mikeash.com: just this guy, you know?

Posted at 2011-10-14 12:01 | RSS feed (Full text feed) | Blog Index
Next article: Friday Q&A 2011-10-28: Generic Block Proxying
Previous article: Friday Q&A 2011-09-30: Automatic Reference Counting
Tags: fridayqna gcd
Friday Q&A 2011-10-14: What's New in GCD
by Mike Ash  

Happy iPhone Day to all! For those of you stuck waiting in line or waiting for the delivery guy, never fear, for I have the solution to your pain. For today's topic, Jon Shier suggested that I talk about the new features in Grand Central Dispatch in Lion and iOS 5.

Prerequisite Reading
If you're new to GCD, you'll want to become familiar with the basics before diving into the new features in Lion. There are many good references out there, including my own series which starts off with Intro to Grand Central Dispatch, Part I: Basics and Dispatch Queues.

GCD is still the same great library that we know and love, but now has a bunch more goodies.

First, there's a new global queue available, which can be accessed by passing DISPATCH_QUEUE_PRIORITY_BACKGROUND. This queue runs at an extremely low priority and disk IO is throttled. This makes it suitable for ongoing tasks which need to have a minimal impact on the system's interactive use.

Next, we now have the ability to create custom concurrent queues. Previously, custom queues were always serial, and the only concurrent queues supported by GCD were the global queues. Custom concurrent queues allow easy suspension of parallelized tasks. Alongside this, GCD now provides dispatch barriers, which allow a custom concurrent queue to be used much like a reader/writer lock.

Finally, we have the long-awaited GCD for IO. Dispatch IO objects can be created for paths or file descriptors. This not only provides a simpler interface for IO compared to the dispatch source API, but it also allows GCD to more intelligently coordinate IO activities to avoid disk thrashing. To go along with this, there's a new dispatch data type, which efficietly manages noncontiguous data.

There isn't really much more to say about the new global background queue, so let's jump straight to....

Custom Concurrent Queues and Barriers
Creating a custom concurrent queue is easy: pass DISPATCH_QUEUE_CONCURRENT to the dispatch_queue_create function. Serial queues can still be obtained by passing NULL or DISPATCH_QUEUE_SERIAL.

Once created, a concurrent queue acts just as you'd expect. Multiple blocks submitted to it can run in parallel if system load and capabilities permit. Unlike the global queues, you can still suspend/resume custom concurrent queues, making it useful for managing a set of parallel operations.

Dispatch barriers go along with custom concurrent queues. They can be used with two functions: dispach_barrier_async and dispatch_barrier_sync. These work just like dispatch_async and dispatch_sync except that, if used on a custom concurrent queue, the block that's submitted with the barrier function doesn't run concurrently with other work on that queue. Instead, it waits until anything currently executing on the queue is finished, then blocks everything else on the queue while the barrier block executes. Once the barrier completes, execution resumes normally.

Note that these barrier functions are pointless on a serial queue, since every unit of work blocks other work on such a queue. They are non-functional when used on the global queues, where they simply do a non-barrier dispatch_async or dispatch_sync. This is because the global queues are a shared resource and it doesn't make sense to allow a single component to block them for everybody.

Custom concurrent queues and barriers allow for the efficient manipulation of data structures which can be read but not written concurrently. If you're familiar with them, they allow essentially the same capabilities as reader/writer locks from more traditional multithreading techniques.

As an example, let's imagine that we have a NSMutableDictionary that's being used as a cache. NSMutableDictionary is thread safe for reading, but doesn't allow any concurrent access while modifying its contents, not even if the other access is simple reading.

This can easily be done using a custom concurrent queue and barriers. First, we'll create the dictionary and the queue:

    _cache = [[NSMutableDictionary alloc] init];
    _queue = dispatch_queue_create("com.mikeash.cachequeue", DISPATCH_QUEUE_CONCURRENT);

To read from the cache, we can just use a dispatch_sync:

    - (id)cacheObjectForKey: (id)key
        __block obj;
        dispatch_sync(_queue, ^{
            obj = [[_cache objectForKey: key] retain];
        return [obj autorelease];

Because the queue is concurrent, this allows for concurrent access to the cache, and therefore no contention between multiple threads in the common case.

