IA010: Principles of Programming Languages
Concurrency
Achim Blumensath
blumens@fi.muni.cz
Faculty of Informatics, Masaryk University, Brno
Terminology
Concurrency
The order in which instructions are executed is non-deterministic.
Parallelism
Several instructions are executed at the same time.
Distributed computing
The computation is distributed over several spacially separated work
stations.
More terminology
Fibre
one path of execution in the program.
Thread, process
a fibre running in parallel.
Coroutine
a non-parallel fibre with explicit switching.
Synchronisation
communication between fibres
Implementing fibres
make_ready : (unit -> unit) -> unit; // make fibre active
schedule : unit -> unit; // switch to next fibre
spawn : (unit -> unit) -> unit; // create new fibre
yield : unit -> unit; // yield control to next fibre
Implementing fibres
let ready_fibres = Queue.make ();
type terminate = | Terminated;
let make_ready(f) {
Queue.push(ready_fibres, f);
};
let schedule() {
if Queue.is_empty(ready_fibres) then
throw Terminated
else {
let f = Queue.pop(ready_fibres);
f();
schedule()
}
};
Implementing fibres
let start_scheduler() {
try
schedule()
catch e => case e
| Terminated => ()
| else => print "uncaught exception" e
};
let spawn(f) {
letcc k {
make_ready(k);
make_ready(f);
schedule()
}
};
let yield() {
letcc k { make_ready(k); schedule() }
};
Synchronisation primitives
type condition(a);
new_condition : unit -> condition(a); // create condition
wait : condition(a) -> a; // wait for conditions
wait_multi : list(condition(a)) -> a; // wait for several
resume : condition(a) -> a -> unit; // wake up waiting fibres
Synchronisation primitives
type trigger(a) = [ cont : a -> void, triggered : bool ];
type condition(a) = [ waiting : list(trigger(a)) ];
let make_trigger(k) {
[ cont = k, triggered = False ]
};
let resume_trigger(t,v) {
if t.triggered then
()
else {
t.triggered := True;
make_ready(fun () { t.cont(v) });
}
};
Synchronisation primitives
let new_condition() {
[ waiting = [] ]
};
let resume(c,v) {
let waiting = c.waiting;
c.waiting := [];
List.iter(fun (k) { resume_trigger(k,v) },
waiting);
};
let wait(c) { let wait_multi(cs) {
letcc k { letcc k {
let t = make_trigger(k); let t = make_trigger(k);
c.waiting := [t | c.waiting]; List.iter(
schedule(); fun (c) {
} c.waiting := [t | c.waiting] },
}; cs);
schedule();
}
};
let c1 = new_condition();
let c2 = new_condition();
let f1() { let f2() {
for i = 1 .. 10 { for i = 1 .. 10 {
print "fibre1:" i; print "fibre2:" i;
resume(c2, ()); resume(c1, ());
wait(c1); wait(c2);
}; };
resume(c2, ()); resume(c1, ());
}; };
spawn(f1);
spawn(f2);
start_scheduler()
Ramifications
Concurrency makes programs much more complicated, in particular
in combination with side-effects.
