Table of Contents
As we write this book, the landscape of CPU architecture is changing more rapidly than it has in decades.
A concurrent program needs to perform several possibly unrelated tasks at the same time. Consider the example of a game server: it is typically composed of dozens of components, each of which has complicated interactions with the outside world. One component might handle multi-user chat; several more will process the inputs of players, and feed state updates back to them; while another performs physics calculations.
The correct operation of a concurrent program does not require multiple cores, though they may improve performance and responsiveness.
In contrast, a parallel program solves a single problem. Consider a financial model that attempts to predict the next minute of fluctuations in the price of a single stock. If we want to apply this model to every stock listed on an exchange, for example to estimate which ones we should buy and sell, we hope to get an answer more quickly if we run the model on five hundred cores than if we use just one. As this suggests, a parallel program does not usually depend on the presence of multiple cores to work correctly.
Another useful distinction between concurrent and parallel programs lies in their interaction with the outside world. By definition, a concurrent program deals continuously with networking protocols, databases, and the like. A typical parallel program is likely to be more focused: it streams data in, crunches it for a while (with little further I/O), then streams data back out.
Many traditional languages further blur the already indistinct boundary between concurrent and parallel programming, because they force programmers to use the same primitives to construct both kinds of program.
In this chapter, we will concern ourselves with concurrent and parallel programs that operate within the boundaries of a single operating system process.
As a building block for concurrent programs, most programming languages provide a way of creating multiple independent threads of control. Haskell is no exception, though programming with threads in Haskell looks somewhat different than in other languages.
In Haskell, a thread is an IO action that
executes independently from other threads. To create a thread,
we import the Control.Concurrent
module and use the
forkIO
function.
ghci>
:m +Control.Concurrent
ghci>
:t forkIO
forkIO :: IO () -> IO ThreadIdghci>
:m +System.Directory
ghci>
forkIO (writeFile "xyzzy" "seo craic nua!") >> doesFileExist "xyzzy"
False
The new thread starts to execute almost immediately, and the thread that created it continues to execute concurrently. The thread will stop executing when it reaches the end of its IO action.
The runtime component of GHC does not specify an order
in which it executes threads. As a result, in our example
above, the file xyzzy
created by the new
thread may or may not have been created
by the time the original thread checks for its existence. If
we try this example once, then remove
xyzzy
and try again, we may get a
different result the second time.
Suppose we have a large file to compress and write to
disk, but we want to handle a user's input quickly enough that
they will perceive our program as responding immediately. If
we use forkIO
to write the file out in a
separate thread, we can do both simultaneously.
-- file: ch24/Compressor.hs import Control.Concurrent (forkIO) import Control.Exception (handle) import Control.Monad (forever) import qualified Data.ByteString.Lazy as L import System.Console.Readline (readline) -- Provided by the 'zlib' package on http://hackage.haskell.org/ import Codec.Compression.GZip (compress) main = do maybeLine <- readline "Enter a file to compress> " case maybeLine of Nothing -> return () -- user entered EOF Just "" -> return () -- treat no name as "want to quit" Just name -> do handle print $ do content <- L.readFile name forkIO (compressFile name content) return () main where compressFile path = L.writeFile (path ++ ".gz") . compress
Because we're using lazy ByteString I/O here, all we really do in the main thread is open the file. The actual reading occurs on demand in the other thread.
The use of handle print
gives us a cheap
way to print an error message if the user enters the name of a
file that does not exist.
The simplest way to share information between two threads is
to let them both use a variable. In our file compression
example, the main
thread shares both the
name of a file and its contents with the other thread. Because
Haskell data is immutable by default, this poses no risks:
neither thread can modify the other's view of the file's name or
contents.
We often need to have threads actively communicate with each other. For example, GHC does not provide a way for one thread to find out whether another is still executing, has completed, or has crashed[54]. However, it provides a synchronizing variable type, the MVar, which we can use to create this capability for ourselves.
An MVar acts like a single-element box: it can be either full or empty. We can put something into the box, making it full, or take something out, making it empty.
ghci>
:t putMVar
putMVar :: MVar a -> a -> IO ()ghci>
:t takeMVar
takeMVar :: MVar a -> IO a
If we try to put a value into an MVar that is already full, our thread is put to sleep until another thread takes the value out. Similarly, if we try to take a value from an empty MVar, our thread is put to sleep until some other thread puts a value in.
-- file: ch24/MVarExample.hs import Control.Concurrent communicate = do m <- newEmptyMVar forkIO $ do v <- takeMVar m putStrLn ("received " ++ show v) putStrLn "sending" putMVar m "wake up!"
The newEmptyMVar
function has a
descriptive name. To create an MVar that starts
out non-empty, we'd use newMVar
.
ghci>
:t newEmptyMVar
newEmptyMVar :: IO (MVar a)ghci>
:t newMVar
newMVar :: a -> IO (MVar a)
Let's run our example in ghci.
ghci>
:load MVarExample
[1 of 1] Compiling Main ( MVarExample.hs, interpreted ) Ok, modules loaded: Main.ghci>
communicate
sending rece
If you're coming from a background of concurrent programming in a traditional language, you can think of an MVar as being useful for two familiar purposes.
GHC's runtime system treats the program's original thread of control differently from other threads. When this thread finishes executing, the runtime system considers the program as a whole to have completed. If any other threads are executing at the time, they are terminated.
As a result, when we have long-running threads that must not be killed, we must make special arrangements to ensure that the main thread doesn't complete until the others do. Let's develop a small library that makes this easy to do.
-- file: ch24/NiceFork.hs import Control.Concurrent import Control.Exception (Exception, try) import qualified Data.Map as M data ThreadStatus = Running | Finished -- terminated normally | Threw Exception -- killed by uncaught exception deriving (Eq, Show) -- | Create a new thread manager. newManager :: IO ThreadManager -- | Create a new managed thread. forkManaged :: ThreadManager -> IO () -> IO ThreadId -- | Immediately return the status of a managed thread. getStatus :: ThreadManager -> ThreadId -> IO (Maybe ThreadStatus) -- | Block until a specific managed thread terminates. waitFor :: ThreadManager -> ThreadId -> IO (Maybe ThreadStatus) -- | Block until all managed threads terminate. waitAll :: ThreadManager -> IO ()
We keep our ThreadManager type abstract using
the usual recipe: we wrap it in a newtype
, and prevent clients
from creating values of this type. Among our module's exports,
we list the type constructor and the IO action that
constructs a manager, but we do not export the data
constructor.
