24 Days of GHC Extensions: Static Pointers

Today we start to wrap up 24 Days of GHC Extensions with the final guest post, this time from Mathieu Boespflug of Tweag I/O. Mathieu, along with his colleagues Facundo Domínguez and Alexander Vershilov, has been working with the GHC team on a new extension - in fact, so new, you won’t even find this in a stable release of GHC! In today’s post, we’ll have a look at a new type of pointer that can be useful for distributed programming.

Distributed Static Programming

GHC already features quite the zoo of pointer types. There are bare Ptr’s (for marshalling to and from foreign objects), ForeignPtr’s (smart pointers that allow automatic memory management of the target object), weak pointers (references to objects that are ignored by the garbage collector), and StablePtr’s (pointers to objects that are pinned to a specific location in memory). GHC 7.10 will add a new beast to this list: StaticPtr, the type of pointers whose value never changes across program runs, even across program runs on different machines. The objects pointed to by static pointers are static, much in the same sense of the word as in other programming languages: their value is known at compile time and their lifetime extends across the entire run of the program. GHC 7.10 also comes with a new language extension to safely create static pointers: StaticPointers.

Why yet another pointer type? And why grace it with yet another extension?

Static pointers turn out to be incredibly useful for distributed programming. Imagine that you have a fleet of networked computers, abstractly called nodes. You’d like these nodes to collaborate, say because you also have a fair amount of data you’d like to crunch through, or because some of these nodes provide services to other nodes. Static pointers help solve the age-old question of distributed programming: how can nodes easily delegate tasks to each other?

For most programming languages, this is a thorny question to ask: support for distributing computations comes as an afterthought, so there is no first class support. But there are exceptions: Erlang is one example of a language that has escaped from research labs one way or another and natively speaks distributed. Erlang supports literally sending the code for any native (non-foreign) function from node to node. Delegating a task called myfun is a case of saying:

spawn(There, myfun)

where There is a variable containing some node identifier. This capability comes at a cost, however. It is in general hard to share optimized compiled code across a cluster of machines, which may not be running the exact same operating system or have the same system libraries available. So Erlang keeps to comparatively slow but easy to handle and easy to distribute interpreted bytecode instead. Moreover, if new code can be loaded into a running program at any moment or existing code monkey patched on-the-go, what tools do we have to reason about the resulting state of the program?

Haskell too natively speaks distributed, at least in its bleeding edge GHC variant. But at much lower cost. In a world where complete systems can be containerized using language agnostic technology, and shipped and deployed within minutes across a full scale cluster, do we really need our language runtimes to distribute code? Are we willing to accept the compromises involved? Perhaps that is a problem best solved once, for all programs in any language, using the likes of Docker or Rocket. And once our entire cluster is running instances of the same program by dint of distributing containers, all we need is a means to control which computations happen when, and where, by sharing references to functions. This works because, if all nodes are running the same program, then they all have access to the same functions.

Turning on -XStaticPointers adds a new keyword static and a new syntactic form to the language for safely creating such references: if expression e has type a, then static e has type StaticPtr a.

Static pointers in practice

For example, here’s a program that obtains a static pointer to f, and prints the info record associated with it:

module Main where

import GHC.StaticPtr

fact :: Int -> Int
fact 0 = 1
fact n = n * fact (n - 1)

main = do
  let sptr :: StaticPtr (Int -> Int)
      sptr = static fact
  print $ staticPtrInfo sptr
  print $ deRefStaticPtr sptr 10

The body of a static form can be any top-level identifier, but also arbitrary expressions, so long as the expression is closed, meaning that all variable names are either bound within the expression itself, or are top-level identifiers. That is, so long as the value of the expression could in principle be computed statically.

Given a static pointer, we can get back the value it points to using

deRefStaticPtr :: StaticPtr a -> a

Notice that we could as well have used a simple string to refer to fact in the above program, construct a string table, so that if the program were distributed we could have each process communicate strings in lieu of functions to commuicate tasks to run remotely, using the string table to map strings back to functions. Something like this:

module Main where

import GHC.StaticPtr
import Data.Dynamic

fact :: Int -> Int
fact 0 = 1
fact n = n * fact (n - 1)

computation1 :: IO ()
computation1 = print $ fact 10

stringTable =
  [ ("fact", toDynamic fact)
  , ("computation1", toDynamic computation1)
  ]

main = do
  send "some-node" "computation1"

where one could imagine node “some-node” running something like

serverLoop :: IO ()
serverLoop = forever $ do
  sptr <- expect
  fromDynamic (stringTable !! sptr)

assuming we have a send function for sending serializable values as messages to nodes and a expect function to receive them available.

Values in the string table are wrapped into Dynamic to make them all have uniform type (that way a simple homegeneous list can do just fine as a datastructure). But there are three problems with this approach:

  1. Constructing the string table is error prone: we might accidentally map the string "fact" to an entirely different function.

  2. No type safety. fromDynamic performs a type cast. This cast might fail if the type of value in the string table doesn’t match the expected type, making the program partial.

