mnrl and mncart - web.eecs.umich.eduangstadt/posters/mnrl-gomactech-2018-poster.pdf · mnrl and...

MNRL and MNCaRTAn Open-Source, Multi-Architecture State Machine Research and Execution Ecosystem

Kevin Angstadt† Jack Wadden‡ Vinh Dang‡ Ted Xie‡ Dan Kramp‡

Westley Weimer† Mircea Stan‡ Kevin Skadron‡

†University of Michigan ‡University of Virginia{angstadt, weimerw}@ umich.edu {wadden, vqd8a, ted.xie, dankramp, mircea, skadron}@ virginia.edu

What is Automata Processing?

I Finite automata are state machines thatmatch patterns of data within an inputdata stream

I Theoretical model used to model andprocess regular expressions

I Able to accelerate a broad range ofapplications: natural languageprocessing, network security, graphanalytics, high-energy physics,bioinformatics, pseudo-random numbergeneration and simulation, data-mining,and machine learning.

The MNCaRT Philosophy

I Previous research and tools werefractured by different file formats andrestrictive licenses

I The MNRL Network Computation andResearch Testbed aims to open-sourcethe end-to-end toolchain

I All tools operate on common file formatI Languages, Compilers, and RuntimesI Support for all major architectures

MNRL: A Unified Representation

I MNRL Network Representation LanguageI Flexible, JSON-based language for

representing state machinesI MNRL file defines a collection of state

machines (network) consisting of one ormore nodes

I Direct support for validation of filesI APIs for both C++ and PythonI Easy to extend for experimental

automata node types

MNCaRT Ecosystem

PCRE RAPIDHigh-Level

Languages ANMLZoo

Benchmarks

MNRL ANMLState Machine

Representations

hscompile AutomataLab VASim ATR REAPR apcompile

AnalysisCompilation

Transformation

HyperscanCPU

Engine

VASimCPU

EngineDFAGE iNFAnt2

GPU Engines

FPGAEngine

MicronD480 AP

Execution

Engines

CPU Engines

I Hyperscan: state-of-the-art regularexpression processing engine, extendedto support MNRL networks

I VASim: supports execution of customautomata designs with experimentalnode types and collection of runtimestatistics and profiles

GPU Engines

I Support for both DFA(DFAGE) and NFA(iNFAnt2) executionon Cuda-enabled graphics cards

I Supports the study of trade-offs betweenthe two execution models

REAPR FPGA Engine

I MNRL networks arecompiled to a custommodule for execution onFPGAs (logic-basedfabrics of LUTs and memory)

I Input/Output handled by Xilinx’s SDAccelFramework

Micron D480 AP

I Prototype architecturefor execution ofautomata

I Memory arrays used as computationalmedium

I Supported by converting MNRL to ANMLrepresentation

Automata Lab

Languages and Benchmarks

I Perl Compatible Regular ExpressionsI RAPID: High-level language for defining

inexact pattern matchesI ANMLZoo: Benchmark suite of diverse

automata and regular expressions withassociated input workloads

VASim & Automata-to-Routing

I Supports transformation, optimization,and execution of finite automata

I Prototype and test new automatadesigns and computational styles

I Experiment with place and routealgorithms for customspatial-reconfigurable architectures

Acknowledgements

This work was supported in part by grants from the NSF (CCF-1116673,CCF-1629450, CCF-1619123, CNS-1619098), AFRL (FA8750-15-2-0075),Jefferson Scholars Foundation, Achievement Rewards for College Scientists(ARCS) Foundation, a grant from Xilinx, and support from C-FAR, one of sixcenters of STARnet, a Semiconductor Research Corporation programsponsored by MARCO and DARPA. Any opinions, findings and conclusions orrecommendations expressed in this material are those of the authors and donot necessarily reflect the views of AFRL.

https://bit.ly/CAP-MNCaRT