grouping multiple method calls into one execution a short time later - ios

I'm writing a custom file system cache component that has an index dictionary that represents important attributes about the files within the folder.
This is for an iOS app, and I'm writing in Objective-C
At various points in the implementation of adding objects / deleting object from the cache, the index dictionary needs to be saved to disk.
In order to stop this operation happening needlessly many times over, for example if objects are added to the cache in a for.. loop, I want to make a system that every time the dictionary is modified, a state is set to ensure that at some point in the future the dictionary will be saved. This should not happen immediately however, in case another change is made quickly, in which case the first 'save' operation should not happen, but another one should be queued up.
In pseudo-code:
//This is the method called by all other parts of the program whenever the dictionary is modified and it needs to be changed
-(void) dispatchSaveIndexDictionary {
//cancel any previous requests to save.
//queue up a save operation some short time later.
}
How I've implemented this:
-(void)saveIndexDictionaryDispatchDelayed
{
NSLog(#"dispatching index save");
[NSObject cancelPreviousPerformRequestsWithTarget:self selector:#selector(saveIndexDictionaryWriteToDisk) object:nil];
//Delay 0 means it gets queued up asap, which means the index dictionary on disk remains in sync whenever possible.
// - This is to solve the case of adding multiple objects in a for... loop
[self performSelector:#selector(saveIndexDictionaryWriteToDisk) withObject:nil afterDelay:0];
}
-(void)saveIndexDictionaryWriteToDisk
{
NSLog(#"Writing cache index to disk : %#", self.cachePath);
[NSKeyedArchiver archiveRootObject:self.indexDictionary
toFile:[OFMFileSystemCache indexDictionaryFullPathWithCachePath:self.cachePath]];
}
Using performSelector:withObject:afterDelay:0
I expected that this would always perform the 'write to disk' method AFTER any of the 'dispatch ' operations, i.e. we could have multiple write operations, if tasks took a long time, but that the 'write' operation would always be the last thing to happen.
I've seen from the logs that this does not always happen, if I do the simple use case of adding 10 files to the cache, then sometimes I get 'dispatching index save' happening and no call afterwards to 'Writing cache index to disk'. I don't really understand how this is possible!
Is there some reason why my implementation isn't a good idea (I guess there must be as it doesn't work very well)?
What do you think is a good secure design for this type of delayed method call, as it's critical that the index remains up to date with the contents of the cache. I.E. write to cache should always happen last, after all modifications have been made.
Thanks.

I've done something similar in my caches in the past. What I end up doing instead of using performSelector:afterDelay: is setup an NSTimer. Something like:
// elsewhere, setup an NSTimer* ivar called saveTimer
-(void) saveToDisk{
saveTimer = nil;
// actually save here
}
-(void)resetSaveTimer{
if(saveTimer) [saveTimer invalidate];
saveTimer = [NSTimer scheduledTimerWithTimeInterval:1.0 target:self selector:#selector(saveToDisk) userInfo:nil repeats:NO];
}
-(void)doStuffWithCache{
// do stuff, add stuff, whatever
[self resetSaveTimer];
}
Depending on your threading, you may also want to add #synchronized(yourCacheDictionary){...} for each of the method bodies, just to make sure you're not trying to write to disk while also editing the dictionary, etc.

Following on from adam.wulf, I wanted to post the solution that I have finally settled on.
This uses NSTimer as suggested, because I had found inconsistent behaviours I couldn't explain with the 'NSObject performSelector:afterDelay:' approach.
After trying the timers approach I needed to modify the solution in some important ways.
1 - make sure the timer is dispatched on a queue with a properly set up runloop for timer execution. I'm using an external caching library for some caching operations, and this calls back on queues that are not set up appropriately. The easiest solution for me was to always dispatch the timer call on the main queue.
2 -Dispatch the actual write operation on a dedicated serial queue, so as not to block the main queue (where the timer method will fire as it is dispatched on that queue).
3 - As suggested wrapping the index Dictionary write to disk methods in #synchronized(indexDictionary) { } to ensure the contents are not modified while being written.
#pragma mark - Saving Index Dictionary
-(void)saveIndexDictionaryDispatchDelayed
{
dispatch_async(dispatch_get_main_queue(), ^{
if (_dispatchSaveTimer) {
[_dispatchSaveTimer invalidate];
_dispatchSaveTimer = nil;
}
_dispatchSaveTimer = [NSTimer scheduledTimerWithTimeInterval:0.5
target:self
selector:#selector(saveIndexDictionaryWriteToDisk)
userInfo:nil
repeats:NO];
});
}
-(void)saveIndexDictionaryWriteToDisk
{
dispatch_async(_cacheOperationQueue, ^{
#synchronized (self.indexDictionary) {
_dispatchSaveTimer = nil;
NSLog(#"Writing cache index to disk : %#", self.cachePath);
[NSKeyedArchiver archiveRootObject:self.indexDictionary
toFile:[OFMFileSystemCache indexDictionaryFullPathWithCachePath:self.cachePath]];
}
});
}

