[CS61C FA20] Lecture 31.3 - I/O: I/O Interrupts — Transcript

Full transcript

1. 0:00[Music]
2. 0:10hello
3. 0:10and welcome back to our discussion about
4. 0:13input output devices
5. 0:14we have introduced polling as one way to
6. 0:18work with input devices but it is
7. 0:20generally
8. 0:21inefficient especially with those that
9. 0:24have a lot of data to
10. 0:26transmit or receive it is
11. 0:29a valid technique and i certainly have
12. 0:32used it
13. 0:33in the past myself um particularly when
14. 0:36doing some prototyping when for example
15. 0:39have a raspberry pi i would like to
16. 0:41figure out if
17. 0:42some new i o device is working properly
18. 0:46the easiest thing that we can do is to
19. 0:49set up a
20. 0:50very short polling loop and see
21. 0:53what does the device do more commonly
22. 0:56the way how we are going to work with
23. 0:58io devices is through unknown mechanism
24. 1:01of interrupts
25. 1:02so we have already highlighted that
26. 1:05polling waste processor resources why
27. 1:08well simply because whether the
28. 1:10input output device has something to
29. 1:12tell us um
30. 1:14or not we anyways go to it and check how
31. 1:18is it doing if it has no data for us we
32. 1:21still spend processor cycles to go pull
33. 1:24it
34. 1:26and there is an analogy in real life you
35. 1:28know when we throw a party
36. 1:29um this wood polling would be equivalent
37. 1:32to
38. 1:33setting a timer and going to the front
39. 1:35door every minute to check if somebody
40. 1:37has arrived we don't do that because
41. 1:40people long time ago have
42. 1:41invented a doorbell to exactly prevent
43. 1:44us from doing that so when the guests
44. 1:47arrive
45. 1:48they ring the doorbell to announce
46. 1:50themselves
47. 1:51similarly an io device rings their
48. 1:54doorbell
49. 1:55when they have something to tell us
50. 1:57their doorbell
51. 1:58is the interrupt so they will raise an
52. 2:02interrupt or throw an interrupt
53. 2:04whenever they have some data to deliver
54. 2:06us or perhaps if
55. 2:08there is some some event that they would
56. 2:10like to tell us about
57. 2:15so the interrupt will
58. 2:18interrupt the current program and then
59. 2:21we will transfer
60. 2:22the control to the trap handler in the
61. 2:25operating system
62. 2:26that will handle that interrupt
63. 2:30um why interrupt surprising polling
64. 2:33because when nothing is happening
65. 2:36no news from the i o device there is
66. 2:38nothing to do
67. 2:39our regular program continues
68. 2:44if there are if the device is not
69. 2:46producing a lot of data
70. 2:48and doesn't have a lot of activity it's
71. 2:50great because we just periodically
72. 2:52will receive these interrupts and handle
73. 2:55them
74. 2:56accordingly but if the device has
75. 2:59a lot of io interrupts are expensive
76. 3:03because
77. 3:04they interrupt the current programs
78. 3:06trash the caches uh
79. 3:08trash the tlbs and we have to go save
80. 3:11the straight
81. 3:12save the state restore the state and
82. 3:14then you know warm up caches
83. 3:16and so on so they're not the best thing
84. 3:19to do generally
85. 3:22they are fine for low rate devices and
86. 3:26they're preferred to polling
87. 3:27so my keyboard and so on will use
88. 3:30interrupts
89. 3:36the in devices that produce a lot of
90. 3:40data
91. 3:41a more common way is that they will
92. 3:44initiate the data transfer with an
93. 3:46interrupt when they have something to
94. 3:48send us
95. 3:48receive the packet over wi-fi
96. 3:51but the transfers are generally if there
97. 3:55is a lot of data to be transferred
98. 3:56it is done by using a different
99. 3:59mechanism
100. 4:01so-called direct memory access or dma
101. 4:04which we'll cover
102. 4:05in the next section before we get to the
103. 4:07next section
104. 4:08there is an older way of essentially
105. 4:11orchestrated polling which is it which
106. 4:14was known as programmed io
107. 4:17or po this
108. 4:20was introduced with older ata
109. 4:23style hard drives probably you do not
110. 4:26have an
111. 4:26ata star uh kind of a hard drive
112. 4:30where essentially it was the role of a
113. 4:33processor
114. 4:35that to to initiate
115. 4:38all the data movement to and from the
116. 4:41hard drive
117. 4:42it was done in a more efficient way than
118. 4:44just using loads and stores
119. 4:46but loads and stores initiated that
120. 4:50so this cpu was typically spending some
121. 4:53number of its cycles to get the data
122. 4:57to and from the disk
123. 5:00and into the main memory but
124. 5:03that was the same cpu that was
125. 5:06being used to do the main computation so
126. 5:10handling the disk was an overhead
127. 5:12typically what would go
128. 5:14to programmed io were about five percent
129. 5:18of processor cycles
130. 5:23so these kind of things happen still
131. 5:26there are still instructions in big
132. 5:29processors that
133. 5:30are being dealt that are being used to
134. 5:32copy
135. 5:33a piece of data from one location to
136. 5:36another
137. 5:37for example if you look at google's uh
138. 5:40workload they report that about five
139. 5:42percent of cpu cycles
140. 5:44go to instructions like mem
141. 5:47copy which copies from one memory
142. 5:49location
143. 5:50to another
144. 5:54we're going to see some other
145. 5:57methods like the dma access
146. 6:01just after the break and we'll use them
147. 6:03whenever we can
148. 6:04when we have a lot of data to move
149. 6:06around see you after a quick break