  3. It is antimodular: each module needs its own string table, which we then need to combine into a global string table for the whole program. If we add a any new module anywhere in the program, we need to also modify the construction of the string table, or accidentally forget to do so, which would constitute a bug.

(Some of these properties can be obtained with some clever Template Haskell hackery, but that solution is still fundamentally anti-modular, as well as contrived to use.)

It is for these three reasons that the StaticPointers language extension comes in handy. There is no need for manually constructing tables. Constructing and dereferencing static pointers is type safe because the type of a static pointer is related to the type of the value that it points to. Separate modules are not a problem, because the compiler takes care of collecting the set of all static pointers in a program into its own internal table that it embeds in the binary.

Pointer serialization

This all sounds rather nice, but the static pointer type is kept abstract, as it should to ensure safety, so how can we serialize a static pointer to send over the wire, and deserialize it on the remote end to reconstruct the static pointer? The GHC.StaticPtr module exports a few primitives to deal with just that. The idea is that each static pointer in a program is assigned a unique key (a StaticKey). We can obtain the key for a static pointer using

type StaticKey = Fingerprint

-- Defined in GHC.Fingerprint.
data Fingerprint = Fingerprint {-# UNPACK #-} !Word64 {-# UNPACK #-} !Word64
  deriving (Generic, Typeable)

staticKey :: StaticPtr a -> StaticKey

The type of keys is concrete (a key is a 128-bit hash), so keys can easily be encoded and decoded on the wire, using the Binary type class provided by the binary package:

-- Automatically derived instance, using `DeriveGeneric`.
instance Binary Fingerprint

Provided a key, we can map it to a StaticPtr using

unsafeLookupStaticPtr :: StaticKey -> Maybe (StaticPtr a)

Hold on a minute! This type is telling us that using unsafeLookupStaticPtr we can map the key to a static pointer of any type, which we can then deRefStaticPtr to a value of arbitrary type… Have we just lost type safety? In GHC 7.10, yes we have! In GHC 7.12, we will have a much safer lookup function:

lookupStaticPtr :: StaticKey
                -> (forall a. Typeable a => StaticPtr a -> b)
                -> Maybe b

(observe that this is a rank-2 type,) or equivalently

data DynStaticPtr = forall a. Typeable a => DynStaticPtr (StaticPtr a)

lookupStaticPtr :: StaticKey -> Maybe DynStaticPtr

This type says, provided a key and a continuation, lookupStaticPtr will resolve the key to a static pointer and if successful feed it to the continuation. The type of the static key is not known a priori, but we can query the type inside the continuation using the supplied Typeable constraint. The reason only the unsafe variant will ship in GHC 7.10 is because the safe variant will require a change to the Data.Typeable API to be truly safe (see here for details), and because we do not yet store Typeable constraints in the internal compiler-generated table mentioned above. In the meantime, this shouldn’t be a problem in practice: higher level libraries like Cloud Haskell and HdPH hide all uses of lookupStaticPtr behind an API that does guarantee type safety - it’s just that we have to trust that their implementations always call lookupStaticPtr at the right type, when ideally we wouldn’t need to entrust type safety to any library code at all, just the compiler.

Static closures

Static pointers turn out to be suprisingly powerful. As it stands, the language extension nominally only allows sharing references to static values across the wire. But it’s easy to build a lot more power on top. In particular, it would be nice if programs could transmit not just static values over the wire, but indeed (nearly) any odd closure. Consider the following main function:

main = do
  putStrLn "Hi! Give me a number..."
  x <- read <$> getLine
  send "some-node" $ closure (static fact) `closureAp` closurePure 10

The idea (first found in the “Towards Haskell in the Cloud” paper) is to introduce a datatype of closures (which we’ll define concretely later), along with three combinators to create Closures from StaticPtrs and from other Closures:

data Closure a

closure :: StaticPtr a -> Closure a
closurePure :: Serializable a => a -> Closure a
closureAp :: Closure (a -> b) -> Closure a -> Closure b

Notice that this datatype is nearly, but not quite, an applicative functor. We can only lift “serializable” values to a closure, not just any value. Given two existing Closures, we can create a new Closure by “applying” one to another. Morally, we are making it possible to pass around not just static pointers to top-level values or purely static expressions, but things that represent (partially) applied static pointers. Closures are not always static: their value may depend on values known only at runtime, as in the example above.

Come to think of it, a Closure very much acts like the closures that one would find deep in the bowels of GHC for representing partially applied functions during program execution. A closure is morally a code pointer paired with an environment, i.e. a list of actual arguments. Closures accumulate arguments as they are applied. In our case, the StaticPtr represents a code pointer, and the environment grows everytime we closureAp a Closure to something else.