Related

Result of calling performSelector: on a "sleeping" thread's current run loop?

I asked myself a question, but have no answer yet so I hope you could have one:
What happened if there is a NSThread having his current NSRunLoop(named runLoop1 for example) in "sleeping" state (well, runLoop1 called [NSThread sleepForTimeInterval:/*...*/]) while an other NSThread is calling [self performSelector:#selector(selector:) onThread:runLoop1 withObject:nil waitUntilDone:NO]?
I hope I'm understandable ^^
The runloop can be explained as an infinite loop:
while(!exit) {
// Do stuff here
}
This runs on a thread and if this thread is in sleep then so is the loop and no events will be called on it.
So then what are performSelector methods:
Imagine there is an array of invocations with that loop which will be executed when appropriate. Since there is a method to perform the selector after a delay there is also a time stamp.
while(!exit) {
NSMutableArray *notExecuted = [NSMutableArray new];
for(Executable *item in [self.pendingExecutables copy]) {
if(item.executionDate && [item.executionDate compare:[NSDate date]] == NSOrderedDescending) {
[notExecuted addObject:item];
}
else {
[item execute];
}
}
self.pendingExecutables = notExecuted;
}
So calling performSelector really does nothing but adds the data needed to perform it into some array. The runloop must be running for the execution to actually happen. So in your case nothing happens, the selector will not be performed because the loop is not executing since the whole thread is sleeping.
You can then also understand what happens if you block the main thread. No touch events, no system notifications, no nothing. It is all just kept in array and it will be called once the thread is unblocked and another loop occurs. The main loop also sends the date on each cycle which is then used for the watch dog. Since the OS works on another thread then your application it is the OS that will check that date and if it is relatively old you get to "application not responding" state. And then the OS may decide to kill your application.
Note that this is oversimplified but it is enough to get a basic understanding on how these things work.

