Why NSOperationQueue is faster than GCD or performSelectorOnMainThread when they process a large number of tasks on the main Thread?

Question

for example, i have 100 times the for loop. and need to update UIImageView,and the last 2 method is same slowly. why? what is the different between they?

//fastest
    [[NSOperationQueue mainQueue] addOperationWithBlock:^{
                       [btnThumb setImage:[UIImage imageWithData:data] forState:UIControlStateNormal];
                        [scrollView addSubview:btnThumb];
                   }];
//slowly
                     dispatch_async(dispatch_get_main_queue(), ^
                    {
                        [btnThumb setImage:[UIImage imageWithData:data] forState:UIControlStateNormal];
                        [scrollView addSubview:btnThumb];
                    });       
//slowly
                   [btnThumb setImage:[UIImage imageWithData:data] forState:UIControlStateNormal];
                   [self performSelectorOnMainThread:@selector(testMethod:) withObject:[NSArray arrayWithObjects:scrollView, btnThumb, nil] waitUntilDone:NO];

    -(void) testMethod:(NSArray*)objs
    {

        UIScrollView *scroll = [objs objectAtIndex:0];
        UIButton *btn = [objs lastObject];
        [scroll addSubview:btn];
    }

Answer 1

For posterity, lest there be any doubt about this, let's consider the following test harness, built in a simple empty application template. It does 1000 operations using each mechanism, and also has a run loop observer, so we can see how our enqueued asynchronous tasks relate to the spinning of the main run loop. It logs to console, but does so asynchronously, so that the cost of NSLog isn't confounding our measurement. It also intentionally blocks the main thread while it enqueues NSOperations/dispatch_asyncs/performSelectors tasks, so that the act of enqueuing also doesn't interfere. Here's the code:

#import "NSAppDelegate.h"

dispatch_queue_t gLogQueue;
#define NSLogAsync(...) dispatch_async(gLogQueue, ^{ NSLog(__VA_ARGS__); });

@implementation NSAppDelegate
{
    dispatch_group_t g;
    NSUInteger numOps;
    useconds_t usleepDuration;
}

static void MyCFRunLoopObserverCallBack(CFRunLoopObserverRef observer, CFRunLoopActivity activity, void *info);

- (void)applicationDidFinishLaunching:(NSNotification *)aNotification
{
    // parameters of test
    numOps = 1000;
    usleepDuration = 1000;

    // Set up a serial queue so we can keep the cost of calling NSLog more or less out of our test case.
    gLogQueue = dispatch_queue_create(NULL, DISPATCH_QUEUE_SERIAL);
    // Group allows us to wait for one test to finish before the next one begins
    g = dispatch_group_create();

    // Insert code here to initialize your application
    CFRunLoopObserverRef rlo = CFRunLoopObserverCreate(NULL, kCFRunLoopAllActivities, YES, 0, MyCFRunLoopObserverCallBack, NULL);
    CFRunLoopAddObserver(CFRunLoopGetCurrent(), rlo, kCFRunLoopCommonModes);
    CFRelease(rlo);

    NSCondition* cond = [[NSCondition alloc] init];


    dispatch_async(dispatch_get_global_queue(0, 0), ^{
        NSTimeInterval start = 0, end = 0;

        // pause the main thread
        dispatch_async(dispatch_get_main_queue(), ^{
            [cond lock];
            [cond signal];
            [cond wait];
            [cond unlock];
        });

        // wait for the main thread to be paused
        [cond lock];
        [cond wait];

        // NSOperationQueue
        for (NSUInteger i = 0; i < numOps; ++i)
        {
            dispatch_group_enter(g);
            [[NSOperationQueue mainQueue] addOperationWithBlock:^{
                NSLogAsync(@"NSOpQ task #%@", @(i));
                usleep(usleepDuration); // simulate work
                dispatch_group_leave(g);
            }];
        }

        // unpause the main thread
        [cond signal];
        [cond unlock];

        // mark start time
        start = [NSDate timeIntervalSinceReferenceDate];
        // wait for it to be done
        dispatch_group_wait(g, DISPATCH_TIME_FOREVER);
        end = [NSDate timeIntervalSinceReferenceDate];
        NSTimeInterval opQDuration = end - start;
        NSLogAsync(@"NSOpQ took: %@s", @(opQDuration));

        // pause the main thread
        dispatch_async(dispatch_get_main_queue(), ^{
            [cond lock];
            [cond signal];
            [cond wait];
            [cond unlock];
        });

        // wait for the main thread to be paused
        [cond lock];
        [cond wait];

        // Dispatch_async
        for (NSUInteger i = 0; i < numOps; ++i)
        {
            dispatch_group_enter(g);
            dispatch_async(dispatch_get_main_queue(), ^{
                NSLogAsync(@"dispatch_async main thread task #%@", @(i));
                usleep(usleepDuration); // simulate work
                dispatch_group_leave(g);
            });
        }

        // unpause the main thread
        [cond signal];
        [cond unlock];

        // mark start
        start = [NSDate timeIntervalSinceReferenceDate];
        dispatch_group_wait(g, DISPATCH_TIME_FOREVER);
        end = [NSDate timeIntervalSinceReferenceDate];
        NSTimeInterval asyncDuration = end - start;
        NSLogAsync(@"dispatch_async took: %@s", @(asyncDuration));

        // pause the main thread
        dispatch_async(dispatch_get_main_queue(), ^{
            [cond lock];
            [cond signal];
            [cond wait];
            [cond unlock];
        });

        // wait for the main thread to be paused
        [cond lock];
        [cond wait];