-- file: ch24/NiceFork.hs module NiceFork ( ThreadManager , newManager , forkManaged , getStatus , waitFor , waitAll ) where
For the implementation of ThreadManager, we maintain a map from thread ID to thread state. We'll refer to this as the thread map.
-- file: ch24/NiceFork.hs newtype ThreadManager = Mgr (MVar (M.Map ThreadId (MVar ThreadStatus))) deriving (Eq) newManager = Mgr `fmap` newMVar M.empty
We have two levels of MVar use here. We keep the Map in an MVar. This lets us “modify” the map by replacing it with a new version. We also ensure that any thread that uses the Map will see a consistent view of it.
For each thread that we manage, we maintain an MVar. A per-thread MVar starts off empty, which indicates that the thread is executing. When the thread finishes or is killed by an uncaught exception, we put this information into the MVar.
To create a thread and watch its status, we must perform a little bit of book-keeping.
-- file: ch24/NiceFork.hs forkManaged (Mgr mgr) body = modifyMVar mgr $ \m -> do state <- newEmptyMVar tid <- forkIO $ do result <- try body putMVar state (either Threw (const Finished) result) return (M.insert tid state m, tid)
The modifyMVar
function that we used
in forkManaged
above is very useful: it's
a safe combination of takeMVar
and
putMVar
.
ghci>
:t modifyMVar
ved "wake up!" modifyMVar :: MVar a -> (a -> IO (a, b)) -> IO b
It takes the value from an MVar, and passes
it to a function. This function can both generate a new value
and return a result. If the function throws an exception,
modifyMVar
puts the original value back
into the MVar, otherwise it puts the new value
in. It returns the other element of the function as its own
result.
When we use modifyMVar
instead of
manually managing an MVar with
takeMVar
and
putMVar
, we avoid two common kinds of
concurrency bug.
Forgetting to put a value back into an MVar. This can result in deadlock, in which some thread waits forever on an MVar that will never have a value put into it.
Failure to account for the possibility that an
exception might be thrown, disrupting the flow of a piece
of code. This can result in a call to
putMVar
that
should occur not actually happening,
again leading to deadlock.
Because of these nice safety properties, it's wise to use
modifyMVar
whenever possible.
We can the take the pattern that
modifyMVar
follows, and apply it to many
other resource management situations. Here are the steps of
the pattern.
Safety aside, this approach has another benefit: it can
make our code shorter and easier to follow. As we can see
from looking at forkManaged
above,
Haskell's lightweight syntax for anonymous functions makes
this style of coding visually unobtrusive.
Here's the definition of modifyMVar
,
so that you can see a specific form of this pattern.
-- file: ch24/ModifyMVar.hs import Control.Concurrent (MVar, putMVar, takeMVar) import Control.Exception (block, catch, throw, unblock) import Prelude hiding (catch) -- use Control.Exception's version modifyMVar :: MVar a -> (a -> IO (a,b)) -> IO b modifyMVar m io = block $ do a <- takeMVar m (b,r) <- unblock (io a) `catch` \e -> putMVar m a >> throw e putMVar m b return r
You should easily be able to adapt this to your particular needs, whether you're working with network connections, database handles, or data managed by a C library.
Our getStatus
function tells us the
current state of a thread. If the thread is no longer managed
(or was never managed in the first place), it returns
Nothing
.
-- file: ch24/NiceFork.hs getStatus (Mgr mgr) tid = modifyMVar mgr $ \m -> case M.lookup tid m of Nothing -> return (m, Nothing) Just st -> tryTakeMVar st >>= \mst -> case mst of Nothing -> return (m, Just Running) Just sth -> return (M.delete tid m, Just sth)
If the thread is still running, it returns Just
Running
. Otherwise, it indicates why the thread
terminated, and stops managing the
thread.
If the tryTakeMVar
function finds
that the MVar is empty, it returns
Nothing
immediately instead of blocking.
ghci>
:t tryTakeMVar
tryTakeMVar :: MVar a -> IO (Maybe a)
Otherwise, it extracts the value from the MVar as usual.
The waitFor
function behaves
similarly, but instead of returning immediately, it blocks
until the given thread terminates before returning.
-- file: ch24/NiceFork.hs waitFor (Mgr mgr) tid = do maybeDone <- modifyMVar mgr $ \m -> return $ case M.updateLookupWithKey (\_ _ -> Nothing) tid m of (Nothing, _) -> (m, Nothing) (done, m') -> (m', done) case maybeDone of Nothing -> return Nothing Just st -> Just `fmap` takeMVar st
It first extracts the MVar that holds the
thread's state, if it exists. The Map type's
updateLookupWithKey
function is useful:
it combines looking up a key with modifying or removing the
value.
ghci>
:m +Data.Map
ghci>
:t updateLookupWithKey
updateLookupWithKey :: (Ord k) => (k -> a -> Maybe a) -> k -> Map k a -> (Maybe a, Map k a)
In this case, we want to always remove the MVar holding the thread's state if it is present, so that our thread manager will no longer be managing the thread. If there was a value to extract, we take the thread's exit status from the MVar and return it.
Our final useful function simply waits for all currently managed threads to complete, and ignores their exit statuses.
-- file: ch24/NiceFork.hs waitAll (Mgr mgr) = modifyMVar mgr elems >>= mapM_ takeMVar where elems m = return (M.empty, M.elems m)
Our definition of waitFor
above is a
little unsatisfactory, because we're performing more or less
the same case analysis in two places: inside the function
called by modifyMVar
, and again on its
return value.
Sure enough, we can apply a function that we came across
earlier to eliminate this duplication. The function in
question is join
, from the
Control.Monad
module.
ghci>
:m +Control.Monad
ghci>
:t join
join :: (Monad m) => m (m a) -> m a
The trick here is to see that we can get rid of the second
case
expression by having the first one return the
IO action that we should perform once we return
from modifyMVar
. We'll use
join
to execute the action.
