Shared Programming Address Space Platforms
Adapted from Ananth Grama, Anshul Gupta, George Karypis, and Vipin Kumar ``Introduction to Parallel Computing'', Addison Wesley, 2003.
Overview
Thread Basics The POSIX Thread API Synchronization Primitives in Pthreads Controlling Thread and Synchronization Attributes Composite Synchronization Constructs OpenMP: a Standard for Directive Based Parallel Programming
Process vs Threads
Thread Basics
Each thread has its own stack, SP, PC, registers, etc. Threads share global variables and heap. Caveat: writing programs in which shared space is treated as a flat address space may give poor performance
Locality is just as important in shared-memory machines as it is in distributed-memory machines
Thread Basics
The logical machine model of a threadbased programming paradigm.
The POSIX Thread API
Commonly referred to as Pthreads, POSIX has emerged as the standard threads API, supported by most vendors. The concepts discussed here are largely independent of the API and can be used for programming with other thread APIs (NT threads, Solaris threads, Java threads, etc.) as well.
Thread Basics: Creation and Termination
Creating Pthreads:
#include <pthread.h> int pthread_create ( pthread_t *thread_handle, const pthread_attr_t *attribute, void * (*thread_function)(void *), void *arg);
Thread is created and it starts to execute thread_function with parameter arg
Terminating threads
Thread terminated when: o it returns from its starting routine, or o it makes a call to pthread_exit() Main thread
exits with pthread_exit(): other threads will continue to execute Otherwise: othereads automatically terminated
Cleanup:
pthread_exit() routine does not close files any files opened inside the thread will remain open after the thread is terminated.
#include <pthread.h> #include <stdio.h> #include <stdlib.h> #define NUM_THREADS 5 void *PrintHello(void *threadid) { printf("\n%d: Hello World!\n", threadid); pthread_exit(NULL); } int main(int argc, char *argv[]) { pthread_t threads[NUM_THREADS]; int rc, t; for(t=0;t<NUM_THREADS;t++){ printf("Creating thread %d\n", t); rc = pthread_create(&threads[t], NULL, PrintHello, (void *)t); if (rc){ printf("ERROR; return code from pthread_create() is %d\n", rc); exit(-1); } } pthread_exit(NULL); }
Output
Creating thread 0 Creating thread 1 0: Hello World! 1: Hello World! Creating thread 2 Creating thread 3 2: Hello World! 3: Hello World! Creating thread 4 4: Hello World!
Synchronizing threads
"Joining" is one way to synchronize threads (not used very often) pthread_join (threadid,status)
The pthread_join() function blocks the calling thread until the specified thread terminates. The programmer can obtain the target thread's termination return status if it was specified in the target thread's call to pthread_exit().
Threads: Example 2
1.0
1.0 Area of circle = pi*0.25 Area of square = 1 So if we shoot randomly into square, probability of hitting circle is pi*0.25 Estimating value of pi:
generate a large number of random values inside the unit square see what fraction of them fall inside circle and multiply by 4
Simple example of Monte Carlo method: estimate some value by repeated sampling of some space Monte Carlo method can be easily parallelized provided each parallel thread generates independent random numbers
Threads: Example2
#include <pthread.h> #include <stdlib.h> #define MAX_THREADS 512 void *compute_pi (void *); .... main() { ... pthread_t p_threads[MAX_THREADS]; pthread_attr_t attr; pthread_attr_init (&attr); for (i=0; i< num_threads; i++) { hits[i] = i; pthread_create(&p_threads[i], &attr, compute_pi, (void *) &hits[i]); } for (i=0; i< num_threads; i++) { pthread_join(p_threads[i], NULL); total_hits += hits[i]; } ... }
Threads: Example2 (contd.)
