Skip to content

Commit

Permalink
Sleeping barbers
Browse files Browse the repository at this point in the history
  • Loading branch information
polytypic committed Dec 24, 2023
1 parent 9a5c521 commit cd03d81
Showing 1 changed file with 265 additions and 0 deletions.
265 changes: 265 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -75,6 +75,7 @@ is distributed under the [ISC license](LICENSE.md).
- [Programming with transactional data structures](#programming-with-transactional-data-structures)
- [The dining philosophers problem](#the-dining-philosophers-problem)
- [A transactional LRU cache](#a-transactional-lru-cache)
- [The sleeping barbers problem](#the-sleeping-barbers-problem)
- [Designing lock-free algorithms with k-CAS](#designing-lock-free-algorithms-with-k-cas)
- [Understand performance](#understand-performance)
- [Minimize accesses](#minimize-accesses)
Expand Down Expand Up @@ -1036,6 +1037,270 @@ val a_cache : (int, string) cache =
As an exercise, implement an operation to `remove` associations from a cache and
an operation to change the capacity of the cache.

#### The sleeping barbers problem

The
[sleeping barber problem](https://en.wikipedia.org/wiki/Sleeping_barber_problem)
is another classic communication and synchronization problem. Let's write a
solution using **Kcas**.

There are
[many ways to solve the problem](https://en.wikipedia.org/wiki/Sleeping_barber_problem#Solutions)
and, in particular, there are concise and subtle implementations using
semaphores or mutexes. Instead of transliterating a solution using semaphores,
our approach uses queues and other concurrent data structures. We also solve the
generalized problem with multiple barbers and we also implement a mechanism to
close the barbershop. In addition, we abstract the concept of a barbershop,
where barbers and customers interact. All of this makes our solution longer than
the well known semaphore based solution. On the other hand, one might argue that
our solution is a more direct transliteration of the problem. Our solution also
avoids the starvation problem by using queues.

Let's begin by abstracting customer

```ocaml
type customer = {
notify_hair_has_been_cut : 'x.xt:'x Xt.t -> unit;
}
```

and barber

```ocaml
type barber = {
wake_up : 'x.xt:'x Xt.t -> customer -> unit;
}
```

actors. The idea is that barbers notify customers after finishing their haircut
and, adhering to the problem description, customers wake up sleeping barbers.

A barbershop consists of any number of barbers and waiting customers and can be
marked as closed:

```ocaml
type barbershop = {
sleeping_barbers : barber Queue.t;
waiting_customers : customer Queue.t;
is_closed : bool Loc.t;
}
```

The barbershop constructor does not limit the number of barbers, which are
assumed to bring their own chairs, but does require a specification of the
number of waiting room chairs for customers:

```ocaml
# let barbershop ~num_waiting_chairs =
let sleeping_barbers = Queue.create ()
and waiting_customers = Queue.create ~capacity:num_waiting_chairs ()
and is_closed = Loc.make false in
{ sleeping_barbers; waiting_customers; is_closed }
val barbershop : num_waiting_chairs:int -> barbershop = <fun>
```

Although the `barbershop` type is not abstract, we treat it as such, so we
provide a transactional predicate to check whether the barbershop is closed or
not:

```ocaml
# let is_closed ~xt bs = Xt.get ~xt bs.is_closed
val is_closed : xt:'a Xt.t -> barbershop -> bool = <fun>
```

To `close` a barbershop we set the `is_closed` location to `true` and clear both
the sleeping barbers and waiting customers queues:

```ocaml
# let close ~xt bs =
Xt.set ~xt bs.is_closed true;
Queue.Xt.clear ~xt bs.sleeping_barbers;
Queue.Xt.clear ~xt bs.waiting_customers
val close : xt:'a Xt.t -> barbershop -> unit = <fun>
```

A barber can try to get a customer sitting on a waiting room chair:

```ocaml
# let get_sitting_customer_opt ~xt bs =
Queue.Xt.take_opt ~xt bs.waiting_customers
val get_sitting_customer_opt : xt:'a Xt.t -> barbershop -> customer option =
<fun>
```

Or may go to sleep on the barber's own chair:

```ocaml
# let sleep ~xt bs barber =
if not (is_closed ~xt bs) then
Queue.Xt.add ~xt barber bs.sleeping_barbers
val sleep : xt:'a Xt.t -> barbershop -> barber -> unit = <fun>
```

Note that the `sleep` transaction uses the `is_closed` predicate. Barbers, as
well as customers, must leave the shop in case it is closed.