-- file: ch24/NiceFork.hs waitFor2 (Mgr mgr) tid = join . modifyMVar mgr $ \m -> return $ case M.updateLookupWithKey (\_ _ -> Nothing) tid m of (Nothing, _) -> (m, return Nothing) (Just st, m') -> (m', Just `fmap` takeMVar st)
This is an interesting idea: we can create a monadic function or action in pure code, then pass it around until we end up in a monad where we can use it. This can be a nimble way to write code, once we develop an eye for when it makes sense.
For one-shot communications between threads, an MVar is perfectly good. Another type, Chan, provides a one-way communication channel. Here is a simple example of its use.
-- file: ch24/Chan.hs import Control.Concurrent import Control.Concurrent.Chan chanExample = do ch <- newChan forkIO $ do writeChan ch "hello world" writeChan ch "now i quit" readChan ch >>= print readChan ch >>= print
If a Chan is empty,
readChan
blocks until there is a value to
read. The writeChan
function never blocks:
it writes a new value into a Chan
immediately.
Like most Haskell container types, both MVar and Chan are non-strict: neither evaluates its contents. We mention this not because it's a problem, but because it's a common blind spot: people tend to assume that these types are strict, perhaps because they're used in the IO monad.
As for other container types, the upshot of a mistaken guess about the strictness of an MVar or Chan type is often a space or performance leak. Here's a plausible scenario to consider.
We fork off a thread to perform some expensive computation on another core.
-- file: ch24/Expensive.hs import Control.Concurrent notQuiteRight = do mv <- newEmptyMVar forkIO $ expensiveComputation_stricter mv someOtherActivity result <- takeMVar mv print result
It seems to do something, and puts its result back into the MVar.
-- file: ch24/Expensive.hs expensiveComputation mv = do let a = "this is " b = "not really " c = "all that expensive" putMVar mv (a ++ b ++ c)
When we take the result from the MVar in the parent thread and attempt to do something with it, our thread starts computing furiously, because we never forced the computation to actually occur in the other thread!
As usual, the solution is straightforward, once we know there's a potential for a problem: we add strictness to the forked thread, to ensure that the computation occurs there. This strictness is best added in one place, to avoid the possibility that we might forget to add it.
-- file: ch24/ModifyMVarStrict.hs {-# LANGUAGE BangPatterns #-} import Control.Concurrent (MVar, putMVar, takeMVar) import Control.Exception (block, catch, throw, unblock) import Prelude hiding (catch) -- use Control.Exception's version modifyMVar_strict :: MVar a -> (a -> IO a) -> IO () modifyMVar_strict m io = block $ do a <- takeMVar m !b <- unblock (io a) `catch` \e -> putMVar m a >> throw e putMVar m b
It's always worth checking Hackage | |
---|---|
In the Hackage package database, you will find a
library, |
The !
pattern above is simple to use, but it
is not always sufficient to ensure that our data is
evaluated. For a more complete approach, see the section called “Separating algorithm from evaluation” below.
Because writeChan
always succeeds
immediately, there is a potential risk to using a
Chan. If one thread writes to a
Chan more often than another thread reads from
it, the Chan will grow in an unchecked manner:
unread messages will pile up as the reader falls further and
further behind.
Although Haskell has different primitives for sharing data between threads than other languages, it still suffers from the same fundamental problem: writing correct concurrent programs is fiendishly difficult. Indeed, several pitfalls of concurrent programming in other languages apply equally to Haskell. Two of the better known problems are deadlock and starvation.
In a deadlock situation, two or more threads get stuck forever in a clash over access to shared resources. One classic way to make a multithreaded program deadlock is to forget the order in which we must acquire locks. This kind of bug is so common, it has a name: lock order inversion. While Haskell doesn't provide locks, the MVar type is prone to the order inversion problem. Here's a simple example.
-- file: ch24/LockHierarchy.hs import Control.Concurrent nestedModification outer inner = do modifyMVar_ outer $ \x -> do yield -- force this thread to temporarily yield the CPU modifyMVar_ inner $ \y -> return (y + 1) return (x + 1) putStrLn "done" main = do a <- newMVar 1 b <- newMVar 2 forkIO $ nestedModification a b forkIO $ nestedModification b a
If we run this in ghci, it will usually—but not always—print nothing, indicating that both threads have gotten stuck.
The problem with the
nestedModification
function is easy to
spot. In the first thread, we take the MVar
a
, then b
. In the
second, we take b
, then
a
. If the first thread succeeds in taking
a
and the second takes
b
, both threads will block: each tries to
take an MVar that the other has already emptied,
so neither can make progress.
Across languages, the usual way to solve an order inversion problem is to always follow a consistent order when acquiring resources. Since this approach requires manual adherence to a coding convention, it is easy to miss in practice.
To make matters more complicated, these kinds of inversion problems can be difficult to spot in real code. The taking of MVars is often spread across several functions in different files, making visual inspection more tricky. Worse, these problems are often intermittent, which makes them tough to even reproduce, never mind isolate and fix.
Concurrent software is also prone to
starvation, in which one thread
“hogs” a shared resource, preventing another from
using it. It's easy to imagine how this might occur: one
thread calls modifyMVar
with a body that
executes for 100 milliseconds, while another calls
modifyMVar
on the same MVar
with a body that executes for 1 millisecond. The second
thread cannot make progress until the first puts a value back
into the MVar.
The non-strict nature of the MVar type can either exacerbate or cause a starvation problem. If we put a thunk into an MVar that will be expensive to evaluate, and take it out of the MVar in a thread that otherwise looks like it ought to be cheap, that thread could suddenly become computationally expensive if it has to evaluate the thunk. This makes the advice we gave in the section called “MVar and Chan are non-strict” particularly relevant.
Fortunately, the APIs for concurrency that we have covered here are by no means the end of the story. A more recent addition to Haskell, Software Transactional Memory, is both easier and safer to work with. We will discuss it in chapter Chapter 28, Software transactional memory.
By default, GHC generates programs that use just one core, even when we write explicitly concurrent code. To use multiple cores, we must explicitly choose to do so. We make this choice at link time, when we are generating an executable program.
