Server applications Performance matters Servers cost (a lot of) money We want as high throughput as possible Scalability: Performance shouldn't degrade (too much) when the number of clients increase Observation: In most (HTTP) servers, the majority of clients are idle at any given point in time
Why Haskell? Simple programming model: Light-weight threads (forkIO) Blocking system calls server = forever $ do sock <- accept serverSock forkIO $ talk sock >> sClose sock talk sock = do req <- recv sock send sock (f req)
Why Haskell? Performance: Lots of concurrency Statically compiled; should perform favorably in comparison with e.g. Python and Ruby Alternative to C++ or Java when performance matters Correctness: Pure functions Strong static typing
What we are missing Support for a large number of concurrent connections Support for a large number of active timeouts Typically one per connection
Implementing light-weight threads Schedule many light-weight threads across a set of OS threads. To avoid blocking the OS threads, use the select system call to monitor multiple file descriptors using a single OS thread.
Non-blocking I/O refresher select: a system call for polling the status of multiple file descriptors. A call to select returns when one or more file descriptors are ready for reading writing, or a timeout occurs. Only call a potentially blocking system call (e.g. recv) when we know it won't block!
Reading data IOReq = Read Fd (MVar ()) | Write Fd (MVar ()) read fd = do waitForReadEvent fd c_read fd waitForReadEvent fd = do m <- newEmptyMVar atomicModifyIORef watechedFds (\xs -> (Read fd m : xs, ())) takeMVar m
Sleeping/timers data DelayReq = Delay USecs (MVar ()) threadDelay time = waitForDelayEvent time waitForDelayEvent usecs = do m <- newEmptyMVar target <- calculateTarget usecs atomicModifyIORef delays (\xs -> (Delay target m : xs, ())) takeMVar m
I/O manager event loop eventLoop delays watchedFds = do now <- getCurrentTime (delays', timeout) <- expire now delays readyFds <- select watchedFds timeout watchedFds' <- wakeupFds readyFds watchedFds eventLoop delays' watchedFds' expire _  = return (, Never) expire now ds@(Delay d m : ds') | d <= now = putMVar m () >> expire now ds' | otherwise = return (ds, Timeout (d - now))
I/O manager event loop cont. wakeupFds readyFds fds = go fds  where go  fds' = return fds' go (Read fd m : fds) fds' | fd `member` readyFds = putMVar m () >> go fds fds' | otherwise = go fds (Read fd m : fds') go (Write fd m : fds) fds' | fd `member` readyFds = putMVar m () >> go fds fds' | otherwise = go fds (Read fd m : fds')
The problem select: ~O(watched file descriptors) Most file descriptors are idle! Limited number of file descriptors (FD_SETSIZE) Iterating through all watched file descriptors every time around the event loop. Timeouts are kept in a list, sorted by time Insertion: O(n) as we need to keep the list sorted
A scalable I/O manager Scalable system calls epoll, kqueue, and some Windows thing... Better data structures Trees and heaps instead of lists
Timeouts New I/O manager uses a priority search queue Insertion: O(log n) Getting all expired timeouts: O(k*(log n - log k)), where k is the number of expired timeouts The API for timeouts is quite limited (to say the least!) One function: threadDelay
Priority search queues allows us to adjust/cancel pending timeouts
Priority search queue performance Used Criterion extensively to benchmark and verify micro optimizations. Biggest performance gains: Specialized to a single type of key/priority Strict sub-trees Unpacked data types Used QuickCheck to make sure that the optimizations didn't break anything.
Light-weight threads vs event loop
Reading/writing Scalable system calls epoll/kqueue: O(active file descriptors) Unblocking threads that are ready to perform I/O O(log n) per thread, using an IntMap from file descriptor to MVar Total running time for k active file descriptors is O(k * log n) instead of O(n).
Send 1M 1-byte messages through pipes
Aside: Good tools are important ThreadScope Helped us find a pathological case in an interaction between atomicModifyIORef and GHC's scheduler
Current status Close to feature complete Needs more testing benchmarking Final step remaining: Integrate into GHC
Conclusions Haskell is (soon) ready for the server! We still need: High performance HTTP server High performance HTML combinator library Composable and secure HTML form generation Formlets + cross-site scripting protection Scalable and distributed data store We could just write binding to an existing one (but where's the fun in that!)