cudf.pandas: the Zero Code Change GPU Accelerator for pandas

Pandas is flexible, but often slow when processing gigabytes of data. Many frameworks promise higher performance, but they often support only a subset of the Pandas API, require significant code change, and struggle to interact with or accelerate third-party code that you can’t change. RAPIDS cuDF enables Pandas users to accelerate their existing workflows and third-party code with zero code change required. You can continue using Pandas on CPUs for small-scale local development and testing, and then enable “Pandas accelerator mode” to run on GPUs when you want better performance. Accelerator mode uses the GPU for supported operations and the CPU otherwise, dispatching under the hood as needed, and is compatible with most third-party libraries.

Transcript

1. cudf.pandas: the Zero Code Change GPU Accelerator for pandas Jacob Tomlinson, Senior Software Engineer | PyData Exeter Meetup Feb 2023
2. NVIDIA RAPIDS PyData Team OSS PyData maintainers hired by NVIDIA to make GPU acceleration ubiquitous and many more…
3. • Pandas and alternatives • Demo • How it works • FAQs and limitations • Conclusion Outline
4. Pandas and alternatives • Pandas is great (but slow) • Why is it slow? o Largely single-threaded o Not a query engine! • Many alternatives : o Faster underlying implementation (C++, Rust, CUDA) o Query engines o SQL-inspired o Distributed computing o Hardware accelerated (GPUs) Results of the H2O.ai benchmark maintained by DuckDB: https://duckdblabs.github.io/db-benchmark/
5. cuDF: GPU DataFrames • Pandas-like API, runs on the GPU • Powered by CUDA and libcudf, a C++ DataFrame library for GPUs • Operations are ~10-100x faster than pandas • Provides 60-75% of the pandas API • Not what this talk is about! cuDF speedups relative to Pandas for a number of different operations NVIDIA A100, AMD EPYC 7642 48-Core Processor
6. "Should I switch from pandas to something else?" • Reasons to use something other than pandas: o Performance above all o Data size o Rewriting code ≠ problem • Reasons to use pandas: o API flexibility o Collaboration o Ecosystem built on pandas o pandas is getting faster
7. What is cudf.pandas? • Lets you keep using pandas o Accelerates it on the GPU with no changes • 100% of the pandas API o Uses the GPU for supported operations o Falls back to using the CPU otherwise • 3rd-party code acceleration o Everything is accelerated. No one changes their code Jupyter/IPython: %load_ext cudf.pandas Command line: python –m cudf.pandas script.py Direct import: import cudf.pandas cudf.pandas.install()
8. Demo Time https://github.com/shwina/pydata-global-2023-demo
9. Demo Recap • Accelerates your code on GPUs with no changes • Key to getting good performance is to minimize CPU execution • 3rd party libraries written to use pandas can be accelerated on the GPUs
10. Under the hood • How does it work?
11. Under the hood • How does it work? o Proxy objects that dispatch to cudf or pandas
12. Under the hood • How does it work? o Proxy objects that dispatch to cudf or pandas o Deep import customization to hijack pandas imports
13. Under the hood • How does it work? o Proxy objects that dispatch to cudf or pandas o Deep import customization to hijack pandas imports • What about...
14. Under the hood • How does it work? o Proxy objects that dispatch to cudf or pandas o Deep import customization to hijack pandas imports • What about... o Duck typing? ▪ Doesn't work for free functions like pd.read_csv ▪ Lots of code doing hard isinstance checks
15. Under the hood • How does it work? o Proxy objects that dispatch to cudf or pandas o Deep import customization to hijack pandas imports • What about... o Duck typing? ▪ Doesn't work for free functions like pd.read_csv ▪ Lots of code doing hard isinstance checks o DataFrame Standard API? ▪ Solves a different problem (developer-focused API) ▪ Exciting possibilities! • Fallback to a faster DataFrame library like Polars?
16. FAQs
17. FAQs • Will my code run up to 100x faster with no code changes? o Yes, with idiomatic pandas usage o The profiler helps you identify where it's falling back to the CPU § As a bonus, you'll likely improve performance on CPUs
18. FAQs • Will my code run up to 100x faster with no code changes? o Yes, with idiomatic pandas usage o The profiler helps you identify where it's falling back to the CPU § As a bonus, you'll likely improve performance on CPUs • How much of the pandas API does this support? o 100%, with the following caveats § Some operations fall back to using the CPU via pandas § There may be small differences from pandas o We test against the pandas unit test suite (94% tests passing)
19. FAQs • Will my code run up to 100x faster with no code changes? o Yes, with idiomatic pandas usage o The profiler helps you identify where it's falling back to the CPU § As a bonus, you'll likely improve performance on CPUs • How much of the pandas API does this support? o 100%, with the following caveats § Some operations fall back to using the CPU via pandas § There may be small differences from pandas o We test against the pandas unit test suite (94% tests passing) • Will cudf.pandas work with <insert third party library>? o Yes, if the library uses pandas in a standard way o Some known limitations: § Isinstance() checks for numpy arrays § Use of the C-API to talk to NumPy or Pandas § Subclassing pd.DataFrame (this kinda works)
20. FAQs • Will my code run up to 100x faster with no code changes? o Yes, with idiomatic pandas usage o The profiler helps you identify where it's falling back to the CPU § As a bonus, you'll likely improve performance on CPUs • How much of the pandas API does this support? o 100%, with the following caveats § Some operations fall back to using the CPU via pandas § There may be small differences from pandas o We test against the pandas unit test suite (94% tests passing) • Will cudf.pandas work with <insert third party library>? o Yes, if the library uses pandas in a standard way o Some known limitations: § Isinstance() checks for numpy arrays § Use of the C-API to talk to NumPy or Pandas § Subclassing pd.DataFrame (this kinda works) • What about working with data larger than GPU memory ? o Right now, this will fall back to using the CPU
21. Get started with cudf.pandas • Code for today's talk: o https://github.com/shwina/pydata-global-2023-demo • Try it on Google Colab: o https://nvda.ws/rapids-cudf • Report issues or feedback on our GitHub repo! o https://github.com/rapidsai/cudf
22. Thank you! Social links at https://jacobtomlinson.dev