To write to the cache, we need a barrier:

    - (void)setCacheObject: (id)obj forKey: (id)key
        dispatch_barrier_async(_queue, ^{
            [_cache setObject: obj forKey: key];

Because this uses the barrier function, it ensures exclusive access to the cache while the block runs. Not only does it exclude all other writes to the cache while it runs, but it also excludes all other reads, making the modification safe.

The gain for such a simple dictionary isn't too compelling, but for more complicated use cases, where the readers need to carry out an expensive sequence of atomic operations, it can make it easy to write fast, safe concurrent code.

Dispatch Data
Dispatch data objects are clearly included to facilitate dispatch IO, but they stand on their own and can be used as general purpose data containers as well. Dispatch data objects are much like NSData objects, in that they're simple object wrappers around a raw pointer and length. What the data means and how you use it is completely up to you.

There's a major difference from NSData, though, which is that dispatch data objects can be noncontiguous. Fundamentally, an NSData is a single buffer. A dispatch data object is a collection of potentially many such buffers. This can dramatically increase performance because there's often no need to copy data around at all. For example, when concatenating two NSData objects together, at least one of the buffers needs to be copied, and likely both. When concatenating dispatch data objects, nothing needs to be copied. Internally, this is implemented by creating a tree of dispatch data objects, with the leaves containing a single contiguous buffer, and others pointing to the objects that contain the individual buffers.

Of course, a lot of code wants to work with contiguous data, but fortunately GCD makes it easy to squash a dispatch data object together so that it becomes a single buffer. For code that's more flexible, it's easy to iterate over the individual buffers contained within the data object.

To create a basic dispatch data object, use the dispatch_data_create function. It takes a pointer, a length, a destructor block, and a queue on which to run the destructor. By requesting the default destructor, the data will be immediately copied into storage that's managed by GCD:

    dispatch_data_t data = dispatch_data_create(buffer, length, NULL, DISPATCH_DATA_DESTRUCTOR_DEFAULT);
    // buffer can now be freed

Of course, the whole point of this stuff is to avoid copies, so it's better to provide an explicit destructor that will free the memory so that it can avoid being copied. A common case is memory allocated with malloc, and the DISPATCH_DATA_DESTRUCTOR_FREE destructor will call free:

    void *buffer = malloc(length);
    // fill out buffer
    dispatch_data_t data = dispatch_data_create(buffer, length, NULL, DISPATCH_DATA_DESTRUCTOR_FREE);
    // buffer will now be freed when data is destroyed

For other types of buffers, we can provide a custom block to do whatever is necessary. For example, here's a simple function that creates a dispatch data object wrapping an NSData object:

    dispatch_data_t CreateDispatchDataFromNSData(NSData *nsdata)
        // copy the data object here, in case it's mutable
        // this will just be a retain if it's immutable
        nsdata = [nsdata copy];

        dispatch_queue_t queue = dispatch_get_global_queue(0, 0);
        return dispatch_data_create([nsdata bytes], [nsdata length], queue, ^{
            // balance the copy at the top
            [nsdata release];

To concatenate two dispatch data objects, just use dispatch_data_create_concat. To extract a piece of a data object, dispatch_data_create_subrange will grab an exact subrange, and dispatch_data_copy_region will grab a region around a specific location, which gives GCD more flexibility to be efficient.

To access the contents of a data object, the simplest way is to call dispatch_data_create_map which concatenates all of the data into a single contiguous buffer:

    const void *buffer;
    size_t length;
    dispatch_data_t tmpData = dispatch_data_create_map(data, &buffer, &length);
    // use buffer and length here

However, this may need to copy the individual data pieces in order to create a contiguous buffer, which is just what this whole system is trying to avoid. For more efficient access to the contents, dispatch_data_apply is available and will walk through the individual pieces, calling a block on each one.

