisc type processors
115. 4:50but this cannot be locked in isolation
116. 4:53the next metric that is
117. 4:55really important is the second component
118. 4:57of our iron law
119. 4:59which is the average number of clock
120. 5:02cycles
121. 5:02that it takes to execute one instruction
122. 5:06so in some architectures in our
123. 5:08architecture we always used one
124. 5:10cycle to execute one instruction but in
125. 5:11some architectures
126. 5:13that number or in some implementations
127. 5:15of the same isa
128. 5:16that number may vary so the number of
129. 5:19clock cycles per instruction depends on
130. 5:22the isa
131. 5:23but within the isa we can have different
132. 5:25processor electric
133. 5:27implementations that have different
134. 5:30cpi average cpi numbers and if you look
135. 5:33at different
136. 5:34you know performance measurements
137. 5:38you'll find out that different
138. 5:40processors from intel
139. 5:41have different cpi's among them
140. 5:45you know higher class ones versus lower
141. 5:47class ones
142. 5:48and then different implementations of
143. 5:52the x86 architecture
144. 5:54may have different cpi's between amd and
145. 5:56intel
146. 6:00in our case we always used one
147. 6:05cycle to execute every instruction so
148. 6:06far but that will change
149. 6:08um in some cases there will be more
150. 6:11complex instructions um
151. 6:12at a higher level language or in the
152. 6:15assembly language
153. 6:16in higher language like string copy in
154. 6:19in
155. 6:20c is going to take way more than one
156. 6:23cycle
157. 6:24to copy a string
158. 6:27finally we are going to soon see
159. 6:30so-called superscalar processors
160. 6:32where that cycle average cycle per
161. 6:34instruction
162. 6:35is much less than one meaning that we
163. 6:37are be simultaneously
164. 6:39executing multiple instructions in one
165. 6:41cycle
166. 6:43finally a third component of our law
167. 6:46is the time for each cycle
168. 6:49or how long does it take to complete the
169. 6:53cycle or one over the frequency of a
170. 6:54processor
171. 6:56this is determined by the
172. 6:57microarchitecture you know
173. 7:00how many logic gates are
174. 7:03in our critical path
175. 7:06but it is not just about counting the
176. 7:09number of logic gates it is
177. 7:11relating the delay of each
178. 7:14logic gates to the technology uh
179. 7:17an inverter in one technology
180. 7:20will be a lot faster than an inverter in
181. 7:24a different technology
182. 7:26we generally use cmos so inverters in
183. 7:29five nanometer technologies
184. 7:30are faster than inverters in 28
185. 7:33nanometer technologies and much faster
186. 7:35than inverters than say
187. 7:3690 nanometer cmos technologies when we
188. 7:39say
189. 7:405 nanometers or um
190. 7:4590 nanometers we refer to the minimum
191. 7:47features of that technology
192. 7:48how short generally a
193. 7:52a gate or a wire can be
194. 7:55in a particular technology
195. 7:58and then finally within the same
196. 8:01technology
197. 8:01we may have different classes of
198. 8:04implementations that have different
199. 8:06power budgets for example we will
200. 8:09encounter desktop processors that
201. 8:11run at clock frequencies that are higher
202. 8:14than four gigahertz but they burn
203. 8:17in excess of 100 watts on the other hand
204. 8:20and
205. 8:20and they have um
206. 8:23you know so they can run at excess of
207. 8:27four gigahertz
208. 8:28on the other hand those processors that
209. 8:30we'll find in our
210. 8:31cell phones will be running at two two
211. 8:34and a half gigahertz
212. 8:35but they're going to have a lot lower
213. 8:38power
214. 8:39and that is going to be largely due
215. 8:42to the fact that they are often running
216. 8:44at a lower supply voltage
217. 8:46lower supply voltage as we'll see in the
218. 8:49next segment
219. 8:50saves us energy but makes things run
220. 8:53slower
221. 8:55okay let's take a look at the
222. 8:58speed trade-off example that puts
223. 9:00together these three components of the
224. 9:02iron law
225. 9:03so this tries to illustrate that just
226. 9:06buying a processor based on clock
227. 9:09frequency
228. 9:10may not always matter you may not get
229. 9:13the best machine for the task so for
230. 9:16example
231. 9:17we have one task that you would like to
232. 9:19implement here which is the image
233. 9:20compression
234. 9:21and that image compression on two
235. 9:24processors because they may be in
236. 9:25different
237. 9:26implemented thing different isas or
238. 9:30may be described in a different
239. 9:33framework
240. 9:33can have different number of
241. 9:35instructions that are needed
242. 9:37to execute to complete the task to
243. 9:38compress the image
244. 9:40so for example in processor a we may
245. 9:43need
246. 9:43just one million instructions and
247. 9:45processor b may need
248. 9:4650 more may need 1.5 million
249. 9:49instructions
250. 9:50now when we look at the average cpi
251. 9:55processor b may be better because it
252. 9:57takes only
253. 9:58one cycle to execute an instruction on
254. 10:00the average on the other hand
255. 10:02processor b takes two and a half cycles
256. 10:04to execute that instruction
257. 10:06and finally um the clock rate of
258. 10:10processor a
259. 10:12may be better may it may run at two and
260. 10:14a half gigahertz processor b
261. 10:15may run at two gigahertz so what's the
262. 10:19what is the total execution time well
263. 10:21when we multiply all of these together
264. 10:22we'll find out that processor b
265. 10:24completes the task faster than the
266. 10:27processor a
267. 10:29although it is
268. 10:32worse in two of its performance metrics
269. 10:35it takes more
270. 10:36instructions to complete the task and it
271. 10:39runs at a slower
272. 10:41clock rate but it its cpi
273. 10:44is significantly better and it helps it
274. 10:47overcome the other disadvantages
275. 10:51so keep that in mind whenever trying to
276. 10:56buy for example a processor
277. 11:00we are going to make a quick break here
278. 11:02and come back
279. 11:03and discuss the energy efficiency
280. 11:05because it also
281. 11:06affects the performance so see you in a
282. 11:09bit