Using multithread to save data on iOS

I am developing an iPhone app which keeps some data. I am using archiving method from NSKeyedArchiver class to save the data to disk. I would like to periodically save the data to disk. The problem is, when the data grows bigger, it takes more time and it actually interrupts the user's current actions.
Hence, I want to use multithreading to solve this problem. From my understanding of multithreading, when I want to save the data to disk, I should create a new thread, run the saving task on the new thread, then terminate the thread. I should also make the thread so that it won't immediately terminate when the app terminates, to finish saving data. This way, the user can continue to interact with the interface.
That being said, I am not familiar with the actual code that does these work...what would the above look like in code?
A couple of thoughts.
You want to use a serial dispatch queue or operation queue.
Note, we probably want it to write to persistent storage serially (if you're saving it to the same filename, for example), i.e. not permit another save to be initiated until the prior save is finished. I suspect that's exceedingly unlikely that your infrequent saves could ever trigger a save while the prior one is still in progress, but as a general principle you should not use concurrent queues unless you write code that supports concurrent operation (which we're not doing here). This means that you do not use the GCD global queues.
For example, to create serial dispatch queue using Grand Central Dispatch (GCD) would be:
#property (nonatomic, strong) dispatch_queue_t queue;
Then instantiate this (e.g. in viewDidLoad):
self.queue = dispatch_queue_create("com.domain.app.savequeue", 0);
Then use this queue
dispatch_async(self.queue, ^{
// do your saving here
});
For a review of concurrency technologies, see the Concurrency Programming Guide. Both dispatch queues (GCD) and operation queues are solid choices.
You might want to be careful about synchronization issues. What if your app proceeds to start changing the data while the save is in progress? There are a bunch of options here, but the easiest is to copy the data to some temporary object(s) in the main queue before you dispatch the save task to the background queue:
// copy the model data to some temporary object(s)
dispatch_async(self.queue, ^{
// save the temporary object(s) here
});
Or, instead of creating a copy of the model, you can alternatively (and this is a little more complicated if you're not familiar with GCD) use a variation of the "reader-writer" pattern that Apple discusses in WWDC 2012 video Asynchronous Design Patterns with Blocks, GCD, and XPC. Bottom line, you can queue to not only perform asynchronous write to persistent storage, but also to synchronize your updates to your model using a "barrier" (see Using Barriers in the GCD reference):
self.queue = dispatch_queue_create("com.domain.app.modelupdates", DISPATCH_QUEUE_CONCURRENT);
Then, when you want to save to disk, you can do
dispatch_async(self.queue, ^{
// save model to persistent storage
});
But, whenever you want to update your model, you should use barrier so that the updating of the model will not happen concurrently with any read/save tasks:
dispatch_barrier_async(self.queue, ^{
// update model here
});
And, whenever you read from your model, you would:
dispatch_sync(self.queue, ^{
// read from model here
});
Theoretically, if you're worried about the possibility that you could conceivably do your save operations so frequently that one save could still be in progress when you initiate the next one, you might actually employ two queues, one serial queue for the saving operation (point 1, above), and the concurrent queue outlined here for the synchronization process.
Finally, Putz1103 is correct, that if it's possible that the app can be terminated while a save is in progress, you might want to add the code to allow the write to persistent storage to complete:
dispatch_async(self.queue, ^{
UIBackgroundTaskIdentifier __block taskId = [[UIApplication sharedApplication] beginBackgroundTaskWithExpirationHandler:^(void) {
// handle timeout gracefully if you can
[[UIApplication sharedApplication] endBackgroundTask:taskId];
taskId = UIBackgroundTaskInvalid;
}];
// save model to persistent storage
// when done, indicate that the task has ended
if (taskId != UIBackgroundTaskInvalid) {
[[UIApplication sharedApplication] endBackgroundTask:taskId];
taskId = UIBackgroundTaskInvalid;
}
});
Adding multi-threading to an application where data is shared between multiple threads (in this case, the data being created by the user and the data you are saving) is a difficult task to manage.
Instead of creating a thread or trying to save off all the data at once, put data to be saved into an internal "this must be saved off" list, and then work it off one N elements at a time periodically in your main thread.
If you get to the point where the user is leaving the screen or the app, then save off all the work that is left in the queue to the database immediately.
You can create a simple timed event (a few times per second) to do the work, which is a very simple approach.
You can explicit control over how many items you save per update.
You should never have concurrency issues.
You should never have to worry about thread start/stop/termination
issues or mutexes.
To create it:
-(void)viewDidAppear:(BOOL)animated
{
self.timer = [NSTimer scheduledTimerWithTimeInterval:1.0f target:self selector:#selector(saveData) userInfo:nil repeats:YES];
[[NSNotificationCenter defaultCenter] addObserver:self selector:#selector(saveAllRemainingItems) name:#"APP EXITING KEY" object:nil]
}
Your update function:
-(void)saveData
{
if([itemsToSave count] > 0)
{
....save off N items, remove from the list
}
}
-(void)saveAllRemainingItems
{
while([itemsToSave count] > 0)
{
...save first item.
[itemsToSave removeObjectAtIndex:0];
}
}
When you leave:
-(void)viewWillDisappear:(BOOL)animated
{
[self.timer invalidate];
[[NSNotificationCenter defaultCenter] removeObserver:self];
[self saveAllRemainingData];
}
To make sure you handle the "app is closing" situation, in your app delegate:
- (void)applicationWillTerminate:(UIApplication *)application
{
[[NSNotificationCenter defaultCenter] postNotificationName:#"APP EXITING KEY" object:nil];
...OTHER CLEANUP ACTIVITIES
}
You can achive mulithreading in IOS with different ways, like NSThread, Operation Queues and GCD.
GCD is the best approch now a days, it uses block. You can execute a code in different thread like this. You can use this in any method.
void performArchiveData{
dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
//Now you are in different thread. You can add your code in here.
});
}

