TU-Berlin-DIMA/fast-interconnects

Research project on scaling GPU-accelerated data management to large data volumes. Code base of two SIGMOD papers.

Project Triton

What is Project Triton?

Project Triton is a research project that aims to scale data management on GPUs to a large data size by utilizing fast interconnects. Fast interconnects such as NVLink 2.0 provide GPUs with high-bandwidth, cache-coherent access to main memory. Thus, we want to unlock higher DBMS query performance with this new class of hardware!

In this project, we rethink database design to take full advantage of fast interconnects. GPUs can store only several gigabytes of data in their on-board memory, while current interconnect technologies (e.g., PCI Express) are too slow to transfer data ad hoc to the GPU. In contrast, CPUs are able to access terabytes of data in main memory. Thus, GPU-based systems run into a data transfer bottleneck.

Fast interconnects provide a path towards querying large data volumes "out-of-core" in main memory. The Triton Project explores the ways in which database management systems can take advantage of fast interconnects to achieve a high data volume scalability.

Guides

We provide a series of guides to setup Project Triton on your hardware, and on how we tuned our code for IBM POWER9 CPUs and Nvidia Volta GPUs:

• Setup Guide

• Benchmarking Guide

• Problems when getting started

• Huge pages tuning

• NUMA in the context of fast interconnects

• POWER9 Microarchitecture Tuning

Code Structure

Project Triton provides the following applications and libraries:

• datagen is a application and library to generate data with data distributions. It is used as a library by data-store and tpch-bench.
• data-store is a library for generating relational data sets. It is used by hashjoin and radix-join.
• hashjoin is an application to execute and benchmark hash joins on CPUs and GPUs.
• microbench is a collection of microbenchmarks for CPUs, GPUs, and GPU interconnects.
• numa-gpu is a library with abstractions and tools to program GPUs with and without NVLink.
• radix-join is an application to execute and benchmark radix joins on CPUs and GPUs. The distinction from hashjoin enables a specialized API for radix joins.
• sql-ops is a library that implements SQL operators. These are used by hashjoin and radix-join.
• tpch-bench is an application to execute and benchmark TPC-H on CPUs and GPUs. Currently, Query 6 is implemented.

Code documentation is available on GitHub Pages, and linked in the above list. You can also build it yourself by running:

cargo doc --document-private-items --no-deps --open

Research

We've published our results from the Triton Project as academic papers:

• Lutz et al., Pump Up the Volume: Processing Large Data on GPUs with Fast Interconnects, SIGMOD 2020

• Lutz et al., Triton Join: Efficiently Scaling to a Large Join State on GPUs with Fast Interconnects, SIGMOD 2022

To cite our works, add these BibTeX snippets to your bibliography:

@InProceedings{lutz:sigmod:2020,
  author        = {Clemens Lutz and Sebastian Bre{\ss} and Steffen Zeuch and
                  Tilmann Rabl and Volker Markl},
  title         = {Pump up the volume: {Processing} large data on {GPUs} with
                  fast interconnects},
  booktitle     = {{SIGMOD}},
  pages         = {1633--1649},
  publisher     = {{ACM}},
  address       = {New York, NY, USA},
  year          = {2020},
  doi           = {10.1145/3318464.3389705}
}

@InProceedings{lutz:sigmod:2022,
  author        = {Clemens Lutz and Sebastian Bre{\ss} and Steffen Zeuch and
                  Tilmann Rabl and Volker Markl},
  title         = {Triton join: {Efficiently} scaling to a large join state
                  on {GPUs} with fast interconnects},
  booktitle     = {{SIGMOD}},
  pages         = {1017–1032},
  publisher     = {{ACM}},
  address       = {New York, NY, USA},
  year          = {2022},
  doi           = {10.1145/3514221.3517911}
}