For example, here is a function which transforms a data object containing ASCII data into an NSString using dispatch_data_apply to avoid unnecessary copies:

    NSString *StringFromDispatchData(dispatch_data_t data)
        NSMutableString *s = [NSMutableString stringWithCapacity: dispatch_data_get_size(data)];
        dispatch_data_apply(data, ^(dispatch_data_t region, size_t offset, const void *buffer, size_t size) {
            [s appendFormat: @"%.*s", size, buffer];
            return (_Bool)true;
        return s;

Note that the applier block returns a boolean indicating whether the apply operation should continue. Since this block continues unconditionally, it just returns true, with some typecasting to placate the compiler.

Dispatch IO
Now we reach the really big new feature in GCD. The original GCD API could integrate with IO through dispatch sources. A dispatch source could be used to monitor a file descriptor and run a handler when data could be read or written. However, this approach still left a lot up to the programmer, who had to manually read and write the data in question and manage file descriptor lifetime. By handing more responsibility over to GCD, it's also able to more intelligently manage multiple concurrent IO operations to reduce thrash and resource contention.

Dispatch IO objects are called channels. A channel wraps a file descriptor. To create one, use the dispatch_io_create function, which takes a channel type (stream or random access), the file descriptor, a queue to associate with the channel, and a cleanup handler. Here's a quick example of making a channel for standard input:

    dispatch_io_t stdinChannel = dispatch_io_create(DISPATCH_IO_STREAM, STDIN_FILENO, dispatch_get_global_queue(0, 0), ^(int error) {
            fprintf(stderr, "got an error from stdin: %d (%s)\n", error, strerror(error));

We'll want to read from this channel, but first we need to configure it. We can tell GCD how often we want data by setting high and low water values for the channel. The low water value sets how much data GCD will try to gather before invoking a read handler. Standard input is often interactive, in which case we want GCD to invoke the read handler when any data comes in no matter how small, so we'll set the low water mark to just one byte:

    dispatch_io_set_low_water(stdinChannel, 1);

There's no reason to limit the maximum amount of data, so we'll leave the high water value at the default of SIZE_MAX, which is essentially infinite. If we had to limit this for some reason, we'd just call dispatch_io_set_high_water.

There's also a dispatch_io_set_interval function, which tells GCD to call the read handler periodically, allowing the code to monitor the progress of the IO operations. This is also unnecessary for simply reading from standard input.

Now that the channel is configured, we'll tell GCD to read from it using dispatch_io_read. This function takes a channel, an offset (ignored in the case of stream channels like this one), a length, a queue, and a handler block:

    dispatch_io_read(stdinChannel, 0, SIZE_MAX, dispatch_get_global_queue(0, 0), ^(bool done, dispatch_data_t data, int error) {
            // process data
            // handle an error, or just let the channel's handler take care of it
            // we've processed all of stdin, so exit

Similarly, we can write to a channel using dispatch_io_write. By creating another channel and inserting a bit of code into the above handler, the program turns into a simple echo tool:

    dispatch_io_t stderrChannel = dispatch_io_create(DISPATCH_IO_STREAM, STDERR_FILENO, dispatch_get_global_queue(0, 0), ^(int error) {
            fprintf(stderr, "got an error from stdout: %d (%s)\n", error, strerror(error));

    dispatch_io_read(stdinChannel, 0, SIZE_MAX, dispatch_get_global_queue(0, 0), ^(bool done, dispatch_data_t data, int error) {
            dispatch_io_write(stderrChannel, 0, data, dispatch_get_global_queue(0, 0), ^(bool done, dispatch_data_t data, int error) {});

In addition to simply dealing with raw file descriptors, GCD also provides dispatch_io_create_with_path, a convenience function for directly getting an IO channel to a file on disk. This essentially combines dispatch_io_create and open, with the convenience of only having to handle errors in one spot.

