Administrivia. Talks and other opportunities: Expect HW on functions in ASM (printing binary trees) soon

Transcription

1 Threads 2/9/18

2 Administrivia Talks and other opportunities: Game designer and developer talk: Wed noon, Alumni Hall Room 302 (extra credit!) Networking, resume, interview: Wed 4pm, Alumni Hall Room 219 Bitcoin and digital currencies, 5-6pm tonight Expect HW on functions in ASM (printing binary trees) soon

3 Thread Unit of program execution has own program counter, stack, etc All threads of a process share address space same global variables, state of open files etc

4 Moore s law Figure: Herb Sutter The free lunch is over: A fundamental turn toward concurrency in software Dr. Dobb's Journal, 30(3), March

5 How hot is your CPU? yer_detailpage&v=7ubncn6v_gk#t=30

6 Parallel computing in the small

7 Why have multiple threads? Performance run like this instead of this

8 Why have multiple threads? Performance run like this instead of this Responsiveness one thread runs user interface while others compute in background (ex: mobile platforms, web servers)

9 Two relevant concepts Parallelism Using more resources to complete job faster Ex: multiple cooks splitting food prep Concurrency Managing access to shared resources Ex: two cooks both trying to get dish into oven

10 Situation: You call a professor to see if they are free to meet. They say yes, but another student is in their office by the time you get there. Is this situation an example of an issue with parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

11 Situation: You call a professor to see if they are free to meet. They say yes, but another student is in their office by the time you get there. Is this situation an example of an issue with parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

12 Situation: For a group project, you and your teammates divvy up the tasks to be completed and set up a meeting to put it all together. Is this situation an example of parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

13 Situation: For a group project, you and your teammates divvy up the tasks to be completed and set up a meeting to put it all together. Is this situation an example of parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

14 Fork-join pattern

15 Speedup Speedup = Serial (non-parallel) running time Parallel running time Linear speedup: speedup equal to the number of processing elements Sublinear speedup: less than this Superlinear speedup: more than this

16 Is speedup a topic related to parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

17 Is speedup a topic related to parallelism or concurrency? A. Primarily parallelism B. Primarily concurrency C. Equally both D. Neither E. What are we talking about again?

18 Why not linear speedup? (1) Some parts of the code can t run in parallel Initialization I/O critical sections: areas where we ensure at most one thread is running

19 Why not linear speedup? (1) If B = fraction of program that must run serially T 1 = total time on 1 processing element What is best possible time on p elements? A. T 1 /p + B B. T 1 B/p C. T 1 (1-B)/p + B D. T 1 (1-B)/p + T 1 B E. None of the above

20 Why not linear speedup? (1) If B = fraction of program that must run serially T 1 = total time on 1 processing element What is best possible time on p elements? A. T 1 /p + B B. T 1 B/p C. T 1 (1-B)/p + B D. T 1 (1-B)/p + T 1 B (called Amdahl s Law) E. None of the above

21 Why not linear speedup? (2) Poor load balance:

22 Why not linear speedup? (3) Overhead Extra instructions needed for running in parallel Examples: creating and destroying threads calls needed to coordinate threads or communicate between them changes to algorithm needed to expose parallelism or improve load balance

23 Race conditions Logic errors caused by interactions through shared variables Example: processing ATM withdrawal Operation Balance Read current value (100) $100 Perform calculation (80) $100 Store new value $80

24 Race conditions Logic errors caused by interactions through shared variables Example: processing ATM withdrawal Operation 1 Operation 2 Balance Read current value (100) $100 Perform calculation (80) Read current value (100) $100 Store new value Perform calculation (80) $80 Store new value $80

25 Solving race conditions One solution: locks acquire: block if lock is held, mark lock as held release: mark lock as not held, unblock one waiting thread (if any)

26 Solving race conditions One solution: locks acquire: block if lock is held, mark lock as held release: mark lock as not held, unblock one waiting thread (if any) Usage: acquire lock do critical section release lock

27 Construction blocks one lane of a two-lane highway so that all traffic must use the other lane. What parallelism/concurrency concept does this illustrate? A. Threads B. Race condition C. Critical section D. Parallel overhead E. I hate construction

28 Construction blocks one lane of a two-lane highway so that all traffic must use the other lane. What parallelism/concurrency concept does this illustrate? A. Threads B. Race condition C. Critical section D. Parallel overhead E. I hate construction

29

30 Multi-threading concepts Threads Parallelism and concurrency Fork-join Speedup Amdahl s law load balance overhead Race conditions critical sections and locks

31 How do we do this in C? pthreads: pthread_create(pthread_t, function to run, etc) pthread_join(pthread_t,etc)