Little tips to improve multicore code

One thing that I love about Obj-C and Cocoa is how seamlessly Apple is introducing funcionalities to take advantage of multicore devices.
One of the most common things that a dev does by hourly basis is enumerate a collection.
In Obj-C we can takle this with different approaches:
1 – old plain C for/while loops
2 – fast enumeration protocol (the for..in)
3 – Create an NSEnumerator object from the collection
4 – Use enumeration methods declared in the object interface

The 4th approach is present in almost each kind of collections and one of its methods is the

- (void)enumerateObjectsWithOptions:(NSEnumerationOptions)opts usingBlock:(void (^)(id obj, BOOL *stop))block

The opts parameter is a bit mask where we can pass this really interesting option:

NSEnumerationConcurrent

As the name states this makes the enumeration process a concurrent task and it’s a pleasure to see in instruments the 2 cores working all togheter. There is one caveat as stated by Apple engineers “the code of the block must be safe against concurrent invocation”.

One other stuff that I really love is the way we can help thread safety using blocks.
Recently for a social-instagram-like app I needed to use extensively the AFNetworking library.
To take trace of the networking operation created by each view controller I created a sort of register of the living connection.
The problem in this scenario is the async nature of the process related to the networking tasks. After adding an operation to the register, is impossible to know when it will be  finished (and consequently removed from the register) and this could happen while adding another one the register. We need to put things in order.

To help us we can use a thread locking mechanism,  I’ve found really uselfull the dispatch_barrier mechanism that has less overhead compared to an exclusive lock.

First we need to create our concurrent queue.
In the reading method we create a synch block that waits for the reading to end, if more cores are avalaible more reading request can occur, in the writing method we use a barrier to be sure that no other code has access until the writing finishes. Since GCD queue are executed in first in – first out order we can guarantee that the block would be executed in the same order we have requested.
As you can see from the picture (taken from the book Effective Obj-C 2.0) while the writing block is executing no reading (or writing) can happen.

Read-Write process

Read-Write process

In code is pretty much something like that.

First we create a our concurrent queue

syncQueue = dispatch_queue_create("it.shootshare.registeroperation", DISPATCH_QUEUE_CONCURRENT);

Then we override our properties:

- (NSString*) name {
  NSString * __block aName = nil;
  dispatch_sync(syncQueue, ^{
     aName = _name;
  });
  return aName;
}

- (void) setName:(NSString*)aName {
   dispatch_barrier_async(syncQueue, ^{
     _name = aName;
  });
}

Sorry, the comment form is closed at this time.

Related Posts