void *compute_pi (void *s) { int seed, i, *hit_pointer; double rand_no_x, rand_no_y; int local_hits; hit_pointer = (int *) s; seed = *hit_pointer; local_hits = 0; for (i = 0; i < sample_points_per_thread; i++) { rand_no_x =(double)(rand_r(&seed))/(double)((2<<14)-1); rand_no_y =(double)(rand_r(&seed))/(double)((2<<14)-1); if (((rand_no_x - 0.5) * (rand_no_x - 0.5) + (rand_no_y - 0.5) * (rand_no_y - 0.5)) < 0.25) local_hits ++; seed *= i; } *hit_pointer = local_hits; pthread_exit(0); }
Synchronizing threads
Style of computing shown in Example 2 is sometimes called fork-join parallelism
fork
join
This style of parallel execution in which threads only synchronize at the end is quite rare Usually, threads need to synchronize during their execution
Need for synchronization
Two common scenarios:
Mutual exclusion
Shared resource such as variable or device Only one thread at a time can access resource Critical section: portion of code that should be executed by only thread at a time
Producer-consumer
One thread (producer) generates a sequence of values Another thread (consumer) reads these values Values are communicated by writing them into a shared buffer Producer must block if buffer is full Consumer must block if buffer is empty
Need for Mutual Exclusion
When multiple threads attempt to manipulate the same data item, the results can often be chaotic if proper care is not taken to synchronize them. Consider:
Assume that there are two threads, the initial value of best_cost is 100, and the values of my_cost are 50 and 75 at threads t1 and t2. Depending on the schedule of the threads, the value of best_cost could be 50 or 75!
Thread 1 reads best_cost (100) Thread 2 reads best_cost (100) Thread 1 writes best_cost (50) Thread 2 writes best_cost (75)
/* each thread tries to update variable best_cost as follows */ if (my_cost < best_cost) best_cost = my_cost;
The value 75 does not seem right because it would not arise in a sequential execution of the same algorithm
General problem
The code in the previous example is called a critical section
Several threads may try to execute code in critical section but only one should succeed at a time
Problem arises very often when writing threaded code
Thread A want to read and write one or more variables in critical section While it is doing that, other threads should be excluded from accessing those variables
Solution: lock
Threads compete for acquiring lock Pthreads implementation guarantees that only one thread will succeed in acquiring lock Successful thread enters critical section, performs its activity When critical section is done, lock is released
Discussion
Lock is implemented by variable with two states: available/not_available When thread tries to acquire a lock and state of lock is available, its state is changed to not_available, and thread is informed that it can proceed Pthreads implementation ensures that this is done atomically: cannot interrupt the processor or contextswitch during the lock acquire Detail:
locks also have queues that hold ids of threads waiting to acquire lock When one thread releases a lock, next thread in queue is informed it has acquired lock, and it can proceed This is more efficient than alternatives like busy-waiting in which a thread repeatedly tries to acquire a lock This is also a way to ensure some notion of fairness any thread that wants to acquire a lock can succeed ultimately even if other threads want to acquire the lock an unbounded number of times
Mutex in Pthreads
The Pthreads API provides the following functions for handling mutex-locks:
Lock creation
int pthread_mutex_init ( pthread_mutex_t *mutex_lock, const pthread_mutexattr_t *lock_attr);
Acquiring lock
int pthread_mutex_lock ( pthread_mutex_t *mutex_lock);
Releasing lock
int pthread_mutex_unlock ( pthread_mutex_t *mutex_lock);
Correct Mutual Exclusion
We can now write our previously incorrect critical section as:
pthread_mutex_t minimum_value_lock; ... main() { .... pthread_mutex_init(&minimum_value_lock, NULL); .... } void *find_min(void *list_ptr) { .... pthread_mutex_lock(&minimum_value_lock); if (my_min < minimum_value) critical section minimum_value = my_min; /* and unlock the mutex */ pthread_mutex_unlock(&minimum_value_lock); }
Critical sections
For performance, it is important to keep critical sections as small as possible While one thread is within critical section, all others threads that want to enter the critical section are blocked It is up to the programmer to ensure that locks are used correctly to protect variables in critical sections Thread A lock(l) x:= ..x.. unlock(l) Thread B Thread C lock(l) x:= ..x.. x: = x unlock(l)
