04 03 Threads and Concurrency Part 3 — Transcript

Full transcript

1. 0:08[Music]
2. 0:22all right so next would be the thread
3. 0:24libraries
4. 0:26so thread libraries provide programmers
5. 0:29with an api for creating and managing
6. 0:31threats
7. 0:32okay
8. 0:33so thread libraries may be implemented
9. 0:36either in the user space or the kernel
10. 0:39space
11. 0:40so the former involves an
12. 0:42api or api functions implemented solely
13. 0:46within the user space with no kernel
14. 0:48support all right
15. 0:51so
16. 0:52the latter involved system calls
17. 0:54okay so and requires a kernel
18. 0:57with tread library support
19. 0:59so there are three main thread libraries
20. 1:02in use today okay so these are the posix
21. 1:05threads
22. 1:06okay our p threads the win32 threads and
23. 1:09the java threads
24. 1:10all right
25. 1:12now let's talk about the p trends okay
26. 1:16so p threads may be provided either a
27. 1:19user level or a kernel level
28. 1:22okay so the posix standard is an ieee
29. 1:251003 that 1c defines the specification
30. 1:29for pthreads not the implementation
31. 1:32okay so specification not an
32. 1:34implementation
33. 1:36so global variables are shared amongst
34. 1:39all threads
35. 1:40okay
36. 1:41and one chart can wait for the others
37. 1:45to rejoin before continuing that's p
38. 1:47threads
39. 1:54all right so next would be a java thread
40. 1:57so java threads are managed by the java
41. 1:59virtual machine okay so all java
42. 2:02programs use a threads okay or this java
43. 2:06threads
44. 2:07even common single threaded ones so the
45. 2:10creation of the new threads requires
46. 2:13objects that
47. 2:14implement
48. 2:16the runnable interface which means
49. 2:19they contain a method public void run so
50. 2:23remember that in programming okay so any
51. 2:26descendant of the thread class will
52. 2:28naturally contain such method so in
53. 2:30practice
54. 2:31okay
55. 2:32so
56. 2:33the run method must be overridden
57. 2:37provided for the thread to have any
58. 2:40practical
59. 2:41functionality
60. 2:42okay so java trends may be created by
61. 2:45extending the thread class implementing
62. 2:47the runnable interfaces
63. 2:49and standard practice is to implement a
64. 2:51runnable interface
65. 2:54so the next one would be implicit
66. 2:55threading
67. 2:57okay
68. 2:59so implicit
69. 3:01threading
70. 3:03shifts the version of addressing the
71. 3:05programming challenges
72. 3:07outlined
73. 3:08on this discussion okay
74. 3:11and the programmer
75. 3:14is using this implicit threading growing
76. 3:17in popularity as the number of threads
77. 3:19increase
78. 3:20okay program correctness
79. 3:22more difficult with explicit threats
80. 3:25okay
81. 3:26so creation and management of the
82. 3:28threads done by compilers
83. 3:32run time libraries rather than
84. 3:34programmers so five methods are explored
85. 3:36we've got the thread pools
86. 3:38fork and join the openmp
87. 3:41the grand central dispatch and the intel
88. 3:43threading building blocks
89. 3:46okay so let's start with thread pulls
90. 3:49so in thread pulls creating a new thread
91. 3:52every time one is needed
92. 3:55and then deleting it when it is done can
93. 3:58be inefficient
94. 4:00and can also lead to a very large or
95. 4:03unlimited number of threads being
96. 4:05created
97. 4:07okay
98. 4:08so advantages well usually slightly
99. 4:10faster to service a request with an
100. 4:12existing thread done creating a new
101. 4:14thread so allows the number of threads
102. 4:17in applications
103. 4:18to be bound to the size of the pool
104. 4:22and separating the task to be performed
105. 4:24from mechanics
106. 4:26of creating tasks allow different uh
107. 4:29strategies
108. 4:30for running tasks okay so example tasks
109. 4:33could be scheduled to run periodically
110. 4:36okay
111. 4:37so the win32 okay the win32 api provides
112. 4:42thread pulls
113. 4:44through the pull function
114. 4:45function
115. 4:46okay so java also provides support for
116. 4:49thread pulls through the java that you
117. 4:52till that concurrent package
118. 4:54and apple supports thread pools under
119. 4:56the grand central dispatch architecture
120. 5:00okay
121. 5:02so what are the threading issues
122. 5:04so semantics
123. 5:06of forks and exec system calls signal
124. 5:09handling thread cancellation of target
125. 5:12thread
126. 5:12okay thread local storage and scheduler
127. 5:16activations
128. 5:17so let's discuss this
129. 5:20uh threading issues on the next slide
130. 5:23okay so let's start with the semantic of
131. 5:25fork and exec okay
132. 5:28so
133. 5:30does fork duplicate only by calling
134. 5:33tread
135. 5:34or all threads
136. 5:35well some unixes have two versions of
137. 5:38word okay so exact usually works as
138. 5:41normal replace the running process