We’ll turn to how Closure is defined in a minute, but first let’s talk about what it really means to be “serializable”:

data Dict c = c => Dict

class (Binary a, Typeable a) => Serializable a where
  serializableDict :: StaticPtr (Dict (Serializable a))

This class definition says that if a value can be encoded/decoded to a ByteString (see the binary package), and it can be queried for a representation of its type at runtime, then the value is serializable. However, serializable values also need to make it possible to obtain concrete “evidence” that the value really is serializable, in the form of a static dictionary. The idea is a neat trick. For all serializable values, we want to be able to obtain a static pointer to the evidence (or “dictionary”) associated with a class constraint. Because if we do, then we can “send” class dictionaries across the wire (or at least references to them)! But we can only take the static pointer of a value, so how does one make dictionary a first class value? The trick is to define a proxy datatype of dictionaries, using the ConstraintKinds extension (the Dict datatype). Any Dict value is a value like any other, but it embeds a constraint in it, which at runtime corresponds to a dictionary.

For example, any concrete value of Dict (Eq Int) carries a dictionary that can be seen as providing evidence that values of Int type can indeed be compared for equality. For any type a, Dict (Serializable a) carries evidence that values of type a are serializable. Any instance of Serializable makes it possible to query for this evidence - for example:

instance Serializable Int where
  serializableDict = static Dict

Now we can turn to the definition of Closure and its combinators:

data Closure a where
  StaticPtr :: StaticPtr b -> Closure b
  Encoded :: ByteString -> Closure ByteString
  Ap :: Closure (b -> c) -> Closure b -> Closure c
  deriving (Typeable)

closure :: StaticPtr a -> Closure a
closure = StaticPtr

closureAp :: Closure (a -> b) -> Closure a -> Closure b
closureAp = Ap

closurePure :: Serializable a => a -> Closure a
closurePure x =
    StaticPtr (static decodeD) `closureAp`
    closure serializableDict `closureAp`
    Encoded (encode x)
  where
    decodeD :: Dict (Serializable a) -> ByteString -> a
    decodeD Dict = decode

(There are many ways to define Closure, but this definition is perhaps most intuitive.)

As we can see from the definition, a Closure is not only a (quasi) applicative functor, but in fact a (quasi) free applicative functor. Using the Ap constructor, we can chain closures into long sequences (i.e. build environments). Using StaticPtr and Encoded, we can further make any serializable value a Closure of the following shape:

Ap (Ap (StaticPtr sptr_decodeD) csdict) bs

where sptr_decodeD is the static pointer to decodeD, csdict is a static serialization dictionary, and bs is a value encoded as a byte string.

Notice that any concrete Closure type is itself serializable:

instance Binary (Closure a) where
  put (Ap (Ap (StaticPtr sptr) dict) (Encoded bs)) =
      putWord8 0 >> put sptr >> put dict >> put bs
  put (StaticPtr sptr) = putWord8 1 >> put sptr
  put (Ap cf cx) = putWord8 2 >> put cf >> put cx

  get = do
    hdr <- getWord8
    case hdr of
      0 -> do sptr <- get
              dict <- get
              bs <- get
              return $ Ap (Ap (StaticPtr sptr) dict) (Encoded bs)
      1 -> StaticPtr <$> get
      2 -> Ap <$> get <*> get

instance Serializable (Closure Int)
  serializableDict = static Dict

(Note that for most types, manually defined Binary instances as above are unnecessary - any datatype with a Generic instance can have its Binary instance derived automatically).

Therefore, suprisingly, adding just static pointers as a primitive datatype in the compiler is all that’s necessary to be able to conveniently send even nearly arbitrary closures down the wire. It turns out that we don’t need to add full blown support for serializing arbitrary closures as an extra primitive to the compiler. That can all be done in user space, and with better control by the user on exactly how. The only limitation is that in effect the environment part of the closure needs to be serializable, but that’s a feature: it means that we can statically rule out accidentally serializing closures that capture gnarly things that we don’t want to serialize down the wire: think file handles, locks, sockets and other system resources, none of which the remote end would be able to make any sense of.

Conclusion

Static pointers are a lightweight extension to GHC, with direct applications to distributed programming, or in general, any form of pointer sharing across processes with distinct address spaces. As first observed in a seminal paper about distributed programming in Haskell, this extension adds just enough power to the GHC compiler and runtime to conveniently and safely send arbitrary serializable closures across the wire.

Distributed programming in Haskell is a reality today: there are several frameworks, most prominently Cloud Haskell, with several industrial applications. But the StaticPointers extension is brand new, and in fact no compiler release including it has shipped yet! Framework and application support for it is still lagging behind, but you can help. In particular, adding support to distributed-static and distributed-process would be a great step forward in the usability Cloud Haskell. Other next steps include: adding support for interoperating multiple versions of a program in a cluster, fully implementing lookupStaticPtr (see above), or improving the robustness and speed of message serialization (see for example these great results for an idea of what’s possible here). Those are just some ideas. If you’re interested in participating, the [GHC wiki][ghc-wiki-dH] contains quite a few pointers, and the cloud-haskell-developers@ and distributed-haskell@ mailing lists are good places to coordinate efforts. See you there!