This program may fail to execute correctly because programmer forgot to use locks in Thread C
Producer-Consumer Using Locks
Two threads
Producer: produces data Consumer: consumes data
Shared buffer is used to communicate data from producer to consumer
Buffer can contain one data value (in this example) Flag is associated with buffer to indicate buffer has valid data
Consumer must not read data from buffer unless there is valid data Producer must not overwrite data in buffer before it is read by consumer
Producer-Consumer Using Locks
pthread_mutex_t data_queue_lock; int data_available; ... main() { .... data_available = 0; pthread_mutex_init(&data_queue_lock, NULL); .... } void *producer(void *producer_thread_data) { .... while (!done()) { inserted = 0; create_data(&my_data); while (inserted == 0) { pthread_mutex_lock(&data_queue_lock); if (data_available == 0) { insert_into_queue(my_data); data_available = 1; inserted = 1; } pthread_mutex_unlock(&data_queue_lock); } } }
Producer-Consumer Using Locks
void *consumer(void *consumer_thread_data) { int extracted; struct data my_data; /* local data structure declarations */ while (!done()) { extracted = 0; while (extracted == 0) { pthread_mutex_lock(&data_queue_lock); if (data_available == 1) { extract_from_queue(&my_data); data_available = 0; extracted = 1; } pthread_mutex_unlock(&data_queue_lock); } process_data(my_data); } }
Types of Mutexes
Pthreads supports three types of mutexes - normal, recursive, and error-check. A normal mutex deadlocks if a thread that already has a lock tries a second lock on it. A recursive mutex allows a single thread to lock a mutex as many times as it wants. It simply increments a count on the number of locks. A lock is relinquished by a thread when the count becomes zero. An error check mutex reports an error when a thread with a lock tries to lock it again (as opposed to deadlocking in the first case, or granting the lock, as in the second case). The type of the mutex can be set in the attributes object before it is passed at time of initialization.
Reducing lock overhead
It is often possible to reduce the idling overhead associated with locks using an alternate function, pthread_mutex_trylock. int pthread_mutex_trylock (
pthread_mutex_t *mutex_lock);
If lock is available, acquire it; otherwise, return a busy error code (EBUSY) Faster than pthread_mutex_lock on typical systems since it does not have to deal with queues associated with locks for multiple threads waiting on the lock.
Alleviating Locking Overhead (Example)
/* Finding k matches in a list */ void *find_entries(void *start_pointer) { /* This is the thread function */ struct database_record *next_record; int count; current_pointer = start_pointer; do { next_record = find_next_entry(current_pointer); count = output_record(next_record); } while (count < requested_number_of_records); } int output_record(struct database_record *record_ptr) { int count; pthread_mutex_lock(&output_count_lock); output_count ++; count = output_count; pthread_mutex_unlock(&output_count_lock); if (count <= requested_number_of_records) print_record(record_ptr); return (count); }
Alleviating Locking Overhead (Example)
/* rewritten output_record function */ int output_record(struct database_record *record_ptr) {
int count; int lock_status; lock_status=pthread_mutex_trylock(&output_count_lock); if (lock_status == EBUSY) { } else {
insert_into_local_list(record_ptr); return(0);
count = output_count; output_count += number_on_local_list + 1; pthread_mutex_unlock(&output_count_lock); print_records(record_ptr, local_list, return(count + number_on_local_list + 1);
requested_number_of_records - count);
}
}
Condition Variables
Condition variables are another construct for more efficient synchronization: permit a thread to be woken up when some predicate on the data is satisifed Example: one thread produces a sequence of data items, and consumer thread must wait till there are more than n items in buffer Busy waiting is inefficient
Better to let waiting thread sleep and get notified when predicate is satisifed Solution: condition variables
Basic operations using condition variables
Thread can wait on condition variable: intuitively, thread blocks until some other thread signals that condition variable Thread can signal condition variable: release one thread waiting on condition variable Condition variables are not boolean variables!
