Bench queues and stacks with heap allocated blocks

[polytypic]

Queues and stacks were previously benchmarked with immediate values only, which unfortunately partially hides the potential cost of write barriers related to setting and clearing elements.

Partially hiding the cost of write barriers makes the queue benchmarks unrealistic. Queues and stack are very rarely used with only immediate values.

[polytypic]

I happened to just realize that benchmarking data structures with immediate values gives potentially unrealistic results in OCaml. The problem is that the write barrier for immediate values is much faster than for non-immediate values and that it is relatively rare to actually use data structures with only immediate values.

I developed a new queue that appeared much faster in benchmarks than some previous queues. However, once put into more realistic use, it performed poorly. Now, after months, I finally realized the reason: the real use case exposed the cost of the write barriers.

I would strongly recommend that the benchmarks of queues and stacks here be changed to use heap allocated elements. Otherwise the benchmark results may be misleading (for most use cases) and may lead to poor choice of data structures and implementation techniques.

[lyrm]

Thanks for the PR. I am reviewing it today. Also, I am working on fixing the CI issue.

[lyrm]

I’ve fixed the CI issue. Could you rebase to check if it’s working on your side as well?

Thanks again for the PR! Have you run any benchmarks to see if this changes how the different queues compare to each other ? What about the two-stack queue for example ?

[polytypic]

Have you run any benchmarks to see if this changes how the different queues compare to each other ?

I changed the benchmarks in multicore-bench and picos to allocate the elements just like in this PR.

What about the two-stack queue for example ?

The two-stack queue doesn't need to clear/set values, which means that allocating the elements only adds the allocation cost.