139. 5:44including all threads
140. 5:46if one of the thread forks is the entire
141. 5:49process copied or is the new process
142. 5:52single threaded
143. 5:54well that's system dependent
144. 5:56okay
145. 5:57what if the new process execs right away
146. 6:01there is no need to copy all other
147. 6:03threads
148. 6:04if it doesn't then the entire process
149. 6:07should be copied well again many
150. 6:10versions of unix provides multiple
151. 6:12versions of fork
152. 6:14for
153. 6:14uh for for the purpose or for that
154. 6:17purpose okay so that's mentioned here
155. 6:20okay
156. 6:21now signal handling
157. 6:25question
158. 6:26so when a multi-threaded process
159. 6:28receives a signal
160. 6:30to what threads should that signal be
161. 6:31delivered
162. 6:34okay
163. 6:35well
164. 6:35there are four major options here okay
165. 6:38so deliver the signal to the thread to
166. 6:41which the signal applies
167. 6:44okay
168. 6:45next is deliver the signal to every
169. 6:48thread in the process
170. 6:50deliver the signal to certain threads in
171. 6:52the process
172. 6:54or assign a specific thread to receive
173. 6:56all signals in the process
174. 7:00okay so the best choice may depend on
175. 7:03which specific signal is involved
176. 7:06all right
177. 7:08okay
178. 7:10so unix allows individual threads to
179. 7:12indicate which signals
180. 7:14they are accepting and which they are
181. 7:16ignoring so however
182. 7:18the signal can only be delivered to one
183. 7:20thread
184. 7:22okay so which is generally the first
185. 7:24thread that is accepting that particular
186. 7:26signal
187. 7:27so unix provides two separate system
188. 7:30calls
189. 7:31you've got the kill
190. 7:32okay the pin signal and the featured
191. 7:35underscore
192. 7:36kill or the tilden signal so for
193. 7:38delivering signals to process s or
194. 7:41specific threads respectively
195. 7:44so windows does not support signals
196. 7:47but they can emulate using asynchronous
197. 7:50procedure calls or the apc
198. 7:54okay
199. 7:55so a pieces are delivered to specific
200. 7:58threads not on the process
201. 8:02okay so another easy start cancellation
202. 8:04how can we cancel a specific thread in
203. 8:06the process
204. 8:07okay so terminating a tread before it
205. 8:10has finished or thread to be cancelled
206. 8:13is the target thread
207. 8:14so two general approaches are being
208. 8:17observed okay we're gonna have the
209. 8:19asynchronous cancellation
210. 8:21which can sell threads immediately okay
211. 8:24and you've got the deferred cancellation
212. 8:27okay sets a flag indicating that thread
213. 8:30should cancel itself when it is
214. 8:32convenient
215. 8:34so it is then
216. 8:35up to the canceled thread to check this
217. 8:37flag periodically
218. 8:39and exit nicely
219. 8:41when it sees the flag set
220. 8:44all right
221. 8:45so
222. 8:46that's thread cancellation in there okay
223. 8:49so the shared or resource allocation
224. 8:52and the inter-threaded
225. 8:54data transfers can be problematic with a
226. 8:56synchronous cancellation
227. 8:58all right
228. 9:02so invoking a chart cancellation request
229. 9:05cancellation but actual cancellation
230. 9:07depends on the trend state
231. 9:10okay so if thread has cancellation
232. 9:12disabled cancellation remains pending
233. 9:15until thread enables it
234. 9:18okay so on linux systems thread
235. 9:21cancellation is handled through signals
236. 9:24all right
237. 9:26next would be the thread local storage
238. 9:29or the tls so the tls allows its thread
239. 9:32to have its own copy of the data so this
240. 9:35would be useful
241. 9:36when you do not have the control over
242. 9:38the thread creation process for example
243. 9:40when you're using thread pool okay
244. 9:43so different from local variables so
245. 9:46local variables visible only during a
246. 9:48single function invocation
247. 9:50well the tls visible across function
248. 9:52invocations
249. 9:54okay so this is similar to static data
250. 9:56so tls is a unique
251. 9:58twitch thread
252. 10:00okay so most data is shared among
253. 10:02threads okay
254. 10:04and this is one of the major benefits of
255. 10:06using threads in the first place however
256. 10:09sometimes threads need to be thread
257. 10:11specific data also
258. 10:13okay so most major thread libraries
259. 10:16including p threads
260. 10:18win32 or java provide support for
261. 10:21trans-specific data known as
262. 10:24this tls here
263. 10:26okay
264. 10:27so note that this is more like static
265. 10:30that local variable or the local
266. 10:33variable as mentioned
267. 10:34okay because it does not cease to exist
268. 10:37when the function ends that's a good
269. 10:39thing about it
270. 10:40all right
271. 10:42okay
272. 10:43so next would be the scheduler
273. 10:46activations
274. 10:47okay