Correct operation of condition variables requires an associated mutex as we will see later
Condition Variable Constructs
Pthreads provides the following functions for condition variables:
int pthread_cond_init(pthread_cond_t *cond, const pthread_condattr_t *attr); int pthread_cond_destroy(pthread_cond_t *cond); int pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex); int pthread_cond_signal(pthread_cond_t *cond); int pthread_cond_broadcast(pthread_cond_t *cond);
Locks associated with condition variables
Correct operation with condition variable requires an associated lock
Wait and signal must be performed while holding lock
Problem:
If thread A holds lock, calls wait on a condition variable, and then goes to sleep, how does thread B acquire lock to signal this condition variable?
Solution:
When thread A calls wait and goes to sleep, pthreads implementation automatically releases associated lock When thread A needs to be woken up in response to signal, pthreads implementation tries to reacquire lock and returns control to application program only after lock has been reacquired Signal and lock reacquire are separate events, so it is good practice to re-check that data predicate after control returns from wait
=> Use a loop around wait (shown in examples)
Producer-Consumer Using Condition Variables
pthread_cond_t cond_queue_empty, cond_queue_full; pthread_mutex_t data_queue_cond_lock; int data_available; /* other data structures here */ main() { /* declarations and initializations */ data_available = 0; pthread_init(); pthread_cond_init(&cond_queue_empty, NULL); pthread_cond_init(&cond_queue_full, NULL); pthread_mutex_init(&data_queue_cond_lock, NULL); /* create and join producer and consumer threads */ }
Producer-Consumer Using Condition Variables
void *producer(void *producer_thread_data) { int inserted; while (!done()) { create_data(); pthread_mutex_lock(&data_queue_cond_lock); while (data_available == 1) pthread_cond_wait(&cond_queue_empty, &data_queue_cond_lock); insert_into_queue(); data_available = 1; pthread_cond_signal(&cond_queue_full); pthread_mutex_unlock(&data_queue_cond_lock); } }
Producer-Consumer Using Condition Variables
void *consumer(void *consumer_thread_data) { while (!done()) { pthread_mutex_lock(&data_queue_cond_lock); while (data_available == 0) pthread_cond_wait(&cond_queue_full, &data_queue_cond_lock);
my_data = extract_from_queue(); data_available = 0; pthread_cond_signal(&cond_queue_empty); pthread_mutex_unlock(&data_queue_cond_lock); process_data(my_data);
}
}
Controlling Thread and Synchronization Attributes
The Pthreads API allows a programmer to change the default attributes of entities using attributes objects. An attributes object is a data-structure that describes entity (thread, mutex, condition variable) properties. Once these properties are set, the attributes object can be passed to the method initializing the entity. Enhances modularity, readability, ease and of modification.
Attributes Objects for Threads
Use pthread_attr_init to create an attributes object. Individual properties associated with the attributes object can be changed using the following functions:
pthread_attr_setdetachstate, pthread_attr_setguardsize_np, pthread_attr_setstacksize, pthread_attr_setinheritsched, pthread_attr_setschedpolicy, pthread_attr_setschedparam
and
Attributes Objects for Mutexes
Initialize the attrributes object using function: The function pthread_mutexattr_settype_np can be used for setting the type of mutex specified by the mutex attributes object.
pthread_mutexattr_settype_np ( pthread_mutexattr_t *attr, int type); PTHREAD_MUTEX_NORMAL_NP PTHREAD_MUTEX_RECURSIVE_NP PTHREAD_MUTEX_ERRORCHECK_NP
pthread_mutexattr_init.
