[inaworld]

I once heard about a very large scale multiprocessor configuration with shared cached virtual memory and or disk storage. I don't know the specific terminology, but a cluster of DEC Alphas sharing a vast mapped memory space, not pure disk caching but something more like object caching. As an analogy imagine running all of Facebook across multiple servers with shared storage, and the cached objects are elements of Facebook pages, such as you and I sharing the same photos and simultaneously updating/tagging them. It is critical that we always retrieve consistent states, but it is not critical that my tagging happens strictly before your tagging.

Locks were not available, so it was simply not an option. So they instead chose a lockless "ethernet with memory space collisions" analog to normal "ethernet with wire timing collisions"

The implementation was to allow concurrent overlapping readers and writers into shared memory, but reading was done backward through memory, and writing was done forward. Readers could thus tell (checksum) if their read got corrupted by an overlapping write, and they could cancel and reinitiate. As long as there are not a large number of collisions, it's a very fast system with no waiting. Again, imagine the intensity and criticality of Facebook object sharing; on a human scale there might be a lot of concurrency and benefit to caching, but at processor speed there will not be much read/write contention.

There was more to the scheme which I don't exactly remember. For example, you can see that a writer can also tell if the write was corrupted by a concurrent write by following with a reread, but that's not what they did. I think there were two pools of shared storage, memory chunks and tables of handles to chunks, and the writers coordinated their lockless shared writes via the table of handles to the chunks using the same forward/backward contention (or really corruption) detection.