• Mutual exclusion
• Deadlocks
• Race conditions
• Starvation
• Lifeness
• Fairness
Message passing
Variants
• synchronous
• asynchronous, buffered
• asynchronous, unbuffered
• one-directional
• bidirectional
Implementation
new_channel : unit -> channel(a) // create new channel
send : channel(a) -> a -> unit // send value over channel
receive : channel(a) -> a // read value from channel
Implementation
type channel_state(a) = | Free | Reading | Written(a);
type channel(a) = [ state : channel_state(a),
readers : condition(a),
writers : condition(unit) ];
let new_channel() {
[ state = Free,
readers = new_condition(),
writers = new_condition() ]
};
let receive(ch) {
case ch.state
| Free => { ch.state := Reading; wait(ch.readers) }
| Written(v) => { ch.state := Free; resume(ch.writers, ()); v }
| Reading => error
};
Implementation
type channel_state(a) = | Free | Reading | Written(a);
type channel(a) = [ state : channel_state(a),
readers : condition(a),
writers : condition(unit) ];
let send(ch,v) {
case ch.state
| Free => { ch.state := Written(v); wait(ch.writers); }
| Written(v) => error
| Reading => { ch.state := Free; resume(ch.readers,v) }
};
Example
let produce(ch) { let consume(ch) {
while True { while True {
let x = get_next_item(); let x = receive(ch);
send(ch,x); process_item(x);
}; };
}; };
let ch = new_channel();
spawn(fun () { produce(ch) });
spawn(fun () { consume(ch) });
start_scheduler()
Inversion of control: callbacks
let event_loop() {
while true {
draw ();
let event = next_event();
case event
| MouseDown(x, y) => mouse_down(x, y)
| MouseUp(x, y) => mouse_up(x, y)
| MouseMove(x, y) => mouse_move(x, y)
}
};
let state = Idle;
let draw() {
draw_window();
case state
| Idle => ()
| Dragging(obj, x, y) => let (mx,my) = get_mouse();
draw_object(obj, mx, my);
};
Inversion of control: callbacks
let mouse_down(x, y) {
case find_object_at(x, y)
| Nothing => ()
| Some(obj) => state := Dragging(obj, x, y)
};
let mouse_up(x, y) {
case state
| Idle => ()
| Dragging(obj, ox, oy) => drop_object(obj, x, y)
};
let mouse_move x y {
()
};
Inversion of control: fibres
let ch = new_channel ();
let event_loop() { let handler() {
while true { while true {
draw (); let (x1,y1) = wait_for_mouse_down();
let event = next_event(); let obj = find_object_at(x1, y1);
send(ch, event); let (x2,y2) = wait_for_mouse_up();
} drop_object(obj, x2, y2);
}; }
};
let wait_for_mouse_down() {
let event = receive(ch);
case event
| MouseDown(x,y) => return (x,y)
| else => wait_for_mouse_down();
};
Discussion
Message passing
• is simple to use, less error-prone
• scales well
• works on distributed systems
• has some overhead
Shared memory
Atomic operations
test_and_set : location -> a -> a;
compare_and_swap : location -> a -> a -> bool;
Locks/mutexes
type lock;
lock : lock -> unit;
unlock : lock -> unit;
• reentrant, non-reentrant
Lock implementation
type lock = [ locked : bool; waiting : condition() ];
let new_lock() {
[ locked = False, waiting = new_condition() ]
};
let lock(l) {
while test_and_set(l.locked, True) {
wait(l.waiting)
}
};
let unlock(l) {
l.locked := False;
resume(l.waiting, ());
};
Special syntax
lock name {
...
}
⇒
lock(name);
...
