U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science Computer Systems Principles Processes & Threads Emery Berger and Mark Corner University of Massachusetts Amherst

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 2 Outline  Processes  Threads  Basic synchronization  Bake-off

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science What is a Process?  One or more threads running in an address space 3

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science What is an address space?  The collection of data and code for the program organized in an organized (addressable) region. 4

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science Why do we need processes?  A process (with a single thread) supports a serial flow of execution  Is that good enough?  What if something goes wrong?  What if something takes a long time?  What if we have more than one user? 5

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 6 Processes vs. Threads…  Both useful –parallel programming & concurrency –Hide latency –Maximize CPU utilization –Handle multiple, asynchronous events  But: different programming styles, performance characteristics, and more

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 7 Processes  Process: execution context (PC, registers) + address space, files, etc.  Basic unit of execution

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 8 Process API  UNIX: –fork() – create copy of current process Different return value Copy-on-write –exec() – replace process w/ executable  Windows: –CreateProcess (…) 10 arguments

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 9 Translation Lookaside Buffer  TLB: fast, fully associative memory –Stores page numbers (key), frame (value) in which they are stored  Copy-on-write: protect pages, copy on first write

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 10 Processes Example Do we need the sleep?

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 11 Quick Activity  Print out fib(0) … fib(100), in parallel –Main loop should fork off children (returns pid of child) int pid; pid = fork(); if (pid == 0) { // I’m child process } else { // I’m parent process } –Wait for children (waitpid(pid,0,0))

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 12 Exercise // Print fib(0)... fib(100) in parallel. int main() { int cids[101]; // Saved child pids. for (int i = 0; i < 101; i++) { // Fork off a bunch of processes to run fib(i). int cid = fork(); if (cid == 0) { // I am the child. cout << "Fib(" << i << ") = " << fib(i) << endl; return 0; // Now the child is done, so exit (IMPORTANT!). } else { // I am the parent. Store the child’s pid somewhere. cids[i] = cid; } // Wait for all the children. for (int i = 0; i < 101; i++) { waitpid (cids[i], 0, 0); // Wait for the ith child. } return 0; }

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 13 Communication  Processes: –Input = state before fork() –Output = return value argument to exit()  But: how can processes communicate during execution?

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 14 IPC  signals –Processes can send & receive ints associated with particular signal numbers Process must set up signal handlers –Best for rare events (SIGSEGV) Not terribly useful for parallel or concurrent programming  pipes –Communication channels – easy & fast –Just like UNIX command line ls | wc -l

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 15 int main() { int pfds[2]; pipe(pfds); if (!fork()) { close(1); /* close normal stdout */ dup(pfds[1]); /* make stdout same as pfds[1] */ close(pfds[0]); /* we don't need this */ execlp("ls", "ls", NULL); } else { close(0); /* close normal stdin */ dup(pfds[0]); /* make stdin same as pfds[0] */ close(pfds[1]); /* we don't need this */ execlp("wc", "wc", "-l", NULL); } } Pipe example

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 16 IPC, continued  sockets –Explicit message passing –Can distribute processes anywhere  mmap (common and aws0m hack) –All processes map same file into fixed memory location –Objects in region shared across processes –Use process-level synchronization Much more expensive than threads…

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 17 Threads  Processes - everything in distinct address space  Threads – same address space (& files, sockets, etc.)

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 18 Threads API  UNIX (POSIX): –pthread_create() – start separate thread executing function –pthread_join() – wait for thread to complete  Windows: –CreateThread (…) only 6 arguments!

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 19 Threads example #include void * run (void * d) { int q = ((int) d); int v = 0; for (int i = 0; i < q; i++) { v = v + expensiveComputation(i); } return (void *) v; } main() { pthread_t t1, t2; int *r1, *r2; pthread_create (&t1, NULL, run, (int *)100); pthread_create (&t2, NLL, run, (int *)100); pthread_join (t1, (void **) &r1); pthread_join (t2, (void **) &r2); cout << "r1 = " << *r1 << ", r2 = " << *r2 << endl; }

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 20 Exercise  Compute fib(100), in parallel –Function definition: void * funName (void * arg) –Creating thread: int tid; pthread_create(&tid, NULL, funName, argument); –Wait for child: pthread_join(tid,&result)

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 21 Exercise

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 22 Communication  In threads, everything shared except: stacks, registers & thread-specific data –Old way: pthread_setspecific pthread_getspecific –New way: __thread static __thread int x; Easier in Java…  Updates of shared state must be synchronized

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 23 Bake-off  Processes or threads? –Performance –Flexibility / Ease-of-use –Robustness

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 24 Scheduling

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 25 Context Switch Cost  Threads – much cheaper –Stash registers, PC (“IP”), stack pointer  Processes – above plus –Process context –TLB shootdown  Process switches more expensive, or require long quanta

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 26 Flexibility / Ease-of-use  Processes – more flexible –Easy to spawn remotely ssh foo.cs.umass.edu “ls -l” –Can communicate via sockets = can be distributed across cluster / Internet –Requires explicit communication or risky hackery  Threads –Communicate through memory – must be on same machine –Require thread-safe code

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 27 Robustness  Processes – far more robust –Processes isolated from other processes Process dies – no effect –Apache 1.x  Threads: –If one thread crashes (e.g., derefs NULL), whole process terminates –Then there’s the stack size problem –Apache 2.x…

U NIVERSITY OF M ASSACHUSETTS A MHERST Department of Computer Science 28 The End