CS计算机代考程序代写 concurrency data structure Java assembly CPSC 213 Introduction to Computer Systems

CPSC 213 Introduction to Computer Systems
Unit 2c
Synchronization
1

Reading
‣ Companion •8
‣Text
• 2ed: 12.4-12.6, parts of 12.7
• 1ed: 13.4-13.5, (no equivalent to 12.6), parts of 13.7
2

Synchronization
CPUs (Cores)
Memory Bus
disk-read thread
disk controller
wait
some other thread
‣ We invented Threads to
• exploit parallelism do things at the same time on different processors • manage asynchrony do something else while waiting for I/O Controller
‣ But, we now have two problems
• coordinating access to memory (variables) shared by multiple threads
• control flow transfers among threads (wait until notified by another thread
‣ Synchronization is the mechanism threads use to • ensure mutual exclusion of critical sections
• wait for and notify of the occurrence of events
notify
Memory
3

The Importance of Mutual Exclusion
‣ Shared data
• data structure that could be accessed by multiple threads • typically concurrent access to shared data is a bug
‣ Critical Sections
• sections of code that access shared data
‣ Race Condition
• simultaneous access to critical section section by multiple threads
• conflicting operations on shared data structure are arbitrarily interleaved
• unpredictable (non-deterministic) program behaviour — usually a bug (a serious bug)
‣ Mutual Exclusion
• a mechanism implemented in software (with some special hardware support) • to ensure critical sections are executed by one thread at a time
• though reading and writing should be handled differently (more later)
‣ For example
• consider the implementation of a shared stack by a linked list …
4

‣ Stack implementation
void push_st (struct SE* e) {
e->next = top;
top = e;
}
struct SE* pop_st () {
struct SE* e = top;
top = (top)? top->next: 0;
return e;
}
‣ Sequential test works
void push_driver (long int n) {
struct SE* e;
while (n–)
push ((struct SE*) malloc (…));
}
push_driver (n);
pop_driver (n);
assert (top==0);
struct SE {
struct SE* next;
};
struct SE *top=0;
void pop_driver (long int n) {
struct SE* e;
while (n–) {
do {
e = pop ();
} while (!e);
free (e); }
}
5

‣ concurrent test doesn’t always work
et = uthread_create ((void* (*)(void*)) push_driver, (void*) n);
dt = uthread_create ((void* (*)(void*)) pop_driver, (void*) n);
uthread_join (et);
uthread_join (dt);
assert (top==0);
malloc: *** error for object 0x1022a8fa0: pointer being freed was not allocated
‣ what is wrong?
void push_st (struct SE* e) {
e->next = top;
top = e;
}
struct SE* pop_st () {
struct SE* e = top;
top = (top)? top->next: 0;
return e;
}
6

‣ The bug
• push and pop are critical sections on the shared stack
• they run in parallel so their operations are arbitrarily interleaved • sometimes, this interleaving corrupts the data structure
top X
void push_st (struct SE* e) {
e->next = top;
top = e;
}
1. e->next = top
6. top = e
struct SE* pop_st () {
struct SE* e = top;
top = (top)? top->next: 0;
return e;
}
2.e =top
3. top = top->next 4. return e
5. free e
7

Mutual Exclusion using locks
‣ lock semantics
• a lock is either held by a thread or available
• at most one thread can hold a lock at a time
• a thread attempting to acquire a lock that is already held is forced to wait
‣ lock primitives
• lock acquire lock, wait if necessary
• unlock release lock, allowing another thread to acquire if waiting
‣using locks for the shared stack
void push_cs (struct SE* e) { lock (&aLock);
push_st (e);
unlock (&aLock);
}
struct SE* pop_cs () {
struct SE* e;
lock (&aLock);
e = pop_st (); unlock (&aLock);
return e; }
8

Implementing Simple Locks
‣Here’s a first cut
• use a shared global variable for synchronization
• lock loops until the variable is 0 and then sets it to 1 • unlock sets the variable to 0
int lock = 0;
void lock (int* lock) {
while (*lock==1) {}
*lock = 1;
}
void unlock (int* lock) {
*lock = 0;
}
• why doesn’t this work?
9

