Reverse Engineering my CPU's Cache Sizes

Sat, 09 May 2026 17:50:18 +0100

In this investigation, I aim to estimate the sizes of my CPU’s L1D and LLC caches.

General Approach and Hypothesis

The approach I took was to process arrays of varying length and measure median cache miss rates for each array size.

Once the array exceeds the capacity of the CPU’s cache, I expect the cache misses to sharply increase since the array no longer fits entirely in the cache, and some reads will need to go to RAM or the next-level cache.

Cyclops

Sun, 03 May 2026 17:50:37 +0100

Project Overview

Cyclops is a framework for investigating CPU behaviour/architecture and custom code performance.

It allows the user to create experiments using low-level resources like PMU counters and rdtscp, with sensible implementations to improve avoid pitfalls. Cyclops makes it easy to design reproducible and accurate experiments that can be easily orchestrated from the command line and visualised to gain insights.

Features

Batches with user-defined warmup runs, batch runs and aggregation
Workload plugin system to easily write custom workloads with scriptable workload parameters (see docs/workload.md)
Parameter sweeps to run multiple batches while varying one parameter and keeping the others fixed
Metric groups including raw metrics like PMU counters (perf_event_open()) and rdtscp, and ratios like IPC (see docs/metrics.md for measurement methodology)
Results are written to stdout & CSV files with metadata for reproducibility
Python scripts for running and visualising some default experiments - see experiments/

Build and Run

# clone repository
git clone https://mj-penney/cyclops.git cyclops

# enter repository
cd cyclops

# build
make

# run (generate help text)
./cyclops -h

Example Usage

Example 1

Single batch
10 warmup runs
20 batch runs
BRANCH workload with default params
IPC metric group

./cyclops -u 10 -r 20 -w BRANCH -m IPC

Example 2

3 warmup runs
5 batch runs
STRIDED_ARRAY workload
L1D_READS metric group
19 batches, sweeping the array-elements param from 10 to 100 in steps of 5

./cyclops -u 3 -r 5 -w STRIDED_ARRAY -m L1D_READS -s array-elements=10:100:5

Experiments

The cyclops tool is designed to be highly scriptable, and make it easy to design performance & microarchitecture experiments.

Matthew Penney