Misframe

Optimizing Concurrent Map Access in Go

Published Mar 31, 2015

One of the more contentious sections of code in Catena, my time series storage engine, is the function that fetches a metricSource given its name. Every insert operation has to call this function at least once, but realistically it will be called potentially hundreds or thousands of times. This also happens across multiple goroutines, so we’ll have to have some sort of synchronization.

The purpose of this function is to retrieve a pointer to an object given its name. If it doesn’t exist, it creates one and returns a pointer to it. The data structure used is a map[string]*metricSource. The key fact to remember is that elements are only inserted into the map.

Here is a simple implementation. I have excluded the function header and return statement to save space.

var source *memorySource
var present bool

p.lock.Lock() // lock the mutex
defer p.lock.Unlock() // unlock the mutex at the end

if source, present = p.sources[name]; !present {
	// The source wasn't found, so we'll create it.
	source = &memorySource{
		name: name,
		metrics: map[string]*memoryMetric{},
	}

	// Insert the newly created *memorySource.
	p.sources[name] = source
}

I have a benchmark that inserts time series points into the database. Again, each insert has to call this function to get the pointer to the metric source it has to update.

This one gets about 1,400,000 inserts / sec with four goroutines running in parallel (i.e. GOMAXPROCS is set to 4). This may seem fast, but it’s actually slower than having one goroutine do all the work. If you’re thinking lock contention, you’re right.

So, what’s the problem here? Let’s consider a simplified case where there are no inserts into the map. Suppose goroutine 1 wants to get source “a” and goroutine 2 wants to get “b”, and assume “a” and “b” are already in the map. With the given implementation, the first one will grab the lock, get the pointer, unlock, and move on. Meanwhile, the other goroutine is stuck waiting to grab the lock. Waiting on that lock seems like a pretty bad use of time! This gets worse and worse as you add more goroutines.

One way to make this faster is to remove the lock and make sure only one goroutine accesses the map. That’s simple enough but you have to give up scalability. Here’s an alternative that’s just as simple and maintains thread-safety.

This change only takes one more line and an additional character, but will keep getting faster as you scale up.

var source *memorySource
var present bool

if source, present = p.sources[name]; !present { // added this line
	// The source wasn't found, so we'll create it.

	p.lock.Lock() // lock the mutex
	defer p.lock.Unlock() // unlock at the end

	if source, present = p.sources[name]; !present {
		source = &memorySource{
			name: name,
			metrics: map[string]*memoryMetric{},
		}

		// Insert the newly created *memorySource.
		p.sources[name] = source
	}
	// if present is true, then another goroutine has already inserted
	// the element we want, and source is set to what we want.

} // added this line

// Note that if the source was present, we avoid the lock completely!

5,500,000 inserts / sec. This is 3.93 times as fast. Recall that I had four goroutines running in parallel, so this increase makes sense.

This works because we’re never deleting sources, and the addresses don’t change. If we have a pointer address in CPU cache, we can use it safely even if the map is changing below us. Notice how we still need the mutex. If we didn’t have it, there would be a race condition where one goroutine will realize that it has to create the source and insert it, but another may insert it in the middle of that sequence. This way, we only hit the lock during inserts into the map, but those are relatively rare.

My colleague John Potocny suggested that I remove the defer because it has nontrivial overhead. He was right. One more very minor change and I was amazed at the result.

var source *memorySource
var present bool

if source, present = p.sources[name]; !present {
	// The source wasn't found, so we'll create it.

	p.lock.Lock() // lock the mutex
	if source, present = p.sources[name]; !present {
		source = &memorySource{
			name: name,
			metrics: map[string]*memoryMetric{},
		}

		// Insert the newly created *memorySource.
		p.sources[name] = source
	}
	p.lock.Unlock() // unlock the mutex
}

// Note that if the source was present, we avoid the lock completely!

This version gets 9,800,000 inserts / sec. That’s 7 times faster with only about 4 lines changed.

Edit:

Is this correct? Unfortunately, no! There is still a race condition, and it’s easy to find using the race detector. We can’t guarantee the integrity of the map for readers while there is a writer.

Here is the race-free, thread-safe, “correct” version. Using an RWMutex, readers won’t block each other but writers will still be synchronized.

var source *memorySource
var present bool

p.lock.RLock()
if source, present = p.sources[name]; !present {
	// The source wasn't found, so we'll create it.
	p.lock.RUnlock()
	p.lock.Lock()
	if source, present = p.sources[name]; !present {
		source = &memorySource{
			name: name,
			metrics: map[string]*memoryMetric{},
		}

		// Insert the newly created *memorySource.
		p.sources[name] = source
	}
	p.lock.Unlock()
} else {
	p.lock.RUnlock()
}

This version is 93.8% as fast as the previous one, so still very good. Of course, the previous version isn’t correct, so there shouldn’t even be a comparison.