DAM: The Dataflow Abstract Machine Simulator Framework

Contents:

1. Using DAM
2. What is DAM?
3. What can I simulate with DAM?
4. Contributing
5. Publications

Using DAM

DAM is a Rust package, currently GitHub-only (awaiting a PR merge into one of the dependencies.).

To use DAM, add the following line into your Cargo.toml under the [dependencies] section:

dam = {git = "https://github.com/stanford-ppl/DAM-RS.git"}

To build the documentation for DAM, run the following command inside of this repository:

cargo +nightly doc

What is DAM

DAM is a framework for building high-performance parallel simulators for dataflow-like systems. DAM comprises of two main components: contexts and channels. Contexts represent the "nodes" in a computational graph, while channels encapsulate the communication between nodes (the "edges").

Context is a trait, which only requires two components:

1. Context::run_falliable(&mut self) -> anyhow::Result<()>, a function which encapsulates the entirety of the execution. This enables the use of near-arbitrary code within.
2. A constructor of any flavor, since #[context_macro] introduces a ContextInfo field used internally within DAM. This is intentional, as will be discussed later.

Communication channels are represented using Sender/Receiver pairs (or just a single Sender when using void channels). In order to keep track of time, users must call Sender::attach_sender(&dyn Context) and Receiver::attach_receiver(&dyn Context) during the initialization phase. We find that the constructor of a Context is the most logical place to insert these calls.

What can I simulate using DAM?

DAM can simulate anything which can be described as things connected by channels, provided that the channels have non-zero latency. Examples include:

• Hardware:
  • Modeling the Sambanova SN40L
• Programming Models:
  • Streaming Tensor Patterns
  • Sparse Abstract Machine

Contributing

If you are interested in contributing to DAM, feel free to shoot Nathan Zhang an email at [email protected].

Publications

The Dataflow Abstract Machine Simulator Framework

ISCA'24

Online PDF

@inproceedings{dam,
  author={Zhang, Nathan and Lacouture, Rubens and Sohn, Gina and Mure, Paul and Zhang, Qizheng and Kjolstad, Fredrik and Olukotun, Kunle},
  booktitle={2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA)}, 
  title={The Dataflow Abstract Machine Simulator Framework}, 
  year={2024},
  volume={},
  number={},
  pages={532-547},
  keywords={Tensors;Machine learning algorithms;Dams;Large language models;Memory management;Machine learning;Parallel processing;Parallel Discrete Event Simulation;Dataflow Accelerators;Modeling},
  doi={10.1109/ISCA59077.2024.00046}}

Papers Using DAM

Implementing and Optimizing the Scaled Dot-Product Attention on Streaming Dataflow

ArXiv

Citation

@misc{sohn2024implementingoptimizingscaleddotproduct,
      title={Implementing and Optimizing the Scaled Dot-Product Attention on Streaming Dataflow}, 
      author={Gina Sohn and Nathan Zhang and Kunle Olukotun},
      year={2024},
      eprint={2404.16629},
      archivePrefix={arXiv},
      primaryClass={cs.AR},
      url={https://arxiv.org/abs/2404.16629}, 
}