Fully Nonblocking Multiproducer/Consumer Queues
The above code still uses two locks, albeit sparingly. How might we eliminate the producer and consumer locks for a fully nonblocking queue implementation that allows multiple concurrent producers and consumers? If that intrigues you, and you're up for some down-and-dirty details, here are two key papers you'll be interested in reading.
In 1996, Michael and Scott published a paper that presented two alternatives for writing an internally synchronized queue. [4] One alternative really is nonblocking; the other uses a producer lock and a consumer lock, much like the examples in this article. In 2003, Herlihy, Luchango and Moir pointed out scalability limitations in Michael and Scott's approach, and presented their own obstruction-free queue implementation. [5]
Both of these papers featured approaches that require a double-width compare-and-swap operations (also known as "DCAS") that can treat a pointer plus an integer counter together as a single atomic unit. That is problematic because not all platforms have a DCAS operation, especially mainstream processors in 64-bit mode which would essentially require a 128-bit CAS. [6] One also requires a special free list allocator to work properly.
Coming Up
We applied four techniques:
- Having two locks, one for each end of the queue.
- Allocating objects on the heap to let us make consumers more concurrent.
- Having consumers remove consumed nodes one at a time for better locality, less contention at the head, and more immediate cleanup than having producers lazily clean up consumed nodes.
- Adding padding to keep data used by different threads on different cache lines, avoiding memory performance penalties due to false sharing or "ping-pong."
But just how much did each of those help, and how much did each help depending on the size of the queued objects? Next month, I'll break down the four techniques by analyzing the successive performance impact of each of these techniques with some pretty graphs. Stay tuned.
Notes
[1] H. Sutter. "Lock-Free Code: A False Sense of Security" (DDJ, June 2008). Available online at http://ddj.com/architect/208200273.
[2] Note that this happens naturally in Java and .NET for reference types, which are always held indirectly via a pointer (which is called an object reference in those environments).
[3] H. Sutter. "Maximize Locality, Minimize Contention" (DDJ, September 2008). Available online at http://ddj.com/architect/208200273.
[4] M. Michael and M. Scott. "Simple, Fast, and Practical Non-Blocking and Blocking Concurrent Queue Algorithms" (Proceedings of the 15th ACM Symposium on Principles of Distributed Computing, 1996)
[5] M. Herlihy, V. Luchango and M. Moir. "Obstruction-Free Synchronization: Double-Ended Queues As an Example" (Proceedings of the 23rd International Conference on Distributed Computing Systems, 2003).
[6] You could try to make it work in 64 bits via heroic efforts to steal from the 64-bit address space, taking ruthless advantage of the knowledge that on mainstream systems today the operating system usually doesn't actually use all 64 bits of addresses and might not notice if you use a few or even a dozen for your own ends. However, that's inherently brittle and nonportable, and a lot of other people (including probably your OS's developers) have had the same idea and tried to grab those bits too in the frenzied 64-bit land rush already in progress. In reality, you generally only get to play this kind of trick if you're the operating system or its close friend.
Herb is a software development consultant, a software architect at Microsoft, and chair of the ISO C++ Standards committee. He can be contacted at www.gotw.ca.