Here, type specifies the type of the mutex and can take one of:
Composite Synchronization Constructs
By design, Pthreads provide support for a basic set of operations. Higher level constructs can be built using basic synchronization constructs. We discuss two such constructs - readwrite locks and barriers.
Read-Write Locks
In many applications, a data structure is read frequently but written infrequently. For such applications, we should use read-write locks. A read lock is granted when there are other threads that may already have read locks. If there is a write lock on the data (or if there are queued write locks), the thread performs a condition wait. If there are multiple threads requesting a write lock, they must perform a condition wait. With this description, we can design functions for read locks mylib_rwlock_rlock, write locks mylib_rwlock_wlock, and unlocking mylib_rwlock_unlock.
Read-Write Locks
The lock data type mylib_rwlock_t holds the following:
a count of the number of readers, the writer (a 0/1 integer specifying whether a writer is present), a condition variable readers_proceed that is signaled when readers can proceed, a condition variable writer_proceed that is signaled when one of the writers can proceed, a count pending_writers of pending writers, and a mutex read_write_lock associated with the shared data structure
Read-Write Locks
typedef struct { int readers; int writer; pthread_cond_t readers_proceed; pthread_cond_t writer_proceed; int pending_writers; pthread_mutex_t read_write_lock; } mylib_rwlock_t; void mylib_rwlock_init (mylib_rwlock_t *l) { l -> readers = l -> writer = l -> pending_writers = 0; pthread_mutex_init(&(l -> read_write_lock), NULL); pthread_cond_init(&(l -> readers_proceed), NULL); pthread_cond_init(&(l -> writer_proceed), NULL); }
Read-Write Locks
void mylib_rwlock_rlock(mylib_rwlock_t *l) { /* if there is a write lock or pending writers, perform condition wait.. else increment count of readers and grant read lock */ pthread_mutex_lock(&(l -> read_write_lock)); while ((l -> pending_writers > 0) || (l -> writer > 0)) pthread_cond_wait(&(l -> readers_proceed), &(l -> read_write_lock)); l -> readers ++; pthread_mutex_unlock(&(l -> read_write_lock)); }
Read-Write Locks
void mylib_rwlock_wlock(mylib_rwlock_t *l) { /* if there are readers or writers, increment pending writers count and wait. On being woken, decrement pending writers count and increment writer count */ pthread_mutex_lock(&(l -> read_write_lock)); while ((l -> writer > 0) || (l -> readers > 0)) { l -> pending_writers ++; pthread_cond_wait(&(l -> writer_proceed), &(l -> read_write_lock)); } l -> pending_writers --; l -> writer ++; pthread_mutex_unlock(&(l -> read_write_lock));
}
Read-Write Locks
void mylib_rwlock_unlock(mylib_rwlock_t *l) { /* if there is a write lock then unlock, else if there are read locks, decrement count of read locks. If the count is 0 and there is a pending writer, let it through, else if there are pending readers, let them all go through */ pthread_mutex_lock(&(l -> read_write_lock)); if (l -> writer > 0) l -> writer = 0; else if (l -> readers > 0) l -> readers --; pthread_mutex_unlock(&(l -> read_write_lock)); if ((l -> readers == 0) && (l -> pending_writers > 0)) pthread_cond_signal(&(l -> writer_proceed)); else if (l -> readers > 0) pthread_cond_broadcast(&(l -> readers_proceed)); }
Barriers
As in MPI, a barrier holds a thread until all threads participating in the barrier have reached it. Barriers can be implemented using a counter, a mutex and a condition variable. A single integer is used to keep track of the number of threads that have reached the barrier. If the count is less than the total number of threads, the threads execute a condition wait. The last thread entering (and setting the count to the number of threads) wakes up all the threads using a condition broadcast.
