Exercices in performance optimizations (expo)

Workshop at GOLAB 2024, 2024-11-11, 14:30, Martin Czygan, LI

Slides

Abstract

The 1 Billion Row Challenge is a simple, data-intensive task, that nonetheless allows to explore many optimization ideas and techniques in Go.

In this workshop, we start with a baseline implementation and interactively improve on the solution, learning about benchmarking, different performance characteristics of standard library types, concurrency patterns, fast data structures, useful operating system facilities and more.

Overview

• Benchmarking
  • writing a benchmark
  • running a benchmark
• Profiling
  • cpu profiling
  • generating a flame graph
• 1BRC problem outline
  • problem description
• A baseline implementation

Variations:

• Caring about allocations
  • ReadString
  • Scanner
  • Scanner buffer size
• Faster string parsing
  • splitting a string
  • parsing a float
  • parsing a float with SWAR
• Parallel processing
  • worker pattern
  • splitting the file
• Using memory-mapped files
  • simplifying the api
• Using a custom hash table
  • custom hash table

Benchmarking mechanics

• benchmark small snippets, separately
• basic benchmark with time

Areas of optimization

• buffered I/O
• allocations, e.g. ReadString vs Scanner
• better buffer sizes, e.g. as passed to Read(...)
• parsing a string
• parsing a float as int
• using a memory-mapped file
• parallel processing
• optimal number of batch size and number of workers
• custom hash function