‣ We now have a race in the lock code
Thread A
void lock (int* lock) {
while (*lock==1) {}
*lock = 1;
Thread B
void lock (int* lock) {
while (*lock==1) {}
*lock = 1;
}} 1. read *lock==0, exit loop
3. *lock = 1
4. return with lock held
2. read *lock==0, exit loop
5. *lock = 1, return
6. return with lock held
Both threads think they hold the lock …
10

‣ The race exists even at the machine-code level
• two instructions acquire lock: one to read it free, one to set it held • but read by another thread and interpose between these two
ld $lock, r1
ld $1, r2
loop: ld (r1), r0
beq r0, free
br loop
free: st r2, (r1)
Thread A
ld (r1), r0
st r2, (r1)
lock appears free
Another thread reads lock
acquire lock
Thread B
ld (r1), r0
st r2, (r1)
11

Atomic Memory Exchange Instruction
‣ We need a new instruction
• to atomically read and write a memory location
• with no intervening access to that memory location from any other thread allowed
‣ Atomicity
• is a general property in systems
• where a group of operations are performed as a single, indivisible unit
‣ The Atomic Memory Exchange
• one type of atomic memory instruction (there are other types)
• group a load and store together atomically
• exchanging the value of a register and a memory location
Name
Semantics
Assembly
atomic exchange
r[v] ← m[r[a]] m[r[a]] ← r[v]
xchg (ra), rv
12

Implementing Atomic Exchange
CPUs (Cores)
Memory Bus
Memory
‣ Can not be implemented just by CPU
• must synchronize accross multiple CPUs
• accessing the same memory location at the same time
‣ Implemented by Memory Bus
• memory bus synchronizes every CPUs access to memory
• the two parts of the exchange (read + write) are coupled on bus
• bus ensures that no other memory transaction can intervene
• this instruction is much slower, higher overhead than normal read or write
13

Spinlock
‣ A Spinlock is
• a lock where waiter spins on looping memory reads until lock is acquired • also called “busy waiting” lock
‣ Implementation using Atomic Exchange • spin on atomic memory operation
• that attempts to acquire lock while • atomically reading its old value
ld $lock, r1
ld $1, r0
loop: xchg (r1), r0
beq r0, held
br loop held:
• but there is a problem: atomic-exchange is an expensive instruction
14

‣ Spin first on normal read
• normal reads are very fast and efficient compared to exchange
• use normal read in loop until lock appears free
• when lock appears free use exchange to try to grab it • if exchange fails then go back to normal read
ld $lock, r1
loop: ld (r1), r0
beq r0, try
br loop
try: ld $1, r0
xchg (r1), r0
beq r0, held
br loop
held:
‣ Busy-waiting pros and cons
• Spinlocks are necessary and okay if spinner only waits a short time • But, using a spinlock to wait for a long time, wastes CPU cycles
15

Blocking Locks
‣ If a thread may wait a long time
• it should block so that other threads can run
• it will then unblock when it becomes runnable (lock available or event notification)
‣ Blocking locks for mutual exclusion
• if lock is held, locker puts itself on waiter queue and blocks
• when lock is unlocked, unlocker restarts one thread on waiter queue
‣ Blocking locks for event notification
• waiting thread puts itself on a a waiter queue and blocks
• notifying thread restarts one thread on waiter queue (or perhaps all)
‣ Implementing blocking locks presents a problem
• lock data structure includes a waiter queue and a few other things
• data structure is shared by multiple threads; lock operations are critical sections • mutual exclusion can be provided by blocking locks (they aren’t implemented yet) • and so, we need to use spinlocks to implement blocking locks (this gets tricky)
16

Implementing a Blocking Lock
‣ Lock data structure
struct blocking_lock {
int spinlock;
int held;
uthread_queue_t waiter_queue;
};
‣ The lock operation
void lock (struct blocking_lock* l) {
spinlock_lock (&l->spinlock);
while (l->held) {
enqueue (&waiter_queue, uthread_self ());
spinlock_unlock (&l->spinlock);
uthread_switch (ready_queue_dequeue (), TS_BLOCKED);
spinlock_lock (&l->spinlock);
}
l->held = 1;
spinlock_unlock (&l->spinlock);
}
17

‣ The unlock operation
void unlock (struct blocking_lock* l) {
uthread_t* waiter_thread;
spinlock_lock (&l->spinlock);
l->held = 0;
waiter_thread = dequeue (&l->waiter_queue);
spinlock_unlock (&->spinlock);
waiter_thread->state = TS_RUNABLE;
ready_queue_enqueue (waiter_thread);
}
18