Using Blocks and GCD to manage tasks

I'm learning iOS and when it comes to GCD, it's confusing.
Let's get it out of the way, I'm writing a small program that fetch data from the internet.
Here is my viewcontroller
NSMutableArray dataArray = [NSMutableArray array];
[querysomethingwithblock:(^ {
//do some stuff here
[otherquerywithblock:( ^ {
//do some stuff here
// Here I got the data from internet
// Do loop action
[dataArray addObject:data];
})];
})];
// here I want to perform some actions only after get data from internet
[self performAction:dataArray];
How can I achieve this purpose. In practical, [self performAction:dataArray] always get fired before I get the data. I tried to play with GCD but no luck.
Here is some patterns I've tried so far
dispatch_async(queue, ^{
// Do query stuff here
dispatch_async(dispatch_get_mainqueue(), ^{
//perform action here
});
{;
Or using dispatch_group_async, dispatch_group_wait, dispatch_group_notify
The only way I can handle right now is to use dispatch_after but the point is the downloading time is variable, it's not good practice to have a specific time here
Thank you so much for any advice.
The part of code called Do query stuff here i assume is async already, why put it inside a dispatch_queue then?
If instead you manage to do a synchronous query, your code (the second snippet) would work, as the dispatch to the main queue would be executed only after the query finished.
If you don't have an option to execute the query in a synchronous manner, then you need some mechanism to register either a block or a callback to be executed when the download is finished.
At the end of the day, it all depends on what kind of query you have in there and what methods it offers for you to register an action to be performed when the download is finished.

NSRunLoop's runMode:beforeDate: - the correct approach for setting the "beforeDate"

