Documentation for method
hyper assembled from the following types:
Returns the object.
method hyper(Int(Cool) : = 64, Int(Cool) : = 4)
Returns another Iterable that is potentially iterated in parallel, with a given batch size and degree of parallelism.
The order of elements is preserved.
hyper in situations where it is OK to do the processing of items in parallel, and the output order should be kept relative to the input order. See
race for situations where items are processed in parallel and the output order does not matter.
degree option (short for "degree of parallelism") configures how many parallel workers should be started. To start 4 workers (e.g. to use at most 4 cores), pass
:4degree to the
race method. Note that in some cases, choosing a degree higher than the available CPU cores can make sense, for example I/O bound work or latency-heavy tasks like web crawling. For CPU-bound work, however, it makes no sense to pick a number higher than the CPU core count.
batch size option configures the number of items sent to a given parallel worker at once. It allows for making a throughput/latency trade-off. If, for example, an operation is long-running per item, and you need the first results as soon as possible, set it to 1. That means every parallel worker gets 1 item to process at a time, and reports the result as soon as possible. In consequence, the overhead for inter-thread communication is maximized. In the other extreme, if you have 1000 items to process and 10 workers, and you give every worker a batch of 100 items, you will incur minimal overhead for dispatching the items, but you will only get the first results when 100 items are processed by the fastest worker (or, for
hyper, when the worker getting the first batch returns.) Also, if not all items take the same amount of time to process, you might run into the situation where some workers are already done and sit around without being able to help with the remaining work. In situations where not all items take the same time to process, and you don't want too much inter-thread communication overhead, picking a number somewhere in the middle makes sense. Your aim might be to keep all workers about evenly busy to make best use of the resources available.
You can also check out this blog post on the semantics of hyper and race
HyperSeq object out of the current one.