unlock(name);
Lock-free stack: non-threaded
type node(a) = [ next : node(a), data : a ];
type stack(a) = [ head : node(a) ];
let push(st : stack(a), x : a) : unit {
let n = [ next = null, data = x ];
let old_head = st.head;
n.next := old_head;
st.head := n;
};
let pop(st : stack(a)) : a {
let old_head = st.head;
if old_head == null then
throw Empty;
let new_head = old_head.next;
st.head := new_head;
return old_head.data;
};
Lock-free stack: incorrectly threadedtype node(a) = [ next : node(a), data : a ];
type stack(a) = [ head : node(a) ];
let push(st : stack(a), x : a) : unit {
let n = [ next = null, data = x ];
while true {
let old_head = atomic_read(&st.head);
n.next := old_head;
if compare_and_swap(&st.head, old_head, n) then
return;
};
};
let pop(st : stack(a)) : a {
while true {
let old_head = atomic_read(&st.head);
if old_head == null then
throw Empty;
let new_head = old_head.next;
if compare_and_swap(&st, old_head, new_head) then
return old_head.data;
};
};
Lock-free stack: correctly threaded
type node(a) = [ next : node(a), data : a ];
type stack(a) = [ head : node(a), count : int ];
let push(st : stack(a), x : a) : unit {
let n = [ next = null, data = x ];
while true {
let old_head = atomic_read(&st.head);
n.next := old_head;
if compare_and_swap(&st.head, old_head, n) then
return;
};
};
Lock-free stack: correctly threaded
let pop(st : stack(a)) : a {
while true {
let old_head = atomic_read(&st.head);
if old_head == null then
throw Empty;
let old_count = atomic_read(&st.count);
let new_head = old_head.next;
let new_count = old_count + 1;
if compare_and_swap2(&st,
old_head, old_count,
new_head, new_count) then {
return old_head.data;
}
};
};
Condition variables
type cvar = [ cond : condition, lock : lock ];
let new_cvar(l) {
[ cond = new_condition(), lock = l ];
};
let wait_cvar(c) {
unlock(c.lock);
wait(c.cond);
lock(c.lock);
};
let resume_cvar(c) {
resume(c.cond, ());
};
Semaphores
type semaphore;
increment : semaphore -> unit;
decrement : semaphore -> unit;
type semaphore = [
count : int,
lock : lock,
waiting : cvar
];
let new_semaphore() {
let l = new_lock();
let cv = new_cvar(l);
[ count = 0,
lock = l,
waiting = cv ]
};
Semaphores
let increment(sem) {
lock(sem.lock);
sem.count := sem.count + 1;
resume_cvar(sem.waiting); // this automatically unlocks the lock
};
let decrement(sem) {
lock(sem.lock);
while sem.count == 0 {
wait_cvar(sem.waiting);
}
sem.count := sem.count - 1;
unlock(sem.lock);
};
Example: bounded buffer
let lock = new_semaphore();
let full = new_semaphore();
let empty = new_semaphore();
let n = 10;
let buffer = new_buffer(n);
for i = 1 .. n { increment(empty) }; // the initial value should be n
increment(lock);
let producer() { let consumer() {
for i = 0 .. 100 { for i = 0 .. 100 {
let x = ... generate a value ...; decrement(full);
decrement(empty); decrement(lock);
decrement(lock); let x = get(buffer);
put(buffer, x); increment(lock);
increment(lock); increment(empty);
increment(full); ... process the value x ...
} }
}; };
Monitors
type node(a) = [ ... ];
type queue(a) = [
lock : lock,
front : node(a),
back : node(a)
];
let pop(q) {
lock(q.lock);
... remove the first node from the list ...
unlock(q.lock);
... return the data stored in the removed node ...
};
let push(q,x) {
lock(q.lock);
... add a new node at the end of the list ...
unlock(q.lock);
};
Transactional memory
⟨expr⟩ = . . . atomic{ ⟨expr⟩ } abort retry
• easy to use
• avoids deadlocks and race conditions
• has considerable overhead
• works well if there is litte contention, but not if there is much
Discussion
Shared-memory
• very efficient
• extremely error prone
• does not scale well
• good primitives for implementation of more high-level
synchronisation mechanisms
Reactive programming
Value of expressions changes with time.
⇒ purly functional computation with streams
let a = ... mouse.x ... mouse.y ... ;
let b = ... time ... ;
let c = ... a ... b ... ;
let num_clicks = count(mouse_button);
let str = sprintf("The mouse is at (%d, %d).", mouse.x, mouse.y);
let fld = make_text_field(str);
Reactive programming
Value of expressions changes with time.
⇒ purly functional computation with streams
let a = ... mouse.x ... mouse.y ... ;
let b = ... time ... ;
let c = ... a ... b ... ;
let num_clicks = count(mouse_button);
let str = sprintf("The mouse is at (%d, %d).", mouse.x, mouse.y);
let fld = make_text_field(str);
Discussion
• easy to use
• high composability
• avoids deadlocks and race conditions
• famework rather restrictive
• mainly intended for user interfaces, in particular web pages