I have a doubt regarding the correct usage of NSRunLoop's runMode:beforeDate method.
I have a secondary, background thread that processes delegate messages as they are received.
Basically, I have process intensive logic that needs to be executed on a background thread.
So, I have 2 objects, ObjectA and AnotherObjectB.
ObjectA initializes AnotherObjectB and tells AnotherObjectB to start doing it's thing. AnotherObjectB works asynchronously, so ObjectA acts as AnotherObjectB's delegate. Now, the code that needs to be executed in the delegate messages, needs to be done on a background thread. So, for ObjectA, I created an NSRunLoop, and have done something like this to set the run loop up:
do {
[[NSRunLoop currentRunLoop] runMode:NSDefaultRunLoopMode beforeDate:[NSDate distantFuture]];
} while (aCondition);
Where aCondition is set somewhere in the "completion delegate message".
I'm getting all my delegate messages and they are being processed on that background thread.
My question being: is this the correct approach?
The reason I ask this is because [NSDate distantFuture] is a date spanning a couple of centuries. So basically, the runLoop won't timeout until "distantFuture" - I definitely won't be using my Mac or this version of iOS till then. >_<
However, I don't want the run loop to run that long. I want the run loop to get done as soon as my last delegate message is called, so that it can properly exit.
Also, I know that I can set repeating timers, with shorter intervals, but that is not the most efficient way since it's akin to polling. Instead, I want the thread to work only when the delegate messages arrive, and sleep when there are no messages. So, is the approach I'm taking the correct approach, or is there some other way of doing it. I read the docs and the guide, and I set this up based off what I understood from reading them.
However, when not completely sure, best to ask this awesome community for an opinion and confirmation.
So, thanks in advance for all your help!
Cheers!
The code is in the docs:
If you want the run loop to terminate, you shouldn't use this method. Instead, use one of the other run methods and also check other arbitrary conditions of your own, in a loop. A simple example would be:
BOOL shouldKeepRunning = YES; // global
NSRunLoop *theRL = [NSRunLoop currentRunLoop];
while (shouldKeepRunning && [theRL runMode:NSDefaultRunLoopMode beforeDate:[NSDate distantFuture]]);
where shouldKeepRunning is set to NO somewhere else in the program.
After your last "message", un-set shouldKeepRunning (on the same thread as the run loop!) and it should finish. The key idea here is that you need to send the run loop an event so it knows to stop.
(Also note that NSRunLoop is not thread-safe; I think you're supposed to use -[NSObject performSelector:onThread:...].)
Alternatively, if it works for your purposes, use a background a dispatch queue/NOperationQueue (but note that code which does this shouldn't touch the run loop; things like starting a NSURLConnection from a dispatch queue/NSOperationQueue worker thread will likely cause problems).
The reason I ask this is because [NSDate distantFuture] is a date spanning a couple of centuries.
The method runMode:beforeDate: will
return NO immediately if there are no sources scheduled on the RunLoop.
return YES whenever an event has been processed.
return YES when the limitDate has been reached.
So even if the limitDate is very high, it will return after every processed event, it will not keep running until limitDate has been hit. It would only wait for that long if no event is ever processed. limitDate is thus like a timeout after that the method will give up on waiting for an event to take place. But if you want to have multiple events in a row handled, you must call this method over and over again, hence the loop.
Think of fetching packets with timeout from a network socket. The fetch call returns when a packet arrives or when the timeout has been hit. Yet if you want to process the next packet, you must call the fetch method again.
The following is unfortunately pretty bad code for two reasons:
// BAD CODE! DON'T USE!
NSDate * distFuture = NSDate.distantFuture;
NSRunLoop * runLoop = NSRunLoop.currentRunLoop;
while (keepRunning) {
[runLoop runMode:NSDefaultRunLoopMode beforDate:distFuture];
}
If no RunLoopSource is yet scheduled on the RunLoop, it will waste 100% CPU time, as the method will return at once just to be called again and that as fast as the CPU is able to do so.
The AutoreleasePool is never renewed. Objects that are autoreleased (and even ARC does that) are added to the current pool but are never released as the pool is never cleared, so memory consumption will raise as long as this loop is running. How much depends on what your RunLoopSources are actually doing and how they are doing it.
A better version would be:
// USE THIS INSTEAD
NSDate * distFuture = NSDate.distantFuture;
NSRunLoop * runLoop = NSRunLoop.currentRunLoop;
while (keepRunning) #autoreleasepool {
BOOL didRun = [runLoop runMode:NSDefaultRunLoopMode beforDate:distFuture];
if (!didRun) usleep(1000);
}
It solves both problems:
An AutoreleasePool is created the first time the loop runs and after every run it is cleared, so memory consumption will not raise over time.
In case the RunLoop didn't really run at all, the current thread sleeps for one millisecond before trying again. This way the CPU load will be pretty low since as as no RunLoopSource is set, this code only runs once every millisecond.
To reliably terminate the loop, you need to do two things:
Set keepRunning to NO. Note that you must declare keepRunning as volatile! If you don't do that, the compiler may optimize the check away and turn your loop into an endless loop since it sees no code in the current execution context that would ever change the variable and it cannot know that some other code somewhere else (and maybe on another thread) may change it in the background. This is why you usually need a memory barrier for these cases (a lock, a mutex, a semaphore, or an atomic operation), as compilers don't optimize across those barriers. However, in that simple case, using volatile is enough, as BOOL is always atomic in Obj-C and volatile tells the compiler "Always check thes value of this variable as it may change behind your back without you seeing that change at compile time".
If the variable has been changed from another thread and not from within an event handler, your RunLoop thread may be sleeping inside the runMode:beforeDate: call, waiting for a RunLoopSource event to take place which may take any amount of time or never happen at all anymore. To force this call to return immediately, just schedule an event after changing the variable. This can be done with performSelector:onThread:withObject:waitUntilDone: as shown below. Performing this selector counts as a RunLoop event and the method will return after the selector was called, see that the variable has changed and break out of the loop.
volatile BOOL keepRunning;
- (void)wakeMeUpBeforeYouGoGo {
// Jitterbug
}
// ... In a Galaxy Far, Far Away ...
keepRunning = NO;
[self performSelector:#selector(wakeMeUpBeforeYouGoGo)
onThread:runLoopThread withObject:nil waitUntilDone:NO];