The “non-threaded” runtime library runs all Haskell threads in a single operating system thread. This runtime is highly efficient for creating threads and passing data around in MVars.
The “threaded” runtime library uses multiple operating system threads to run Haskell threads. It has somewhat more overhead for creating threads and using MVars.
If we pass the -threaded
option to the
compiler, it will link our program against the threaded runtime
library. We do not need to use -threaded
when
we are compiling libraries or source files, only when we are
finally generating an executable.
Even when we select the threaded runtime for our program, it will still default to using only one core when we run it. We must explicitly tell the runtime how many cores to use.
We can pass options to GHC's runtime system on the
command line of our program. Before handing control to our
code, the runtime scans the program's arguments for the
special command line option +RTS
. It
interprets everything that follows, until the special option
-RTS
, as an option for the runtime system,
not our program. It hides all of these options from our code.
When we use the System.Environment
module's
getArgs
function to obtain our command
line arguments, we will not find any runtime options in the
list.
The threaded runtime accepts an option
-N
[55]. This takes one argument, which specifies the
number of cores that GHC's runtime system should use. The
option parser is picky: there must be no spaces between
-N
and the number that follows it. The option
-N4
is acceptable, but -N 4
is not.
The module GHC.Conc
exports a variable,
numCapabilities
, that tells us how many
cores the runtime system has been given with the
-N
RTS option.
-- file: ch24/NumCapabilities.hs import GHC.Conc (numCapabilities) import System.Environment (getArgs) main = do args <- getArgs putStrLn $ "command line arguments: " ++ show args putStrLn $ "number of cores: " ++ show numCapabilities
If we compile and run the above program, we can see that the options to the runtime system are not visible to the program, but that it can see how many cores it can run on.
$
ghc -c NumCapabilities.hs
$
ghc -threaded -o NumCapabilities NumCapabilities.o
$
./NumCapabilities +RTS -N4 -RTS foo
command line arguments: ["foo"] number of cores: 4
The decision of which runtime to use is not completely clear cut. While the threaded runtime can use multiple cores, it has a cost: threads and sharing data between them are more expensive than with the non-threaded runtime.
Furthermore, the garbage collector used by GHC as of version 6.8.3 is single threaded: it pauses all other threads while it runs, and executes on one core. This limits the performance improvement we can hope to see from using multiple cores[56].
In many real world concurrent programs, an individual thread will spend most of its time waiting for a network request or response. In these cases, if a single Haskell program serves tens of thousands of concurrent clients, the lower overhead of the non-threaded runtime may be helpful. For example, instead of having a single server program use the threaded runtime on four cores, we might see better performance if we design our server so that we can run four copies of it simultaneously, and use the non-threaded runtime.
Our purpose here is not to dissuade you from using the threaded runtime. It is not much more expensive than the non-threaded runtime: threads remain amazingly cheap compared to the runtimes of most other programming languages. We merely want to make it clear that switching to the threaded runtime will not necessarily result in an automatic win.
We will now switch our focus to parallel programming. For many computationally expensive problems, we could calculate a result more quickly if we could divide up the solution, and evaluate it on many cores at once. Computers with multiple cores are already ubiquitous, but few programs can take advantage of the computing power of even a modern laptop.
In large part, this is because parallel programming is traditionally seen as very difficult. In a typical programming language, we would use the same libraries and constructs that we apply to concurrent programs to develop a parallel program. This forces us to contend with the familiar problems of deadlocks, race conditions, starvation, and sheer complexity.
While we could certainly use Haskell's concurrency features to develop parallel code, there is a much simpler approach available to us. We can take a normal Haskell function, apply a few simple transformations to it, and have it evaluated in parallel.
The familiar seq
function evaluates
an expression to what we call head normal
form (abbreviated HNF). It stops once it reaches
the outermost constructor (the “head”). This is
distinct from normal form (NF), in which
an expression is completely evaluated.
You will also hear Haskell programmers refer to weak head normal form (WHNF). For normal data, weak head normal form is the same as head normal form. The difference only arises for functions, and is too abstruse to concern us here.
Here is a normal Haskell function that sorts a list using a divide-and-conquer approach.
-- file: ch24/Sorting.hs sort :: (Ord a) => [a] -> [a] sort (x:xs) = lesser ++ x:greater where lesser = sort [y | y <- xs, y < x] greater = sort [y | y <- xs, y >= x] sort _ = []
This function is inspired by the well-known Quicksort algorithm, and it is a classic among Haskell programmers: it is often presented as a one-liner early in a Haskell tutorial, to tease the reader with an example of Haskell's expressiveness. Here, we've split the code over a few lines, to make it easier to compare the serial and parallel versions.
Here is a very brief description of how
sort
operates.
It chooses an element from the list. This is called the pivot. Any element would do as the pivot; the first is merely the easiest to pattern match on.
It creates a sublist of all elements less than the pivot, and recursively sorts them.
It creates a sublist of all elements greater than or equal to the pivot, and recursively sorts them.
The parallel version of the function is only a little more complicated than the initial version.
-- file: ch24/Sorting.hs module Sorting where import Control.Parallel (par, pseq) parSort :: (Ord a) => [a] -> [a] parSort (x:xs) = force greater `par` (force lesser `pseq` (lesser ++ x:greater)) where lesser = parSort [y | y <- xs, y < x] greater = parSort [y | y <- xs, y >= x] parSort _ = []
We have barely perturbed the code: all we have added are
three functions, par
,
pseq
, and
force
.
The par
function is
provided by the Control.Parallel
module. It
serves a similar purpose to seq
: it
evaluates its left argument to weak head normal form, and
returns its right. As its name suggests,
par
can evaluate its left argument in
parallel with whatever other evaluations are occurring.
As for pseq
, it is similar
to seq
: it evaluates the expression on
the left to WHNF before returning the expression on the right.
The difference between the two is subtle, but important for
parallel programs: the compiler does not
promise to evaluate the left argument of
seq
if it can see that evaluating the
right argument first would improve performance. This
flexibility is fine for a program executing on one core, but
it is not strong enough for code running on multiple cores.
In contrast, the compiler guarantees that
pseq
will evaluate its left argument
before its right.
