Is this really a good strategy? I'm sure it works in a lots of cases, but reading the paper (https://www.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf https:/

Is this really a good strategy? I'm sure it works in a lots of cases, but reading the paper (https://www.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf https://www.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf) I wonder if they wouldn't often be better served by a faster locking scheme. If the collisions truly are rare, an optimized optimistic CAS nonblocking locking scheme seems like it shouldn't add much overhead. At least in the relatively common case of a single CPU with a shared L3, I'm surprised that locking would be the bottleneck.