Instance-Optimized Mapping with Portfolio Methods

Mappings are ubiquitous in computer systems, such as translating virtual memory to physical memory, file paths to inode numbers, database keys to data locations. Traditional system mappings are often hand-crafted and data-agnostic. In this paper, we explore the use of neural networks as learned mappings that are automatically generated and data-dependent, optimizing performance for specific workloads and scenarios. Unlike prior learned structures, we employ a portfolio method consisting of a set of independent neural networks, each responsible for making sole decisions. Our preliminary results indicate that these portfolio mappings can generalize across multiple applications.