KernMLOps

This repository serves as the mono-repo for the KernMLOps research project.

Currently, it only contains scripts for data collection of kernel performance.

See CONTRIBUTING.md for an overview of how to expand this tool.

WARNING: Do not clone submodules, this will clone the linux kernel. Cloning the kernel is prohibitively expensive. That submodule is only necessary if you plan to use in-kernel inference.

Quick Setup

You will need docker. You can install from docker's website. The tools should work with other container engines but have not been tested yet.

# Create and run your management container -- Expected Time: ~5 mins
make docker-image

# Install YCSB benchmark (Run inside Container) -- Expected Time: ~3 mins
make docker
make install-ycsb

# Copy a simple starting redis script (Run Outside Container) -- Expected Time: ~1 mins
cp config/start_overrides.yaml overrides.yaml

# Capture the simple overrides (Run Outside the Container) -- Expected Time: ~1 mins
make collect

Jupyter Setup

You will need docker and pipenv

You can install docker from docker's website. The tools should work with other container engines but have not been tested yet.

You can install this by using either pip or apt on Ubuntu. A helpful link for this should be here. Or on later versions of ubuntu sudo apt install pipenv. From then on launch jupyter notebook like so:

# Create and run your management container -- Expected Time: ~5 mins
make docker-image

# Create your virtual environment and invoke jupyter -- Expected Time: ~5 mins
pipenv shell jupyter notebook

From here open up the Jupyter not

Capturing Data -> Processing in Python

For this example you need to open two terminals.

In Terminal 1 navigate to your cloned version of KernMLOps.

make docker
make collect-raw

You are looking for output that looks like this:

Hit Ctrl+C to terminate...
Started benchmark faux

This tells you that the probes have been inserted and data collection has begun.

In Terminal 2, start the application. You will eventually need the pid, the terminal can get you that as shown below.

./app arg1 arg2 arg3 &
echo $!
wait

The result of the command should be a pid. The pid can be used later to filter results. When the wait call finishes the program app has exited. Then in Terminal 1 press CTRL+C. Now the data should be collected in ... Now in Terminal 1 you can exit the docker container, enter python and import the last data collection.

The data is now under data/curated in the KernMLOps directory.

You can import that to python by doing the following: Note that you need at least Python 3.12 (as is provided in the container) Change to the python/kernmlops/ directory

cd python/kernmlops/
python3
>>> import data_import as di
>>> r = di.read_parquet_dir("<path-to-data-curated>")
>>> r.keys()

In this case r is a dictionary containing a dataframe per key. If we want to explore for example the dtlb_misses for our program we can do the following:

>>> dtlb_data = r['dtlb_misses']
>>> import polars as pl
>>> dtlb_data.filter(pl.col("tgid") == <pid for program>)

Tools

Python-3.12

This is here to make the minimum python version blatant.

pre-commit

A left shifting tool to consistently run a set of checks on the code repo. Our checks enforce syntax validations and formatting. We encourage contributors to use pre-commit hooks.

# install all pre-commit hooks
make hooks

# run pre-commit on repo once
make pre-commit

perf

Perf counters are used for low-level insights into performance.

When using a new machine it is likely the counters used will be different from those already explicitly supported. Developers can run python python/kernmlops collect perf-list to get the names, descriptions, and umasks of the various hardware events on the new machine. From there developers can add the correct name, umask pair to an existing counter config that already exists.

It is simplest to run the above command inside a container.

Dependencies

Python

Python is required, at least version 3.12 is required for its generic typing support. This is the default version on Ubuntu 24.04.

Python package dependencies are listed in requirements.txt and can be installed via:

# On some systems like Ubuntu 24.04 without a virtual environment
# `--break-system-packages` may be necessary
pip install [--break-system-packages] -r requirements.txt

Contributing

Developers should verify their code passes basic standards by running:

make lint

Developers can automatically fix many common styling issues with:

make format

Usage

Users can run data collection with:

make collect-raw

Configuration

All default configuration options are shown in defaults.yaml, this can be generated via make defaults.

To configure collection, users can create and modify a overrides.yaml file with just the overrides they wish to set, i.e.:

---
benchmark_config:
  generic:
    benchmark: gap
  gap:
    trials: 7

Then make collect or make collect-data will use the overrides set.

If an unknown configuration parameter is set (i.e. benchmark_cfg) and error will be thrown before collection begins.

Troubleshooting: Or How I Learned to Shoot My Foot

eBPF Programs

eBPF Programs are statically verified when the python scripts attempt to load them to into the kernel and that is where errors will manifest. When a program fails to compile the error will be the usual sort of C-compiler error. i.e.

/virtual/main.c:53:3: error: call to undeclared function '__bpf_builtin_memset';
    ISO C99 and later do not support implicit function declarations
    [-Wimplicit-function-declaration]
   53 |   __bpf_builtin_memset(&data, 0, sizeof(data));
      |   ^
1 error generated.

For verification errors the entire compiled bytecode will be printed, look for something along the lines of:

invalid indirect read from stack R4 off -32+20 size 24
processed 59 insns (limit 1000000) max_states_per_insn 0
    total_states 3 peak_states 3 mark_read 3

eBPF Padding

The error:

invalid indirect read from stack R4 off -32+20 size 24
processed 59 insns (limit 1000000) max_states_per_insn 0
    total_states 3 peak_states 3 mark_read 3

Indicates that a read in the program is reading uninitialized memory.

That error came from:

struct quanta_runtime_perf_event data;
data.pid = pid;
data.tgid = tgid;
data.quanta_end_uptime_us = ts / 1000;
data.quanta_run_length_us = delta / 1000;
quanta_runtimes.perf_submit(ctx, &data, sizeof(data));

The invalid read was perf_submit since there was extra padding in the data struct that was not formally initialized. To be as robust as possible this should be handled with an explicit __builtin_memset as in:

struct quanta_runtime_perf_event data;
__builtin_memset(&data, 0, sizeof(data));
data.pid = pid;
data.tgid = tgid;
data.quanta_end_uptime_us = ts / 1000;
data.quanta_run_length_us = delta / 1000;
quanta_runtimes.perf_submit(ctx, &data, sizeof(data));

This gives the most robust handling for multiple systems, see here.