NSOperation deadlocks and blocks NSOperationQueue

I use a subclass of NSOperation to upload large files to AWS S3 using Amazon's iOS SDK (v1.3.2). This all works fine, but some beta testers experience deadlocks (iOS 5.1.1). The result is that the NSOperationQueue in which the operations are scheduled is blocked as only one operation is allowed to run at one time. The problem is that I cannot reproduce the issue whereas the beta testers experience this problem every single time.
The operation is quite complex due to how the AWS iOS SDK works. However, the problem is not related to the AWS iOS SDK as far as I know based on my testing. The operation's main method is pasted below. The idea of the operation's main method is based on this Stack Overflow question.
- (void)main {
// Operation Should Terminate
_operationShouldTerminate = NO;
// Notify Delegate
dispatch_async(dispatch_get_main_queue(), ^{
[self.delegate operation:self isPreparingUploadWithUuid:self.uuid];
});
// Increment Network Activity Count
[self incrementNetworkActivityCount];
// Verify S3 Credentials
[self verifyS3Credentials];
while (!_operationShouldTerminate) {
if ([self isCancelled]) {
_operationShouldTerminate = YES;
} else {
// Create Run Loop
[[NSRunLoop currentRunLoop] runMode:NSDefaultRunLoopMode beforeDate:[NSDate distantFuture]];
}
}
// Decrement Network Activity Count
[self decrementNetworkActivityCount];
NSLog(#"Operation Will Terminate");
}
The method that finalizes the multipart upload sets the boolean _operationShouldTerminate to YES to terminate the operation. That method looks like this.
- (void)finalizeMultipartUpload {
// Notify Delegate
dispatch_async(dispatch_get_main_queue(), ^{
[self.delegate operation:self didFinishUploadingUploadWithUuid:self.uuid];
});
// Operation Should Terminate
_operationShouldTerminate = YES;
NSLog(#"Finalize Multipart Upload");
}
The final log statement is printed to the console, but the while loop in the main method does not seem to exit as the final log statement in the operation's main method is not printed to the console. As a result, the operation queue in which the operation is scheduled, is blocked and any scheduled operations are not executed as a result.
The operation's isFinished method simply returns _operationShouldTerminate as seen below.
- (BOOL)isFinished {
return _operationShouldTerminate;
}
It is odd that the while loop is not exited and it is even more odd that it does not happen on any of my own test devices (iPhone 3GS, iPad 1, and iPad 3). Any help or pointers are much appreciated.
The solution to the problem is both complex and simple as it turns out. What I wrongly assumed was that the methods and delegate callbacks of the operation were executed on the same thread, that is, the thread on which the operation's main method was called. This is not always the case.
Even though this was true in my test and on my devices (iPhone 3GS), which is why I did not experience the problem myself. My beta testers, however, used devices with multicore processors (iPhone 4/4S), which caused some of the code to be executed on a thread different from the thread on which the operation's main method was invoked.
The result of this is that _operationShouldTerminate was modified in the finalizeMultipartUpload method on the wrong thread. This in turn means that the while loop of the main method was not exited properly resulting in the operation deadlocking.
In short, the solution is to update _operationShouldTerminate on the same thread as the main method was invoked on. This will properly exit the while loop and exit the operation.
There are a number of problems with your code, and I can offer two solutions:
1) read up on Concurrent NSOperations in Apple's Concurrency Programming Guide. To keep the runLoop "alive" you have to add either a port or schedule a timer. The main loop should contain a autorelease pool as you may not get one (see Memory Management in that same memo). You need to implement KVO to let the operationQueue know when your operation is finished.
2) Or, you can adopt a small amount of field tested hardened code and reuse it. That Xcode project contains three classes of interest to you: a ConcurrentOperation file that does well what you are trying to accomplish above. The Webfetcher.m class shows how to subclass the concurrent operation to perform an asynchronous URL fetch from the web. And the OperationsRunner is a small helper file you can add to any kind of class to manage the operations queue (run, cancel, query, etc). All of the above are less than 100 lines of code, and provide a base for you to get your code working. The OperationsRunner.h file provide a "how to do" too.

Resources