Blocking Lock Example Scenario
Thread A
1. calls lock()
3. grabs spinlock
5. acquires blocking lock 6. releases spinlock
7. returns from lock()
Thread B
2. calls lock()
4. tries to grab spinlock, but spins
3. grabs spinlock
4. queues itself on watier list 5. releases spinlock
6. blocks
Thread C
9. calls unlock()
10. grabs spinlock
11. releases lock
12. restarts a Thread B 13. releases spinlock 14. returns from unlock()
16. scheduled
17. grabs spinlock
18. acquires blocking lock 19. releases spinlock
20. returns from lock()
15. yields, blocks or stops
thread running spinlock held blocking lock held
8. scheduled
19

Blocking vs Busy Waiting
‣ Spinlocks
• Pros and Cons
– uncontended locking has low overhead – contending for lock has high cost
• Use when
– critical section is small
– contention is expected to be minimal
– event wait is expected to be very short – when implementing Blocking locks
‣ Blocking Locks
• Pros and Cons
– uncontended locking has higher overhead – contending for lock has no cost
• Use when
– lock may be head for some time – when contention is high
– when event wait may be long
20

Monitors and Conditions
‣ Mutual exclusion plus inter-thread synchronization • introduced by Tony Hoare and Per Brinch Hansen circ. 1974
• basis for synchronization primitives in Java etc.
‣ Monitor
• is a mutual-exclusion lock
• primitives are enter (lock) and exit (unlock)
‣ Condition Variable
• allows threads to synchronize with each other
• wait
• notify
• notify_all
• can only be accessed from inside of a monitor (i.e, with monitor lock held)
blocks until a subsequent signal operation on the variable unblocks waiter, but continues to hold monitor (Hansen) unblocks all waiters and continues to hold monitor
21

Using Conditions
‣ Basic formulation
• one thread enters monitor and may wait for a condition to be established
monitor {
while (!x)
wait (); }
• another thread enters monitor, establishes condition and signals waiter
monitor {
x = true;
notify (); }
‣ wait exists the monitor and blocks thread
• before waiter blocks, it exists monitor to allow other threads to enter
• when wait unblocks, it re-enters monitor, waiting/blocking to enter if necessary • note: other threads may have been in monitor between wait call and return
22

‣ notify awakens one thread • does not release monitor
• waiter does not run until notifier exits monitor
• a third thread could intervene and enter monitor before waiter
• waiter must thus re-check wait condition
And not
monitor {
x = true;
notify ();
monitor {
while (!x)
}}
wait ();
‣ notify_all awakens all threads
• may wakeup too many
• okay since threads re-check wait condition and re-wait if necessary
monitor {
x += n;
monitor {
while (!x)
notify_all (); }}
wait ();
monitor {
if (!x)
wait (); }
23

Drinking Beer Example
‣ Beer pitcher is shared data structure with these operations • pour from pitcher into glass
• refill pitcher
‣ Implementation goal
• synchronize access to the shared pitcher
• pouring from an empty pitcher requires waiting for it to be filled • filling pitcher releases waiters
void pour () {
monitor {
while (glasses==0)
wait ();
glasses–;
void refill (int n) {
monitor {
for (int i=0; inext = 0;
if (queue->tail)
queue->tail->next = thread;
queue->tail = thread;
if (queue->head==0)
queue->head = queue->tail;
}
uthread_t* dequeue (uthread_queue_t* queue) {
uthread_t* thread;
if (queue->head) {
thread = queue->head;
queue->head = queue->head->next;
if (queue->head==0)
queue->tail=0;
} else
thread=0;
return thread;
}
27

‣ Adding Mutual Exclusion
void enqueue (uthread_queue_t* queue, uthread_t* thread) { uthread_monitor_enter (&queue->monitor);
thread->next = 0;
if (queue->tail)
queue->tail->next = thread;
queue->tail = thread;
if (queue->head==0)
queue->head = queue->tail;
uthread_monitor_exit (&queue->monitor);
}
uthread_t* dequeue (uthread_queue_t* queue) { uthread_t* thread;
uthread_monitor_enter (&queue->monitor);
if (queue->head) {
thread = queue->head;
queue->head = queue->head->next;
if (queue->head==0)
queue->tail=0;
} else
thread=0;
uthread_monitor_exit (&queue->monitor); return thread;
}
28

