작업을 청크로 나누기

Launching a thread per element is wasteful. Instead you split the range into a few larger chunks, one per worker.

Each chunk is just a contiguous slice of indices, like 0 to 99 or 100 to 199. One worker owns one slice.

A common choice is one chunk per core. You ask the machine for that number with num_physical_cores.

Divide the total length by the worker count to get how many items each worker should handle.

Inside the work function, the chunk index tells you where to begin. Multiply it by the chunk size.

The last chunk may run past the data, so clamp its end to the total length with min to stay in bounds.

Each worker runs a normal loop over just its own range. Inside the chunk it is plain, ordinary code.

Now you call parallelize with the number of chunks, not the number of elements. Each call processes one whole chunk.

Equal-sized chunks keep every core equally busy. If one chunk is far bigger, that worker finishes late.

When some items cost more than others, using more, smaller chunks lets fast workers grab extra work and stay busy.

The best chunk count depends on your data and machine. Try a few sizes and keep the one your benchmark likes.

You split a length-n array into one chunk per worker.

You divide the range into balanced chunks, clamp each end with min, loop inside a chunk, and call parallelize over the chunk count for even load. 🚀