These changes to our code are remarkable for all the things we have not needed to say.
The key to getting decent performance out of parallel
Haskell code is to find meaningful chunks of work to perform
in parallel. Non-strict evaluation can get in the way of
this, which is why we use the force
function in our parallel sort. To best explain what the
force
function is for, we will first look
at a mistaken example.
-- file: ch24/Sorting.hs sillySort (x:xs) = greater `par` (lesser `pseq` (lesser ++ x:greater)) where lesser = sillySort [y | y <- xs, y < x] greater = sillySort [y | y <- xs, y >= x] sillySort _ = []
Take a look at the small changes in each use of
par
. Instead of force
lesser
and force greater
, here we
evaluate lesser
and greater
.
Remember that evaluation to WHNF only computes enough of
an expression to see its outermost
constructor. In this mistaken example, we evaluate each
sorted sublist to WHNF. Since the outermost constructor in
each case is just a single list constructor, we are in fact
only forcing the evaluation of the first element of each
sorted sublist! Every other element of each list remains
unevaluated. In other words, we do almost no useful work in
parallel: our sillySort
is nearly
completely sequential.
We avoid this with our force
function
by forcing the entire spine of a list to be evaluated before
we give back a constructor.
-- file: ch24/Sorting.hs force :: [a] -> () force xs = go xs `pseq` () where go (_:xs) = go xs go [] = 1
Notice that we don't care what's in the list;
we walk down its spine to the end, then use
pseq
once. There is clearly no magic
involved here: we are just using our usual understanding of
Haskell's evaluation model. And because we will be using
force
on the left hand side of
par
or pseq
, we
don't need to return a meaningful value.
Of course, in many cases we will need to force the evaluation of individual elements of the list, too. Below, we will discuess a typeclass-based solution to this problem.
The par
function does not actually
promise to evaluate an expression in parallel with another.
Instead, it undertakes to do so if it “makes
sense”. This wishy-washy non-promise is actually
more useful than a guarantee to always evaluate an expression
in parallel. It gives the runtime system the freedom to act
intelligently when it encounters a use of
par
.
For instance, the runtime could decide that an expression is too cheap to be worth evaluating in parallel. Or it might notice that all cores are currently busy, so that “sparking” a new parallel evaluation will lead to there being more runnable threads than there are cores available to execute them.
This lax specification in turn affects how we write parallel
code. Since par
may be somewhat
intelligent at runtime, we can use it almost wherever we like,
on the assumption that performance will not be bogged down by
threads contending for busy cores.
To try our code out, let's save sort
,
parSort
, and
parSort2
to a module named
Sorting.hs
. We create a small driver
program that we can use to time the performance of one of
those sorting function.
-- file: ch24/SortMain.hs module Main where import Data.Time.Clock (diffUTCTime, getCurrentTime) import System.Environment (getArgs) import System.Random (StdGen, getStdGen, randoms) import Sorting -- testFunction = sort -- testFunction = seqSort testFunction = parSort -- testFunction = parSort2 2 randomInts :: Int -> StdGen -> [Int] randomInts k g = let result = take k (randoms g) in force result `seq` result main = do args <- getArgs let count | null args = 500000 | otherwise = read (head args) input <- randomInts count `fmap` getStdGen putStrLn $ "We have " ++ show (length input) ++ " elements to sort." start <- getCurrentTime let sorted = testFunction input putStrLn $ "Sorted all " ++ show (length sorted) ++ " elements." end <- getCurrentTime putStrLn $ show (end `diffUTCTime` start) ++ " elapsed."
For simplicity, we choose the sorting function to
benchmark at compilation time, via the
testFunction
variable.
Our program accepts a single optional command line argument, the length of the random list to generate.
Non-strict evaluation can turn performance measurement and analysis into something of a minefield. Here are some potential problems that we specifically work to avoid in our driver program.
Measuring several things, when we think we
are looking at just one. Haskell's default
pseudorandom number generator (PRNG) is slow, and the
randoms
function generates random
numbers on demand.
Before we record our starting time, we force every element of the input list to be evaluated, and we print the length of the list: this ensures that we create all of the random numbers that we will need in advance.
If we were to omit this step, we would interleave the generation of random numbers with attempts to work with them in parallel. We would thus be measuring both the cost of sorting the numbers and, less obviously, the cost of generating them.
Invisible data dependencies. When we generate the list of random numbers, simply printing the length of the list would not perform enough evaluation. This wouls evaluate the spine of the list, but not its elements. The actual random numbers would not be evaluated until the sort compares them.
This can have serious consequences for performance. The value of a random number depends on the value of the preceding random number in the list, but we have scattered the list elements randomly among our processor cores. If we did not evaluate the list elements prior to sorting, we would suffer a terrible “ping pong” effect: not only would evaluation bounce from one core to another, performance would suffer.
Try snipping out the application of
force
from the body of
main
above: you should find that the
parallel code can easily end up three times
slower than the non-parallel
code.
Benchmarking a thunk, when we believe that
the code is performing meaningful work. To
force the sort to take place, we print the length of the
result list before we record the ending time. Without
putStrLn
demanding the length of the
list in order to print it, the sort would not occur at
all.
When we build the program, we enable optimization and GHC's threaded runtime.
$
ghc -threaded -O2 --make SortMain
[1 of 2] Compiling Sorting ( Sorting.hs, Sorting.o ) [2 of 2] Compiling Main ( SortMain.hs, SortMain.o ) Linking SortMain ...
When we run the program, we must tell GHC's
runtime how many cores to use. Initially, we try the original
sort
, to establish a performance
baseline.
$
./Sorting +RTS -N1 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 3.178941s elapsed.
Enabling a second core ought to have no effect on performance.
$
./Sorting +RTS -N2 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 3.259869s elapsed.
If we recompile and test the performance of
parSort
, the results are less than
stellar.
$
./Sorting +RTS -N1 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 3.915818s elapsed.$
./Sorting +RTS -N2 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 4.029781s elapsed.
We have gained nothing in performance. It seems that this
could be due to one of two factors: either
par
is intrinsically expensive, or we are
using it too much. To help us to distinguish between the two
possibilities, here is a sort is identical to
parSort
, but it uses
pseq
instead of
par
.