When done with an IO channel, simply call dispatch_io_close to explicitly close the channel, and don't forget a dispatch_release to balance the create call.

For simple use cases, GCD also provides dispatch_read and dispatch_write calls, which do simple, GCD-based IO from a file descriptor without having to set up channels. This makes it simple to just read a chunk of data all at once, although for more complicated uses, creating a channel is more efficient and easier to work with.

Dispatch IO channels can talk to just about any kind of file descriptor, making this API useful not only for manipulating files, but also sockets and pipes.

Lion and iOS 5 bring some exciting and long-awaited additions to GCD. The new global background queue is useful for long-running, low impact tasks. The ability to create custom concurrent queues allows much better management of parallelized tasks, and barriers allow for safe parallel access to shared data while reading and exclusive access for writing. Finally, the dispatch IO API brings GCD's smarts and system-wide integration to file access and networking.

That's it for today. You may now return to waiting for your shiny new iPhone. Friday Q&A is driven by reader ideas, so if you happen to come up with something you'd like to see covered here while you're playing with your new toy, send it in!

Did you enjoy this article? I'm selling a whole book full of them. It's available for iBooks and Kindle, plus a direct download in PDF and ePub format. It's also available in paper for the old-fashioned. Click here for more information.


Brent Gulanowski at 2011-10-14 13:36:53:
So, when do we get foundation containers and file abstractions that have (optional) integrated support for re-entrancy based on this? Seems like a no-brainer to encapsulate the custom dispatch queue right into the object. Until then, I guess we use custom wrappers. (NSMutableDictionary<-(NS)ReentrantDictionary?)

mikeash at 2011-10-14 21:56:41:
The trouble with thread-safe (note, not reentrant, subtle difference) containers is that they're difficult to create, tend to be slow, and leave their users open to nasty threading bugs anyway.

For example, you might write a method to getthe first item of an array:

- (id)firstItem { return [array count] ? [array objectAtIndex: 0] : nil; }

However, with an array being used by multiple threads, this code is unsafe even if the array itself is nominally thread safe, because another thread could remove the last object from the array between the check and the fetch.

Fundamentally, thread safety doesn't compose. It's painful, but there we are.

What you'd really need is something that allowed for transactions on these containers, which would be really useful but fairly complex, or something which allowed locking around multiple operations which need to be atomic, which works but basically puts you right back to square one.

Tom M at 2011-10-16 02:22:24:
Great summary of the new GCD features.

One minor point: block capture in CreateDispatchDataFromNSData will result in a redundant retain/release taking place for the copied nsdata. Storing the copy in a __block NSData* auto variable would eliminate that.

mikeash at 2011-10-16 04:01:23:
True, but it's not a problem. Are you suggesting that using __block will be faster? I'd want to see measurements on that before accepting it if that's the case, since __block captured variables have additional overhead compared to the regular kind.

Steven Fusco at 2011-10-17 18:10:36:
You can get to priority background and the new barrier API in iOS 4.3, no need to wait for 5. It's marked __OSX_AVAILABLE_STARTING(__MAC_10_7,__IPHONE_4_3)
 in the header

Tom M at 2011-10-18 07:37:19:
It was more an observation, and as you say, it would be foolish to assume which is faster. You got me curious enough to have a quick look, and as it happens it seems to be a wash. Certainly for NSData objects (of the size I was testing, which were pretty small) the difference wasn't statistically different. On multiple runs, it went either way. A quick look with Instruments shows that it's just not spending enough time in the block copy helper to be meaningful compared to other big ticket items (like objc_msgSend, and memory operations).

mikeash at 2011-10-18 14:30:18:
Makes sense, it's going to be pretty fast either way.

Jean-Daniel at 2011-10-19 17:22:24:
To follow up Tom suggestion, using __block has an other big advantage, it works with ARC, where explicit release call is prohibited.

mikeash at 2011-10-19 17:59:43:
Under ARC you need to explicitly reference the data object in the block in some other way, like [nsdata self]. It makes no difference at all whether the captured variable is decorated with __block or not.