‣ Now have dequeue wait for item if queue is empty • classical producer-consumer model with each in different thread
– e.g., producer enqueues video frames consumer thread dequeues them for display
void enqueue (uthread_queue_t* queue, uthread_t* thread) { uthread_monitor_enter (&queue->monitor);
thread->next = 0;
if (queue->tail)
queue->tail->next = thread;
queue->tail = thread;
if (queue->head==0)
queue->head = queue->tail;
uthread_cv_notify (&queue->not_empty);
uthread_monitor_exit (&queue->monitor); }
uthread_t* dequeue (uthread_queue_t* queue) { uthread_t* thread;
uthread_monitor_enter (&queue->monitor);
while (queue->head==0)
uthread_cv_wait (&queue->not_empty);
thread = queue->head;
queue->head = queue->head->next;
if (queue->head==0)
queue->tail=0; uthread_monitor_exit (&queue->monitor); return thread;
}
29

Some Questions About Example
uthread_t* dequeue (uthread_queue_t* queue) {
uthread_t* thread;
uthread_monitor_enter (&queue->monitor);
while (queue->head==0)
uthread_cv_wait (&queue->not_empty);
thread = queue->head;
queue->head = queue->head->next;
if (queue->head==0)
queue->tail=0;
uthread_monitor_exit (&queue->monitor);
return thread;
}
‣ Why is does dequeue have a while loop to check for non-empty?
‣ Why must condition variable be associated with specific monitor?
‣ Why can’t use use condition variable outside of monitor? • this is called a naked used of the condition variable
• this is actually required sometimes … can you think where (BONUS)? – Experience with Processes and Monitors with Mesa, Lampson and Redell, 1980
30

Implementing Condition Variables
‣ Some key observations
• wait, notify and notify_all are called while monitor is held
• the monitor must be held when they return
• wait must release monitor before locking and re-acquire before returning
‣ Implementation • in the lab
• look carefully at the implementations of monitor enter and exit
• understand how these are similar to wait and notify
• use this code as a guide
• you also have the code for semaphores, which you might also find helpful
31

Reader-Writer Monitors
‣ If we classify critical sections as • reader if only reads the shared data • writer if updates the shared data
‣ Then we can weaken the mutual exclusion constraint • writers require exclusive access to the monitor
• but, a group of readers can access monitor concurrently
‣ Reader-Writer Monitors • monitor state is one of
– free, held-for-reading, or held • monitor_enter ()
– waits for monitor to be free then sets its state to held
• monitor_enter_read_only ()
– waits for monitor to be free or held-for-reading, then sets is state to head-for-reading – increment reader count
Monitor
• monitor_exit ()
– if held, then set state to free
– if held-for-reading, then decrement reader count and set state to free if reader count is 0
32

‣ Policy question
• monitor state is head-for-reading
• thread A calls monitor_enter() and blocks waiting for monitor to be free • thread B calls monitor_enter_read_only(); what do we do?
‣ Disallowing new readers while writer is waiting
• is the fair thing to do
• thread A has been waiting longer than B, shouldn’t it get the monitor first?
‣ Allowing new readers while writer is waiting
• may lead to faster programs by increasing concurrency
• if readers must WAIT for old readers and writer to finish, less work is done
‣ What should we do
• normally either provide a fair implementation
• or allow programmer to choose (that’s what Java does)
33

Semaphores
‣ Introduced by Edsger Dijkstra for the THE System circa 1968 • recall that he also introduced the “process” (aka “thread”) for this system
• was fearful of asynchrony, Semaphores synchronize interrupts • synchronization primitive provide by UNIX to applications
‣ A Semaphore is
• an atomic counter that can never be less than 0
• attempting to make counter negative blocks calling thread
‣ P (s)
• try to decrement s (prolaag for probeer te varlagen in Dutch)
• atomically blocks until s >0 then decrement s
‣ V (s)
• increment s (verhogen in Dutch)
• atomically increase s unblocking threads waiting in P as appropriate
34

