When we wrapped up our discussion of feruscore in the previous chapter, we'd constructed a program that could discover corewars warriors through simulated natural selection. This was done by writing evolved warriors out to disk, using Rust's OS process interface to call out to pmars—the de facto standard MARS—and competing them to discover their relative fitness. We used Rayon—Rust's very convenient data parallelism library—to distribute the workload of competitions between available CPUs. Unfortunately, the implementation was pretty slow. Building a tournament selection criteria was maybe more difficult to express than we might have hoped—though I'm sure there a bright-spark of a reader out there who will improve that substantially and wow me. The real pain point was serializing every warrior to disk multiple times, allocating similar structures repeatedly to establish each round, and then eating pmars' allocation and parsing overhead. It's this last step, calling out to an external program to run competitions, is something we'll address in this chapter. We will also address the other two issues because, well, why not go all-in?