-- file: ch24/Sorting.hs seqSort :: (Ord a) => [a] -> [a] seqSort (x:xs) = lesser `pseq` (greater `pseq` (lesser ++ x:greater)) where lesser = seqSort [y | y <- xs, y < x] greater = seqSort [y | y <- xs, y >= x] seqSort _ = []
We also drop the use of force
, so
compared to our original sort
, we should
only be measuring the cost of using pseq
.
What effect does pseq
alone have on
performance?
$
./Sorting +RTS -N1 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 3.848295s elapsed.
This suggests that par
and
pseq
have similar costs. What can we do
to improve performance?
In our parSort
, we perform twice as
many applications of par
as there are
elements to sort. While par
is
cheap, as we have seen, it is not
free. When we recursively apply
parSort
, we eventually apply
par
to individual list elements. At this
fine granularity, the cost of using par
outweighs any possible usefulness. To reduce this effect, we
switch to our non-parallel sort
after
passing some threshold.
-- file: ch24/Sorting.hs parSort2 :: (Ord a) => Int -> [a] -> [a] parSort2 d list@(x:xs) | d <= 0 = sort list | otherwise = force greater `par` (force lesser `pseq` (lesser ++ x:greater)) where lesser = parSort2 d' [y | y <- xs, y < x] greater = parSort2 d' [y | y <- xs, y >= x] d' = d - 1 parSort2 _ _ = []
Here, we stop recursing and sparking new parallel evaluations at a controllable depth. If we knew the size of the data we were dealing with, we could stop subdividing and switch to the non-parallel code once we reached a sufficiently small amount of remaining work.
$
./Sorting +RTS -N2 -RTS 700000
We have 700000 elements to sort. Sorted all 700000 elements. 2.947872s elapsed.
On a dual core system, this gives us roughly a 25% speedup. This is not a huge number, but consider the number of changes we had to make in return for this performance improvement: just a few annotations.
This sorting function is particularly resistant to good
parallel performance. The amount of memory allocation it
performs forces the garbage collector to run frequently. We
can see the effect by running our program with the
-sstderr
RTS option, which prints garbage
collection statistics to the screen. This indicates that our
program spends roughly 40% of its time collecting garbage.
Since the garbage collector in GHC 6.8 stops all threads and
runs on a single core, it acts as a bottleneck.
You can expect more impressive performance improvements
from less allocation-heavy code when you use
par
annotations. We have seen some
simple numerical benchmarks run 1.8 times faster on a dual
core system than with a single core. As we write this book, a
parallel garbage collector is under development for GHC,
which should help considerably with the performance of
allocation-heavy code on multicore systems.
Within the programming community, one of the most famous software systems to credit functional programming for inspiration is Google's MapReduce infrastructure for parallel processing of bulk data.
We can easily construct a greatly simplified, but still useful, Haskell equivalent. To focus our attention, we will look at the processing of web server log files, which tend to be both huge and plentiful[57].
As an example, here is a log entry for a page visit recorded by the Apache web server. The entry originally filled one line; we have split it across several lines to fit.
201.49.94.87 - - [08/Jun/2008:07:04:20 -0500] "GET / HTTP/1.1" 200 2097 "http://en.wikipedia.org/wiki/Mercurial_(software)" "Mozilla/5.0 (Windows; U; Windows XP 5.1; en-GB; rv:1.8.1.12) Gecko/20080201 Firefox/2.0.0.12" 0 hgbook.red-bean.com
While we could create a straightforward implementation without much effort, we will resist the temptation to dive in. If we think about solving a class of problems instead of a single one, we may end up with more widely applicable code.
When we develop a parallel program, we are always faced with a few “bad penny” problems, which turn up no matter what the underlying programming language is.
Our algorithm quickly becomes obscured by the details of partitioning and communication. This makes it difficult to understand code, which in turn makes modifying it risky.
Choosing a “grain size”—the smallest unit of work parceled out to a core— can be difficult. If the grain size is too small, cores spend so much of their time on book-keeping that a parallel program can easily become slower than a serial counterpart. If the grain size is too large, some cores may lie idle due to poor load balancing.
In parallel Haskell code, the clutter that would arise
from communication code in a traditional language is replaced
with the clutter of par
and
pseq
annotations. As an example, this
function operates similarly to map
, but
evaluates each element to weak head normal form (WHNF) in
parallel as it goes.
-- file: ch24/ParMap.hs import Control.Parallel (par) parallelMap :: (a -> b) -> [a] -> [b] parallelMap f (x:xs) = let r = f x in r `par` r : parallelMap f xs parallelMap _ _ = []
The type b
might be a list,
or some other type for which evaluation to WHNF doesn't do a
useful amount of work. We'd prefer not to have to write a
special parallelMap
for lists, and for
every other type that needs special handling.
To address this problem, we will begin by considering a simpler problem: how to force a value to be evaluated. Here is a function that forces every element of a list to be evaluated to WHNF.
-- file: ch24/ParMap.hs forceList :: [a] -> () forceList (x:xs) = x `pseq` forceList xs forceList _ = ()
Our function performs no computation on the list. (In
fact, from examining its type signature, we can tell that it
cannot perform any computation, since it
knows nothing about the elements of the list.) Its only
purpose is to ensure that the spine of the list is evaluated
to head normal form. The only place that it makes any sense to
apply this function is in the first argument of
seq
or par
, for
example as follows.
-- file: ch24/ParMap.hs stricterMap :: (a -> b) -> [a] -> [b] stricterMap f xs = forceList xs `seq` map f xs
This still leaves us with the elements of the list evaluated only to WHNF. We address this by adding a function as parameter that can force an element to be evaluated more deeply.
-- file: ch24/ParMap.hs forceListAndElts :: (a -> ()) -> [a] -> () forceListAndElts forceElt (x:xs) = forceElt x `seq` forceListAndElts forceElt xs forceListAndElts _ _ = ()
The Control.Parallel.Strategies
module generalizes this idea into something we can use as a
library. It introduces the idea of an evaluation
strategy.
-- file: ch24/Strat.hs type Done = () type Strategy a = a -> Done
An evaluation strategy performs no computation; it simply
ensures that a value is evaluated to some extent. The
simplest strategy is named r0
, and does
nothing at all.