A customer can try to find a sleeping barber:

```ocaml
# let get_sleeping_barber_opt ~xt bs =
Queue.Xt.take_opt ~xt bs.sleeping_barbers
val get_sleeping_barber_opt : xt:'a Xt.t -> barbershop -> barber option =
<fun>
```

Or sit on a waiting room chair:

```ocaml
# let try_sitting ~xt bs customer =
not (is_closed ~xt bs) &&
Queue.Xt.try_add ~xt customer bs.waiting_customers
val try_sitting : xt:'a Xt.t -> barbershop -> customer -> bool = <fun>
```

The above `try_sitting` transaction is non-blocking. In case the
`waiting_customers` queue is full, it will return `false`. With the `customer`
actor implementation we'll look at shortly this would mean that customers would
busy-wait, which works, but potentially wastes energy. Here is a blocking
version of `try_sitting`:

```ocaml
# let try_sitting ~xt bs customer =
not (is_closed ~xt bs) &&
begin
Queue.Xt.add ~xt customer bs.waiting_customers;
true
end
val try_sitting : xt:'a Xt.t -> barbershop -> customer -> bool = <fun>
```

Both of the above `try_sitting` transactions work with the `customer` actor
we'll see shortly, but with the latter blocking version we avoid busy-wait.

The above constitutes the barbershop abstraction which is a kind of passive
concurrent data structure. Let's then implement the active participants of the
problem.

A customer tries to get a haircut. When a customer enter the barbershop he first
tries to find a sleeping barber. If none is available, the customer then tries
to sit on a waiting room chair. If both fail, then the customer has no option
except to retry. Otherwise the customer waits to get a haircut. If the shop is
closed, the customer exits. Here is the `customer` actor:

```ocaml
# let customer shop cuts =
let clean = Mvar.create None in
let self = { notify_hair_has_been_cut = Mvar.Xt.put clean true } in
while not (Xt.commit { tx = is_closed shop }) do
let get_barber_opt ~xt =
match get_sleeping_barber_opt ~xt shop with
| None ->
try_sitting ~xt shop self
| Some barber ->
barber.wake_up ~xt self;
true
in
if Xt.commit { tx = get_barber_opt } then
let try_await_haircut ~xt =
not (is_closed ~xt shop) &&
Mvar.Xt.take ~xt clean
in
if Xt.commit { tx = try_await_haircut } then
Loc.incr cuts
done
val customer : barbershop -> int Loc.t -> unit = <fun>
```

A barber tries to get a customer to give a haircut. A barber first looks for a
customer from the waiting room. If none is available, the barber goes to sleep
waiting for a wakeup from a customer. After obtaining a customer in either way,
the barber gives a haircut to the customer. Otherwise the shop must be closed
and the barber exits. Here is the `barber` actor:

```ocaml
# let barber shop cuts =
let customer = Mvar.create None in
let self = { wake_up = Mvar.Xt.put customer } in
while not (Xt.commit { tx = is_closed shop }) do
let cut customer =
Xt.commit { tx = customer.notify_hair_has_been_cut };
Loc.incr cuts
in
let get_customer_opt ~xt =
match get_sitting_customer_opt ~xt shop with
| Some _ as some -> some
| None ->
sleep ~xt shop self;
None
in
match Xt.commit { tx = get_customer_opt } with
| Some customer -> cut customer
| None ->
let await_wakeup_opt ~xt =
if is_closed ~xt shop then None
else Some (Mvar.Xt.take ~xt customer)
in
match Xt.commit { tx = await_wakeup_opt } with
| Some customer -> cut customer
| None -> ()
done
val barber : barbershop -> int Loc.t -> unit = <fun>
```

To run the problem, a barbershop is created with given number of waiting room
chairs, is populated by given number of barbers, and a given number of customers
are spawned. Once each barber has given and each customer has received a given
number of haircuts the shop is closed. This termination condition seeks to
demonstrate that no actor is starved. Here is the `sleeping_barbers` setup:

```ocaml
# let sleeping_barbers ~barbers
~num_waiting_chairs
~customers
~cuts_per_actor =
assert (0 < barbers
&& 0 <= num_waiting_chairs
&& 0 <= customers
&& 0 <= cuts_per_actor);
let shop = barbershop ~num_waiting_chairs in
let barbers = Array.init barbers @@ fun _ ->
let cuts = Loc.make 0 in
(cuts, Domain.spawn @@ (fun () -> barber shop cuts))
and customers = Array.init customers @@ fun _ ->
let cuts = Loc.make 0 in
(cuts, Domain.spawn @@ (fun () -> customer shop cuts))
in
let agents = Array.append barbers customers in
while agents
|> Array.map fst
|> Array.exists @@ fun c ->
Loc.get c < cuts_per_actor do
Domain.cpu_relax ()
done;
Xt.commit { tx = close shop };
agents
|> Array.map snd
|> Array.iter Domain.join
val sleeping_barbers :
barbers:int ->
num_waiting_chairs:int -> customers:int -> cuts_per_actor:int -> unit =
<fun>
```

Finally, let's try our solution:

```ocaml
# sleeping_barbers ~barbers:2
~num_waiting_chairs:1
~customers:4
~cuts_per_actor:10
- : unit = ()
```

Like mentioned in the beginning, this is not the most concise solution of the
sleeping barbers problem, but hopefully this solution can be understood
relatively easily with respect to the problem description.

## Designing lock-free algorithms with k-CAS

The key benefit of k-CAS, or k-CAS-n-CMP, and transactions in particular, is
Expand Down

0 comments on commit cd03d81

Please sign in to comment.