[CS61C FA20] Lecture 38.1 - Dependability- Parity, ECC, RAID: Intro — Transcript

Full transcript

1. 0:03[Music]
2. 0:11hello
3. 0:12welcome back to 61c it's a brand new
4. 0:15module
5. 0:16on dependability it's a fairly short
6. 0:19module
7. 0:20we're going to just touch on it remember
8. 0:24our early story about six great ideas
9. 0:27in computer architecture so we have
10. 0:30talked about
11. 0:31the layers of abstraction that enable us
12. 0:33to build these complex systems
13. 0:35out of many many components
14. 0:38and we do that by layering many
15. 0:41layers of abstraction we have talked
16. 0:44about moore's law that enabled us to
17. 0:46integrate so many devices and build
18. 0:48these complex systems
19. 0:50we have talked about principle of
20. 0:52locality
21. 0:53and memory hierarchy that enabled us to
22. 0:56build these memory systems that look
23. 0:59infinitely fast and infinitely big
24. 1:02we have talked about parallelism as a
25. 1:04way to improve the performance
26. 1:07in the in power limited regime
27. 1:12we have also throughout the course
28. 1:14talked about
29. 1:15the this concept of of performance
30. 1:18measurement and improvements
31. 1:20primarily through the iron law of
32. 1:23compute
33. 1:24and average memory access times
34. 1:27but we haven't talked about the sixth
35. 1:30great idea
36. 1:31which is dependability by a redundancy
37. 1:35so that's what we're going to cover now
38. 1:38and why is that important well we
39. 1:42rely on our compute systems more and
40. 1:46more
41. 1:47in everyday life and in domains that are
42. 1:51just
43. 1:52not purely computing all
44. 1:55our financial transactions are performed
45. 1:58by computers
46. 1:59we drive a collection of computers
47. 2:03many people's lives are supported by
48. 2:07some sort of a computer so
49. 2:10we really need to care about that
50. 2:13dependability
51. 2:14but the fact is that the computers fail
52. 2:18they might fail transiently and we have
53. 2:21seen before i
54. 2:23i'm afraid these blue screens of that
55. 2:26and many other failure modes for
56. 2:29for computers often after a computer
57. 2:32crashes
58. 2:35um it can come back and we'll call that
59. 2:39a transient mode of failure this failure
60. 2:42may happen
61. 2:43because of a bad code well you know
62. 2:45somebody you know
63. 2:46it's not just students who forget to
64. 2:49clean the exit
65. 2:50functions production software does that
66. 2:53sometimes as well they do
67. 2:55you know send you know new release and
68. 2:56it works better but these things happen
69. 2:59but sometimes these failures transit
70. 3:01failures happen because
71. 3:03something did not go quite right in the
72. 3:06hardware
73. 3:07hardware made an error
74. 3:10if these errors persist something really
75. 3:14fails
76. 3:15in permanently in a computer
77. 3:18then often we discard them or we try to
78. 3:22repair them
79. 3:24if they are repairable
80. 3:27so in this module we are going to spend
81. 3:30some time
82. 3:32talking about how do we mitigate these
83. 3:35hardware failures by using
84. 3:38redundant components so
85. 3:42we have briefly touched on that early on
86. 3:45in the in the introductory module
87. 3:47when we can use
88. 3:51redundancy to replace a failing
89. 3:55part of a system this redundancy
90. 3:58is often encountered in memory systems
91. 4:00we have spare stuff in memories
92. 4:02um but we
93. 4:06in parallel computers we may have a
94. 4:09spare processor core you'll find out
95. 4:13that on the market you can buy now
96. 4:17eight core gpus or seven
97. 4:20core gpus well when you look at the chip
98. 4:23they're exactly the same
99. 4:25except that eight core was disabled
100. 4:29and they'll sell you that chip a little
101. 4:32bit
102. 4:33cheaper why was it disabled because it
103. 4:35was no good it was failing
104. 4:37that's okay you just have a little bit
105. 4:39less of a performance you have seven
106. 4:40instead of eight
107. 4:41and you're willing to pay less for that
108. 4:45so the way how it works is
109. 4:48that you'll have multiple replicas
110. 4:52of hardware and there'll be some kind of
111. 4:55a voting mechanism
112. 4:56where two out of three perhaps will
113. 4:59agree
114. 5:00that one plus one is equal to two and
115. 5:02we'll take that
116. 5:04as an answer this is
117. 5:07this is made a lot easier by
118. 5:10the advances in integration transistor
119. 5:13integration densities
120. 5:14so we can have more of these components
121. 5:18now there is another cache there as the
122. 5:20components are smaller as the threshold
123. 5:21or smaller
124. 5:22they have a tendency you know to more
125. 5:24frequently fail
126. 5:26but that is not increasing the
127. 5:30that the failure rate is not increasing
128. 5:32at the speed as which at which we can
129. 5:34integrate more of them
130. 5:35so we can add redundancy the other way
131. 5:40how we can improve things
132. 5:43how we can mitigate these transient
133. 5:46failures
134. 5:47is through
135. 5:50another form of redundancy
136. 5:54we can temporarily you know things can
137. 5:57temporarily go out
138. 5:58and we can still
139. 6:02continue computing without them for
140. 6:04example you know at the very
141. 6:06top level the data center may go out i
142. 6:08mean
143. 6:09there may be a hurricane um on
144. 6:12the east coast
145. 6:15the power may be out the internet may be
146. 6:17out data center goes out
147. 6:19the weather system goes away data center
148. 6:22is back up
149. 6:24the other thing that we have encountered
150. 6:26is often we use these arrays of disks
151. 6:29if a disk if a mechanical disk fails
152. 6:33it's okay because there is a way
153. 6:36to cover for that disc by having
154. 6:38redundancy
155. 6:39on the shelf so a shelf of disk disks
156. 6:43will have spares and finally
157. 6:47we can use a particular type of coding
158. 6:51to add
159. 6:54redundancy to our dram so our dram chip
160. 6:57this is a
161. 6:58dim that goes into computers if you look
162. 7:00at it it has
163. 7:01nine memory chips that correspond to
164. 7:03nine bits instead of
165. 7:05eight and the reason for that is that
166. 7:08the ninth one
167. 7:08is there to represent parity and that
168. 7:11parity is going to be used to indicate
169. 7:13if something went wrong
170. 7:16so we're going to take a quick break now
171. 7:19and then we're going to take a
172. 7:21look at how do we measure dependability
173. 7:24see you after a quick break