David H at 2011-11-16 15:21:37:
Your information on "concurrent queue and barriers" is just EXCELLENT and a huge help to me (who is using gobs of operations and async dispatches on an iPhone app.

Note that DISPATCH_QUEUE_PRIORITY_BACKGROUND is also available on iOS 4.


Don Fulano at 2012-04-04 03:48:59:
I am little unclear why the write block is dispatch_barrier_async and not a dispatch_barrier_sync. Don't you need it to be sync to make sure the data is finished writing before returning from the block. Won't the async return before the block is completed.

mikeash at 2012-04-04 14:31:08:
It doesn't matter if the data isn't written before the method completes. The only way to retrieve data from the cache is to use the dispatch queue, which will wait until all previous writes are complete.

tikhop at 2012-08-12 11:10:22:
Why I should use dispatch_barrier_async instead of @synchronized. I try both approach and @synchronized method little bit faster.

Andrew at 2012-08-13 12:45:57:
This gentleman on stackoverflow has posted the header files for some of these APIs which suggest that these tools were actually available in iOS 4, not iOS 5. i'm a little confused so any assistance is appreciate; many resources around the web suggest only Lion and iOS 5 feature some of these options:


mikeash at 2012-08-13 18:14:06:
tikhop: The semantics are quite different. A custom concurrent queue using barriers will allow multiple active readers simultaneously, or one writer, while @synchronized only ever allows a single thread to be within the synchronized block at a time (with the same lock object). Additionally, using a dispatch queue means that your locked operations can be run asynchronously, whereas code must always wait for a @synchronized block to complete before continuing. There's nothing wrong with @synchronized, and if it fits your needs then by all means use it.

Andrew: I don't understand the relevance of posting the declaration for dispatch_get_global_queue. In any case, the documentation is quite clear that the dispatch_barrier_async function and its friends are available on 10.7+ and 4.3+.

tikhop at 2012-08-21 09:18:42:
@mikeash become much clearer, thanks for the answer!

I also had another test and it showed a lot more interesting: GCD 2x time faster.


roshni at 2013-11-14 12:41:40:
@mikeash I am bit confuse when exactly to use sync and async,Can you elaborate on the same . thanks

HS at 2014-02-07 09:19:30:
@mikeash: I was wondering how you get such type of information. You are an amazingly genius!

jasna at 2015-05-23 20:43:39:
well, it's way late but replying to don fulano:

I am little unclear why the write block is dispatch_barrier_async and not a dispatch_barrier_sync. Don't you need it to be sync to make sure the data is finished writing before returning from the block. Won't the async return before the block is completed.

assuming this is about thread-safe cache code above, you can afford to do async and return immediately, because all your getters are protected by dispatch_sync on the same dispatch queue; so your getter will block until earlier dispatch_barrier_async operation completes

bbum at 2016-03-07 18:36:54:

The collection class documentation does not make any claims as to the thread safety of NSMutableDictionary in a read-only role.


Valeriy Van at 2016-05-10 12:47:23:
Description of dispatch_data_apply doesn't give guarantee applier block will not be fired concurrently. In that case StringFromDispatchData function will crash.

James Alan Bush at 2016-09-06 16:48:59:
Terrific post! Very concise, accurate. One question, though: how would one construct a CMSampleBuffer from a dispatch_data_t object?

That's a big one, I know; but, it would solve a big problem for those needing to read assets (videos) without using AVFoundation...

Comments RSS feed for this page

Add your thoughts, post a comment:

Spam and off-topic posts will be deleted without notice. Culprits may be publicly humiliated at my sole discretion.

Web site:
Formatting: <i> <b> <blockquote> <code>. URLs are automatically hyperlinked.
Code syntax highlighting thanks to Pygments.
Hosted at DigitalOcean.