https://github.com/coastalwhite/amdahls-lie

A demonstration of breaking Amdahl's Law with Caches
https://github.com/coastalwhite/amdahls-lie

Last synced: about 1 month ago
JSON representation

A demonstration of breaking Amdahl's Law with Caches

• Host: GitHub
• URL: https://github.com/coastalwhite/amdahls-lie
• Owner: coastalwhite
• License: mit
• Created: 2024-03-21T15:16:22.000Z (8 months ago)
• Default Branch: main
• Last Pushed: 2024-04-02T11:32:44.000Z (7 months ago)
• Last Synced: 2024-04-02T13:47:15.629Z (7 months ago)
• Language: Rust
• Size: 11.7 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Amdahl's Lie - Speeding up code beyond the theoretical limit

This repository provides a proof-of-concept that shows how to break Amdahl's Law using caches. This tries to show that `n`-threaded applications can be more than `n` times as fast. This works for applications that can partition their work into memory slices.

This is by no means supposed to be a criticism of Amdahl's Law which is extremely useful in parallel programming.

## Explanation

This program is just a proof-of-concept to idea behind breaking Amdahl's Law.

The program is divided into a `lib.rs` and `main.rs` as to simplify benchmarking. The `lib.rs` contains all the interesting logic. There are three implementations: `singlethreaded`, `singlethreaded_batched` and `multithreaded`. To each of these implementations, you pass (i) a `set` of data, (ii) a list of `requests`, and (iii) a configuration (`cfg`). Each implementation will then call `handle_request` for each request and return a vector with all the results.

The `handle_request` function aims to do a large calculation over data that is present in the `set`. This is not that weird of a situation as the same might happen in for example a database. In this`handle_request` implementation, we unpredictably load a bunch of bytes and create a sum over them. This range of bytes that we work on is dependent on the `request.section`.

## License

Licensed under an MIT license.