-- file: ch24/Strat.hs r0 :: Strategy a r0 _ = ()
Next is rwhnf
, which evaluates a
value to weak head normal form.
-- file: ch24/Strat.hs rwhnf :: Strategy a rwhnf x = x `seq` ()
To evaluate a value to normal form, the module provides a
typeclass with a method named rnf
.
-- file: ch24/Strat.hs class NFData a where rnf :: Strategy a rnf = rwhnf
For the basic types, such as Int, weak head
normal form and normal form are the same thing, which is why
the NFData typeclass uses
rwhnf
as the default implementation of
rnf
. For many common types, the
Control.Parallel.Strategies
module provides
instances of NFData.
-- file: ch24/Strat.hs instance NFData Char instance NFData Int instance NFData a => NFData (Maybe a) where rnf Nothing = () rnf (Just x) = rnf x {- ... and so on ... -}
From these examples, it should be clear how you might
write an NFData instance for a type of your own.
Your implementation of rnf
must handle
every constructor, and apply rnf
to every
field of a constructor.
From these strategy building blocks, we can construct more
elaborate strategies. Many are already provided by
Control.Parallel.Strategies
. For instance,
parList
applies an evaluation strategy in
parallel to every element of a list.
-- file: ch24/Strat.hs parList :: Strategy a -> Strategy [a] parList strat [] = () parList strat (x:xs) = strat x `par` (parList strat xs)
The module uses this to define a parallel
map
function.
-- file: ch24/Strat.hs parMap :: Strategy b -> (a -> b) -> [a] -> [b] parMap strat f xs = map f xs `using` parList strat
This is where the code becomes interesting. On the left
of using
, we have a normal application of
map
. On the right, we have an evaluation
strategy. The using
combinator tells us
how to apply a strategy to a value, allowing us to keep the
code separate from how we plan to evaluate it.
-- file: ch24/Strat.hs using :: a -> Strategy a -> a using x s = s x `seq` x
The Control.Parallel.Strategies
module
provides many other functions that provide fine control over
evaluation. For instance, parZipWith
that applies zipWith
in parallel, using
an evaluation strategy.
-- file: ch24/Strat.hs vectorSum' :: (NFData a, Num a) => [a] -> [a] -> [a] vectorSum' = parZipWith rnf (+)
We can quickly suggest a type for a
mapReduce
function by considering what it
must do. We need a map component, to
which we will give the usual type a -> b. And
we need a reduce; this term is a synonym
for fold. Rather than commit ourselves
to using a specific kind of fold, we'll use a more general
type, [b] -> c. This type lets us use a left
or right fold, so we can choose the one that suits our data
and processing needs.
If we plug these types together, the complete type looks like this.
-- file: ch24/MapReduce.hs simpleMapReduce :: (a -> b) -- map function -> ([b] -> c) -- reduce function -> [a] -- list to map over -> c
The code that goes with the type is extremely simple.
-- file: ch24/MapReduce.hs simpleMapReduce mapFunc reduceFunc = reduceFunc . map mapFunc
Our definition of simpleMapReduce
is
too simple to really be interesting. To make it useful, we
want to be able to specify that some of the work should occur
in parallel. We'll achieve this using strategies, passing in
a strategy for the map phase and one for the reduction phase.
-- file: ch24/MapReduce.hs mapReduce :: Strategy b -- evaluation strategy for mapping -> (a -> b) -- map function -> Strategy c -- evaluation strategy for reduction -> ([b] -> c) -- reduce function -> [a] -- list to map over -> c
Both the type and the body of the function must grow a little in size to accommodate the strategy parameters.
-- file: ch24/MapReduce.hs mapReduce mapStrat mapFunc reduceStrat reduceFunc input = mapResult `pseq` reduceResult where mapResult = parMap mapStrat mapFunc input reduceResult = reduceFunc mapResult `using` reduceStrat
To achieve decent performance, we must ensure that the
work that we do per application of par
substantially outweighs its book-keeping costs. If we are
processing a huge file, splitting it on line boundaries gives
us far too little work compared to overhead.
We will develop a way to process a file in larger chunks in a later section. What should those chunks consist of? Because a web server log file ought to contain only ASCII text, we will see excellent performance with a lazy ByteString: this type is highly efficient, and consumes little memory when we stream it from a file.
-- file: ch24/LineChunks.hs module LineChunks ( chunkedReadWith ) where import Control.Exception (bracket, finally) import Control.Monad (forM, liftM) import Control.Parallel.Strategies (NFData, rnf) import Data.Int (Int64) import qualified Data.ByteString.Lazy.Char8 as LB import GHC.Conc (numCapabilities) import System.IO data ChunkSpec = CS { chunkOffset :: !Int64 , chunkLength :: !Int64 } deriving (Eq, Show) withChunks :: (NFData a) => (FilePath -> IO [ChunkSpec]) -> ([LB.ByteString] -> a) -> FilePath -> IO a withChunks chunkFunc process path = do (chunks, handles) <- chunkedRead chunkFunc path let r = process chunks (rnf r `seq` return r) `finally` mapM_ hClose handles chunkedReadWith :: (NFData a) => ([LB.ByteString] -> a) -> FilePath -> IO a chunkedReadWith func path = withChunks (lineChunks (numCapabilities * 4)) func path
We consume each chunk in parallel, taking careful advantage of lazy I/O to ensure that we can stream these chunks safely.
Lazy I/O poses a few well known hazards that we would like to avoid.
We may invisibly keep a file handle open for longer than necessary, by not forcing the computation that pulls data from it to be evaluated. Since an operating system will typically place a small, fixed limit on the number of files we can have open at once, if we do not address this risk, we can accidentally starve some other part of our program of file handles.
If we do not explicitly close a file handle, the garbage collector will automatically close it for us. It may take a long time to notice that it should close the file handle. This poses the same starvation risk as above.
We can avoid starvation by explicitly closing a file handle. If we do so too early, though, we can cause a lazy computation to fail if it expects to be able to pull more data from a closed file handle.