Using Semaphores to Drink Beer
‣ Use semaphore to store glasses head by pitcher • set initial value of empty when creating it
uthread_semaphore_t* glasses = uthread_create_semaphore (0); ‣ Pouring and refilling don’t require a monitor
void pour () {
uthread_P (glasses);
}
void refill (int n) {
for (int i=0; iresult = tuple->arg0 + tuple->arg1;
uthread_V (tuple->barrier);
return 0;
}
uthread_semaphore_t* barrier = uthread_semaphore_create (0);
struct arg_tuple a0 = {1,2,0,barrier};
struct arg_tuple a1 = {3,4,0,barrier};
uthread_init (1);
uthread_create (add, &a0);
uthread_create (add, &a1);
uthread_P (barrier);
uthread_P (barrier);
printf (“%d %d\n”, a0.result, a1.result);
‣ Barrier (global)
• In a system of N threads with no parent
• All threads must arrive, before any can continue … and should work repeatedly
37

‣ Implementing Monitors • initial value of semaphore is 1
• lock is P()
• unlock is V()
‣ Implementing Condition Variables • this is the warm beer problem
• it took until 2003 before we actually got this right
• for further reading
– Andrew D. Birrell. “Implementing Condition Variables with Semaphores”, 2003. – Google “semaphores condition variables birrell”
38

Synchronization in Java (5)
‣ Monitors using the Lock interface
• a few variants allow interruptibility, just trying lock, …
Lock l = …;
l.lock ();
try {

} finally {
l.unlock ();
}
• multiple-reader single writer locks
ReadWriteLock l = …;
Lock rl = l.readLock ();
Lock wl = l.writeLock ();
Lock l = …;
try {
l.lockInterruptibly ();
try {

} finally {
l.unlock ();
} catch (InterruptedException ie) {}
}
39

‣ Condition variables
• await is wait (replaces Object wait)
• signal or signalAll is “notify” (replaces Object notify, notifyAll)
class Beer {
Lock l = …;
Condition notEmpty = l.newCondition ();
int glasses = 0;
void pour () throws InterruptedException {
l.lock ();
try {
while (glasses==0)
notEmpty.await ();
glasses–;
} finaly {
l.unlock ();
}
}
void refill (int n) throws InterruptedException {
l.lock ();
try {
glasses += n;
notEmpty.signalAll ();
} finaly {
l.unlock ();
}}}
40

‣ Semaphore class
• acquire () or acquire (n) is P() or P(n) • release () or release (n) is V() or V(n)
class Beer {
Semaphore glasses = new Semaphore (0);
void pour () throws InterruptedException {
glasses.acquire ();
}
void refill (int n) throws InterruptedException {
glasses.release (n);
} }
‣ Lock-free Atomic Variables
• AtomicX where X in {Boolean, Integer, IntegerArray, Reference, …}
• atomic operations such as getAndAdd(), compareAndSet(), …
– e.g., x.compareAndSet (y,z) atomically sets x=z iff x==y and returns true iff set occurred
41

Lock-Free Atomic Stack in Java
‣ Recall the problem with concurrent stack
void push_st (struct SE* e) {
e->next = top;
top = e;
}
struct SE* pop_st () {
struct SE* e = top;
top = (top)? top->next: 0;
return e;
}
• a pop could intervene between two steps of push, corrupting linked list
top X
• we solved this problem using locks to ensure mutual exclusion
• now … solve without locks, using atomic compare-and-set of top
42

class Element {
Element* next;
}
class Stack {
AtomcReference top;
Stack () {
top.set (NULL);
}
void push () {
Element t;
Element e = new Element ();
do {
t = top.get ();
e.next = t; X } while (!top.compareAndSet (t, e));
} }
top X
43

Problems with Concurrency
‣ Race Condition
• competing, unsynchronized access to shared variable – from multiple threads
– at least one of the threads is attempting to update the variable
• solved with synchronization
– guaranteeing mutual exclusion for competing accesses
– but the language does not help you see what data might be shared — can be very hard
‣ Deadlock
• multiple competing actions wait for each other preventing any to complete
• what can cause deadlock? – MONITORS
– CONDITION VARIABLES – SEMAPHORES
44

