In the article, Introduction to nonblocking algorithms, author Brian Goetz explains that compared to blocking algorithms, non-blocking algorithms take a more optimistic approach:

[Non-blocking algorithms] proceed with the assumption that there will be no interference. If interference is detected, they back off and retry.

As a result, they tend to outperform blocking algorithms when the number of threads vying for the resource is relatively low:

Under light to moderate contention, nonblocking algorithms tend to outperform blocking ones because most of the time the CAS [compare-and-swap operation] succeeds on the first try, and the penalty for contention when it does occur does not involve thread suspension and context switching, just a few more iterations of the loop....Under high contention—when many threads are pounding on a single memory location—lock-based algorithms start to offer better throughput than nonblocking ones because when a thread blocks, it stops pounding and patiently waits its turn, avoiding further contention.

Although a non-blocking algorithm may tend to perform better than its blocking counterpart in common situations, non-blocking algorithms may also tend to be more complex:

Nonblocking algorithms tend to be far more complicated than lock-based ones.

When do you feel it is appropriate to reach for a non-blocking algorithm? If you think the non-blocking approach is more complex than an equivalent blocking approach, would you consider using the non-blocking approach a performance optimization, one that you should reach for only if you discover you have a real performance problem with a blocking alternative?