zyx-db / search-perf

baby's first cpp perf benchmarking

Performance Benchmarking for sorted search

The goal of this project is to measure the performance of searching for an item within a static set. The factors being compared are the data layout and branchless vs branching code.

Measured Structures / Algorithms

There are 2 data structures I am measuring performance of, a sorted set and a sorted map. For each data structure I am implementing 4 versions:

• backed by sorted array and binary search
• backed by sorted array and branchless binary search
• backed by eytzinger layout
• backed by eytzinger layout and a branchless search

In addition for the Map, I am comparing performance when the <Key, Value> is stored together versus in separate sections.

Benchmarking strategy

I will benchmark the structures with varying sizes, to account for performance when memory does and doesnt fit in cache lines. The sizes I plan are:

• powers of 2 from 2^1 to 2^24
• multiples of 1000, from 1000 to 100,000

(i am including the powers of 2 due to potential issues with cache coherence)

Tools

• perf: record cpu cycles and cache misses

Setup

1. generate n random integers
2. populate all structures
3. generate 1 million queries

Runs

1. start with a dry run, not measured in benchmarks
2. start timer
3. run the 1 million queries
4. end timer

Interpretation

I will be looking

Results

WIP.

Inspiration

I recently came across Algorithms for Modern Hardware by Sergey Slotin. I found it to be a great read! He does a great job of gathering lots of useful information in one place. This inspired me to give performance benchmarking in cpp a try, specifically for the eytzinger layout and branch predication.

I have also came across multiple blogs on data oriented design, and wanted to test the performance impact on a structure like a map.