Recursive Monitor Entry
‣ What should we do for a program like this
void foo () { uthread_monitor_enter (mon);
count–;
if (count>0)
foo();
uthread_monitor_exit (mon);
}
‣ Here is implementation of lock, is this okay?
void lock (struct blocking_lock* l) {
spinlock_lock (&l->spinlock);
while (l->held) {
enqueue (&waiter_queue, uthread_self ());
spinlock_unlock (&l->spinlock);
uthread_switch (ready_queue_dequeue (), TS_BLOCKED);
spinlock_lock (&l->spinlock);
}
l->held = 1;
spinlock_unlock (&l->spinlock);
}
45

‣ if we try to lock the monitor again it is a deadlock • the thread will hold the monitor when it tries to enter
• the thread will wait for itself, and thus never wake up
‣ allow a thread that holds the monitor to enter again
void uthread_monitor_enter (uthread_monitor_t* monitor) { spinlock_lock (&monitor->spinlock);
while (monitor->holder && monitor->holder!=uthread_self()) {
enqueue (&monitor->waiter_queue, uthread_self ());
spinlock_unlock (&monitor->spinlock);
uthread_stop (TS_BLOCKED);
spinlock_lock (&monitor->spinlock);
}
monitor->holder = uthread_self ();
spinlock_unlock (&monitor->spinlock);
}
46

Systems with multiple monitors
‣ We have already seen this with semaphores
‣ Consider a system with two monitors a, and b
void foo() {
uthread_monitor_enter (a);
uthread_monitor_exit (a);
void bar() {
uthread_monitor_enter (b);
}}
uthread_monitor_exit (b);
void x() {
uthread_monitor_enter (a);
bar();
uthread_monitor_exit (a);
void y() {
uthread_monitor_enter (b);
}}
Any problems so far?
foo();
uthread_monitor_exit (b);
What about now?
47

Waiter Graph Can Show Deadlocks
‣ Waiter graph
• edge from lock to thread if thread HOLDs lock
• edge from thread to lock if thread WANTs lock • a cycle indicates deadlock
void foo() {
uthread_monitor_enter (a);
uthread_monitor_exit (a);
}
void bar() {
uthread_monitor_enter (b);
uthread_monitor_exit (b);
}
void x() {
uthread_monitor_enter (a);
bar();
uthread_monitor_exit (a);
}
void y() {
uthread_monitor_enter (b);
foo();
uthread_monitor_exit (b);
}
x
ab
y
48

The Dining Philosophers Problem
‣ Formulated by Edsger Dijkstra to explain deadlock (circa 1965) • 5 computers competed for access to 5 shared tape drives
‣ Re-told by Tony Hoare
• 5 philosophers sit at a round table with fork placed in between each
– fork to left and right of each philosopher and each can use only these 2 forks
• they are either eating or thinking
– while eating they are not thinking and while thinking they are not eating – they never speak to each other
• large bowl of spaghetti at centre of table requires 2 forks to serve – dig in …
• deadlock
– every philosopher holds fork to left waiting for fork to right (or vice versa) – how might you solve this problem?
• starvation (aka livelock)
– philosophers still starve (ever get both forks) due to timing problem, but avoid deadlock – for example:
49

Avoiding Deadlock
‣ Don’t use multiple threads
• you’ll have many idle CPU cores and write asynchronous code
‣ Don’t use shared variables
• if threads don’t access shared data, no need for synchronization
‣ Use only one lock at a time
• deadlock is not possible, unless thread forgets to unlock
‣ Organize locks into precedence hierarchy • each lock is assigned a unique precedence number
• before thread X acquires a lock i, it must hold all higher precedence locks • ensures that any thread holding i can not be waiting for X
‣ Detect and destroy
• if you can’t avoid deadlock, detect when it has occurred
• break deadlock by terminating threads (e.g., sending them an exception)
50

Synchronization Summary
‣ Spinlock
• one acquirer at a time, busy-wait until acquired
• need atomic read-write memory operation, implemented in hardware • use for locks held for short periods (or when minimal lock contention)
‣ Monitors and Condition Variables
• blocking locks, stop thread while it is waiting
• monitor guarantees mutual exclusion
• condition variables wait/notify provides control transfer among threads
‣ Semaphores
• blocking atomic counter, stop thread if counter would go negative • introduced to coordinate asynchronous resource use
• use to implement barriers or monitors
• use to implement something like condition variables, but not quite
‣ Problems, problems, problems
• race conditions to be avoided using synchronization
• deadlock/livelock to be avoided using synchronization carefully
51