A Parallel Cow in a Vacuum, or a Problem of Choosing Parallelization Technology (for CPU and a Shared Memory)

Writing a highly performant code, which fully utilizes modern CPUs, is hard. Even more, this is impossible without using parallelism. This inevitably raises two questions: how many worker threads to create and how to distribute workload among them.

The second question is the most interesting one. To answer it, one should choose from two main paradigms: Work Sharing and Work Stealing. Choosing the right one might be not as easy as it looks. For a middleware library, it is an even more tricky task, as the library can not control (or even know) the end-user application it used in.

We will take a look at both approaches and discuss their benefits and drawbacks. This talk will let you know what the "oversubscription", "works stealing" and other terms mean.