On top of these well-known risks, we cannot use a single file handle to supply data to multiple threads. A file handle has a single “seek pointer” that tracks the position from which it should be reading, but when we want to read multiple chunks, each needs to consume data from a different position in the file.
With these ideas in mind, let's fill out the lazy I/O picture.
-- file: ch24/LineChunks.hs chunkedRead :: (FilePath -> IO [ChunkSpec]) -> FilePath -> IO ([LB.ByteString], [Handle]) chunkedRead chunkFunc path = do chunks <- chunkFunc path liftM unzip . forM chunks $ \spec -> do h <- openFile path ReadMode hSeek h AbsoluteSeek (fromIntegral (chunkOffset spec)) chunk <- LB.take (chunkLength spec) `liftM` LB.hGetContents h return (chunk, h)
We avoid the starvation problem by explicitly closing file handles. We allow multiple threads to read different chunks at once by supplying each one with a distinct file handle, all reading the same file.
The final problem that we try to mitigate is that of a
lazy computation having a file handle closed behind its
back. We use rnf
to force all of our
processing to complete before we return from
withChunks
. We can then close our file
handles explicitly, as they should no longer be read from.
If you must use lazy I/O in a program, it is often best to
“firewall” it like this, so that it cannot
cause problems in unexpected parts of your code.
Processing chunks via a fold | |
---|---|
We can adapt the fold-with-early-termination technique from the section called “Another way of looking at traversal” to stream-based file processing. While this requires more work than the lazy I/O approach, it nicely avoids the above problems. |
Since a server log file is line-oriented, we need an efficient way to break a file into large chunks, while making sure that each chunk ends on a line boundary. Since a chunk might be tens of megabytes in size, we don't want to scan all of the data in a chunk to determine where its final boundary should be.
Our approach works whether we choose a fixed chunk size or a fixed number of chunks. Here, we opt for the latter. We begin by seeking to the approximate position of the end of a chunk, then scan forwards until we reach a newline character. We then start the next chunk after the newline, and repeat the procedure.
-- file: ch24/LineChunks.hs lineChunks :: Int -> FilePath -> IO [ChunkSpec] lineChunks numChunks path = do bracket (openFile path ReadMode) hClose $ \h -> do totalSize <- fromIntegral `liftM` hFileSize h let chunkSize = totalSize `div` fromIntegral numChunks findChunks offset = do let newOffset = offset + chunkSize hSeek h AbsoluteSeek (fromIntegral newOffset) let findNewline off = do eof <- hIsEOF h if eof then return [CS offset (totalSize - offset)] else do bytes <- LB.hGet h 4096 case LB.elemIndex '\n' bytes of Just n -> do chunks@(c:_) <- findChunks (off + n + 1) let coff = chunkOffset c return (CS offset (coff - offset):chunks) Nothing -> findNewline (off + LB.length bytes) findNewline newOffset findChunks 0
The last chunk will end up a little shorter than its predecessors, but this difference will be insignificant in practice.
This simple example illustrates how to use the scaffolding we have built.
-- file: ch24/LineCount.hs module Main where import Control.Monad (forM_) import Data.Int (Int64) import qualified Data.ByteString.Lazy.Char8 as LB import System.Environment (getArgs) import LineChunks (chunkedReadWith) import MapReduce (mapReduce, rnf) lineCount :: [LB.ByteString] -> Int64 lineCount = mapReduce rnf (LB.count '\n') rnf sum main :: IO () main = do args <- getArgs forM_ args $ \path -> do numLines <- chunkedReadWith lineCount path putStrLn $ path ++ ": " ++ show numLines
If we compile this program with ghc -O2 --make
-threaded
, it should perform well after an initial
run to “warm” the filesystem cache. On a dual
core laptop, processing a log file 248 megabytes (1.1 million
lines) in size, this program runs in 0.576 seconds using a
single core, and 0.361 with two (using +RTS
-N2
).
In this example, we count the number of times each URL is accessed. This example comes from [Google08], Google's original paper discussing MapReduce. In the map phase, for each chunk, we create a Map from URL to the number of times it was accessed. In the reduce phase, we union-merge these maps into one.
-- file: ch24/CommonURLs.hs module Main where import Control.Parallel.Strategies (NFData(..), rwhnf) import Control.Monad (forM_) import Data.List (foldl', sortBy) import qualified Data.ByteString.Lazy.Char8 as L import qualified Data.ByteString.Char8 as S import qualified Data.Map as M import Text.Regex.PCRE.Light (compile, match) import System.Environment (getArgs) import LineChunks (chunkedReadWith) import MapReduce (mapReduce) countURLs :: [L.ByteString] -> M.Map S.ByteString Int countURLs = mapReduce rwhnf (foldl' augment M.empty . L.lines) rwhnf (M.unionsWith (+)) where augment map line = case match (compile pattern []) (strict line) [] of Just (_:url:_) -> M.insertWith' (+) url 1 map _ -> map strict = S.concat . L.toChunks pattern = S.pack "\"(?:GET|POST|HEAD) ([^ ]+) HTTP/"
To pick a URL out of a line of the log file, we use the bindings to the PCRE regular expression library that we developed in Chapter 17, Interfacing with C: the FFI.
Our driver function prints the ten most popular URLs. As with the line counting example, this program runs about 1.8 times faster with two cores than with one, taking 1.7 seconds to process the a log file containing 1.1 million entries.
Given a problem that fits its model well, the MapReduce programming model lets us write “casual” parallel programs in Haskell with good performance, and minimal additional effort. We can easily extend the idea to use other data sources, such as collections of files, or data sourced over the network.
In many cases, the performance bottleneck will be streaming data at a rate high enough to keep up with a core's processing capacity. For instance, if we try to use either of the above sample programs on a file that is not cached in memory or streamed from a high-bandwidth storage array, we will spend most of our time waiting for disk I/O, gaining no benefit from multiple cores.
[54] As we will show later, GHC threads are extraordinarily lightweight. If the runtime were to provide a way to check the status of every thread, the overhead of every thread would increase, even if this information were never used.
[55] The non-threaded runtime does not understand this option, and will reject it with an error message.
[56] As we write this book, the garbage collector is being retooled to use multiple cores, but we cannot yet predict its future effect.
[57] The genesis of this idea comes from Tim Bray.