        // performSelector:
        for (NSUInteger i = 0; i < numOps; ++i)
        {
            dispatch_group_enter(g);
            [self performSelectorOnMainThread: @selector(selectorToPerfTask:) withObject: @(i) waitUntilDone: NO];
        }

        // unpause the main thread
        [cond signal];
        [cond unlock];

        // mark start
        start = [NSDate timeIntervalSinceReferenceDate];
        dispatch_group_wait(g, DISPATCH_TIME_FOREVER);
        end = [NSDate timeIntervalSinceReferenceDate];
        NSTimeInterval performDuration = end - start;
        NSLogAsync(@"performSelector took: %@s", @(performDuration));

        // Done.
        dispatch_async(dispatch_get_main_queue(), ^{
            CFRunLoopRemoveObserver(CFRunLoopGetCurrent(), rlo, kCFRunLoopCommonModes);
            NSLogAsync(@"Done. NSOperationQueue: %@s dispatch_async: %@s performSelector: %@", @(opQDuration), @(asyncDuration), @(performDuration));
        });
    });
}

- (void)selectorToPerfTask: (NSNumber*)task
{
    NSLogAsync(@"performSelector task #%@", task);
    usleep(usleepDuration); // simulate work
    dispatch_group_leave(g);
}

static NSString* NSStringFromCFRunLoopActivity(CFRunLoopActivity activity)
{
    NSString* foo = nil;
    switch (activity) {
        case kCFRunLoopEntry:
            foo = @"kCFRunLoopEntry";
            break;
        case kCFRunLoopBeforeTimers:
            foo = @"kCFRunLoopBeforeTimers";
            break;
        case kCFRunLoopBeforeSources:
            foo = @"kCFRunLoopBeforeSources";
            break;
        case kCFRunLoopBeforeWaiting:
            foo = @"kCFRunLoopBeforeWaiting";
            break;
        case kCFRunLoopAfterWaiting:
            foo = @"kCFRunLoopAfterWaiting";
            break;
        case kCFRunLoopExit:
            foo = @"kCFRunLoopExit";
            break;
        default:
            foo = @"ERROR";
            break;

    }
    return foo;
}

static void MyCFRunLoopObserverCallBack(CFRunLoopObserverRef observer, CFRunLoopActivity activity, void *info)
{
    NSLogAsync(@"RLO: %@", NSStringFromCFRunLoopActivity(activity));
}


@end

In the output of this code, we see the following (with irrelevant/repeating portions removed):

RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
RLO: kCFRunLoopBeforeWaiting
RLO: kCFRunLoopAfterWaiting
NSOpQ task #0
RLO: kCFRunLoopExit
RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
RLO: kCFRunLoopBeforeWaiting
RLO: kCFRunLoopAfterWaiting
NSOpQ task #1
RLO: kCFRunLoopExit

... pattern repeats ...

RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
RLO: kCFRunLoopBeforeWaiting
RLO: kCFRunLoopAfterWaiting
NSOpQ task #999
RLO: kCFRunLoopExit
NSOpQ took: 1.237247049808502s
RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
RLO: kCFRunLoopBeforeWaiting
RLO: kCFRunLoopAfterWaiting
dispatch_async main thread task #0
dispatch_async main thread task #1

... pattern repeats ...

dispatch_async main thread task #999
dispatch_async took: 1.118762016296387s
RLO: kCFRunLoopExit
RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
performSelector task #0
performSelector task #1

... pattern repeats ...

performSelector task #999
performSelector took: 1.133482992649078s
RLO: kCFRunLoopExit
RLO: kCFRunLoopEntry
RLO: kCFRunLoopBeforeTimers
RLO: kCFRunLoopBeforeSources
Done. NSOperationQueue: 1.237247049808502s dispatch_async: 1.118762016296387s performSelector: 1.133482992649078

What this shows us is that NSOperations enqueued on the main queue get executed one per pass of the run loop. (Incidentally, this is going to allow views to draw for each operation, so if you're updating a UI control in these tasks as the OP was, this will allow them to draw.) With dispatch_async(dispatch_get_main_queue(),...) and -[performSelectorOnMainThread:...] all enqueued blocks/selectors are called one after the other without letting views draw or anything like that. (If you don't forcibly pause the main runloop while you enqueue tasks, you sometimes can see the run loop spin once or twice during the enqueueing process.)

In the end, the results are about what I expected they would be:

• NSOperationQueue: 1.2372s
• dispatch_async: 1.1188s
• performSelector: 1.1335s

NSOperationQueue will always be slower, because spinning the run loop isn't free. In this test harness, the run loop isn't even doing anything of substance, and it's already 10% slower than dispatch_async. If it were doing anything of substance, like redrawing a view, it would be much slower. As for dispatch_async vs performSelectorOnMainThread: both execute all enqueued items in one spin of the run loop, so the difference is pretty marginal. I expect it's down to message send overhead and managing the retain/releases on the target and argument of the performSelector....

So contrary to the implication of the question, NSOperationQueue is not, objectively, the fastest of the three mechanisms, but rather the slowest. My suspicion is that in OP's case, NSOperationQueue appears faster because its "time to first visible change" is going to be much shorter, whereas for dispatch_async and performSelector all enqueued operations are going to be executed, and only then will the view redraw and show the new state. In the pathological case, I would expect this to mean that only the last frame was ever seen, although if you don't block the main thread while enqueueing, you could get a handful of visible frames (but you'll effectively be dropping most of the frames on the ground.)

Regardless of which asynchronous execution mechanism is objectively fastest, they're all pretty crappy ways to do animation.