275. 10:48so many implementations of threads
276. 10:51provide a virtual processor as an
277. 10:53interface between the user thread
278. 10:55and the kernel thread so particularly
279. 10:58for the many too many or the two-tier
280. 11:01models
281. 11:02okay so the virtual processor is known
282. 11:04as the life that process
283. 11:07or lwp
284. 11:09okay
285. 11:10so there is one-to-one correspondence
286. 11:13between the lwp
287. 11:16and the kernel threads
288. 11:18okay so the number of kernel threads
289. 11:20available
290. 11:21and hence the number of lwp's may change
291. 11:24dynamically so the application user
292. 11:29level thread library maps user threads
293. 11:32onto an available lwp's
294. 11:34okay
295. 11:35so this communication allows an
296. 11:37application to maintain the correct
297. 11:39number of kernel threads
298. 11:41right
299. 11:43okay so next would be the operating
300. 11:46system example so we'll be talking about
301. 11:48the windows and linux threads here
302. 11:50okay so for the windows threads
303. 11:53so we are using the windows api or the
304. 11:56win32 okay
305. 11:58win32 api thread library that supports
306. 12:01one to one thread model
307. 12:03so win32
308. 12:05also provides a fiverr
309. 12:08uh library
310. 12:10which supports the many-to-many model
311. 12:13okay so the win32 thread component
312. 12:15includes
313. 12:17the thread id okay registers
314. 12:21uh a user stack used in user mode and
315. 12:26the kernel stock used in kernel mode
316. 12:29okay
317. 12:31and
318. 12:32a private storage area used by various
319. 12:34runtime libraries and dynamic link
320. 12:36libraries
321. 12:38okay
322. 12:40so the primary data structures of thread
323. 12:43includes the is thread k thread and
324. 12:47this thread environment block here
325. 12:49okay so the e thread is the executive
326. 12:51thread block includes pointer to process
327. 12:54to which thread belongs to
328. 12:57and to k-thread internal space
329. 13:00so as a k-thread internal thread block
330. 13:03scheduling and synchronization
331. 13:04information
332. 13:06kernel block stop pointer to peb in a
333. 13:09kernel space
334. 13:11okay and the last one is the thread
335. 13:13environment block
336. 13:15thread id user mode stack thread local
337. 13:18storage in the user space
338. 13:20okay
339. 13:21so each thread is in the kernel space
340. 13:23k thread is the kernel space and teb is
341. 13:26in the user space
342. 13:28okay so let's talk about the linux
343. 13:30styles now
344. 13:31so linux refers to the or refers them as
345. 13:35tasks rather than threads
346. 13:38okay so in windows in other operating
347. 13:40systems we call it trends in linux it's
348. 13:42called tasks
349. 13:44all right so linux does not distinguish
350. 13:47between the process and trends
351. 13:50okay
352. 13:51so it uses the more generic term tasks
353. 13:54here so when you have heard of the term
354. 13:56task that pertains to linux threads
355. 13:59okay
356. 14:00so the traditional fork system call
357. 14:04completely duplicates a process or task
358. 14:07as described earlier okay
359. 14:09so
360. 14:11don't be confused with the term okay
361. 14:14so an alternative system called clone
362. 14:17okay
363. 14:18which allows a child task to share the
364. 14:20address space of the parent tasks
365. 14:23or process
366. 14:25okay
367. 14:26so calling clone with no flags
368. 14:28is set equivalent to fork
369. 14:31all right
370. 14:33and calling clone
371. 14:35with clone fs
372. 14:37all right you've got clone fs here clone
373. 14:40vm
374. 14:42clone
375. 14:44sync hand
376. 14:45and clone files
377. 14:48these are all equivalent to creating a
378. 14:50trend
379. 14:51okay
380. 14:52as all these data structures will be
381. 14:55shared
382. 14:57all right so the struck task struck
383. 14:59points or process data structures shared
384. 15:02or
385. 15:03unique
386. 15:05okay
387. 15:07now linux implements this using the
388. 15:09structure task strap
389. 15:13which essentially provides a level of
390. 15:15indirection
391. 15:17to task resources so when the flags are
392. 15:20not set
393. 15:21then the resources pointed to the
394. 15:23structure are copied
395. 15:25but if the flags are set
396. 15:27then only the pointers and the resources
397. 15:29are copied and hence the resources are
398. 15:32shared
399. 15:33okay
400. 15:36so think about the deep copy versus the
401. 15:39swallow copy and object oriented
402. 15:41programming
403. 15:42okay so that's how it works
404. 15:45okay supposex compliant
405. 15:47support for smp or the symmetric
406. 15:50multiprocessing
407. 15:51you've got the numa or numa the
408. 15:54non-uniform memory access
409. 15:56and a multi-core processors
410. 15:59also supports for hundreds of thousands
411. 16:02of trends
412. 16:13[Music]
413. 16:21you