Barriers
typedef struct { pthread_mutex_t count_lock; pthread_cond_t ok_to_proceed; int count; } mylib_barrier_t; void mylib_init_barrier(mylib_barrier_t *b) { b -> count = 0; pthread_mutex_init(&(b -> count_lock), NULL); pthread_cond_init(&(b -> ok_to_proceed), NULL); }
Barriers
void mylib_barrier (mylib_barrier_t *b, int num_threads) {
pthread_mutex_lock(&(b -> count_lock)); b -> count ++; if (b -> count == num_threads) { } else
b -> count = 0; pthread_cond_broadcast(&(b -> ok_to_proceed));
while (pthread_cond_wait(&(b -> ok_to_proceed),
&(b -> count_lock)) != 0);
}
pthread_mutex_unlock(&(b -> count_lock));
Barriers
The barrier described above is called a linear barrier. The trivial lower bound on execution time of this function is therefore O(n) for n threads. This implementation of a barrier can be speeded up using multiple barrier variables organized in a tree. We use n/2 condition variable-mutex pairs for implementing a barrier for n threads. At the lowest level, threads are paired up and each pair of threads shares a single condition variable-mutex pair. Once both threads arrive, one of the two moves on, the other one waits. This process repeats up the tree. This is also called a log barrier and its runtime grows as O(log p).
Barrier
Execution time of 1000 sequential and logarithmic barriers as a function of number of threads on a 32 processor SGI Origin 2000.
Tips for Designing Asynchronous Programs
Never rely on scheduling assumptions when exchanging data. Never rely on liveness of data resulting from assumptions on scheduling. Do not rely on scheduling as a means of synchronization. Where possible, define and use group synchronizations and data replication.
Types of threads
Thread implementations:
User-level threads:
Implemented by user-level runtime library OS is unaware of threads Portable, thread scheduling can be tuned to application requirements Problem: cannot leverage multiprocessors, entire process blocks when one thread blocks
Kernel-level threads:
OS is aware of each thread and schedules them Thread operations are performed by OS Can leverage multiprocessors Problem: higher overhead, usually not quite as portable
Hybrid-level threads: Solaris
OS provides some number of kernel level threads, and each of these can create multiple user-level threads Problem: complexity
OpenMP: a Standard for Directive Based Parallel Programming
OpenMP is a directive-based API that can be used with FORTRAN, C, and C++ for programming shared address space machines. OpenMP directives provide support for concurrency, synchronization, and data handling while obviating the need for explicitly setting up mutexes, condition variables, data scope, and initialization.
OpenMP Programming Model
OpenMP directives in C and C++ are based on the #pragma compiler directives. A directive consists of a directive name followed by clauses. OpenMP programs execute serially until they encounter the parallel directive, which creates a group of threads. The main thread that encounters the parallel directive becomes the master of this group of threads and is assigned the thread id 0 within the group.
#pragma omp parallel [clause list] /* structured block */ #pragma omp directive [clause list]
OpenMP Programming Model
The clause list is used to specify conditional parallelization, number of threads, and data handling.
Conditional Parallelization: The clause if (scalar expression) determines whether the parallel construct results in creation of threads. Degree of Concurrency: The clause num_threads(integer expression) specifies the number of threads that are created. Data Handling: The clause private (variable list) indicates variables local to each thread. The clause firstprivate (variable list) is similar to the private, except values of variables are initialized to corresponding values before the parallel directive. The clause shared (variable list) indicates that variables are shared across all the threads.
OpenMP Programming Model
A sample OpenMP program along with its Pthreads translation that might be performed by an OpenMP compiler.
OpenMP Programming Model
#pragma omp parallel if (is_parallel== 1) num_threads(8) \ private (a) shared (b) firstprivate(c) { /* structured block */ }
If the value of the variable is_parallel equals one, eight threads are created. Each of these threads gets private copies of variables a and c, and shares a single value of variable b. The value of each copy of c is initialized to the value of c before the parallel directive. The default state of a variable is specified by the clause default (shared) or default (none).
Reduction Clause in OpenMP
The reduction clause specifies how multiple local copies of a variable at different threads are c...