Skip to content
forked from zpzim/SCAMP

The fastest way to compute matrix profiles on CPU and GPU!

License

Notifications You must be signed in to change notification settings

albertoazzari/SCAMP

 
 

Repository files navigation

Build and Test RTD Build Status

Docker Build and Push Docker Image Version (latest semver) Docker Image Size (latest semver) Docker Pulls

PyPI PyPI - Downloads

Conda (channel only) Conda Conda

Conda (channel only) Conda Conda

DOI

SCAMP: SCAlable Matrix Profile

Table of Contents

Overview
Documentation
Performance
Python Module
Run Using Docker
Distributed Operation
Reference

Overview

This is a GPU/CPU implementation of the SCAMP algorithm. SCAMP takes a time series as input and computes the matrix profile for a particular window size. You can read more at the Matrix Profile Homepage This is a much improved framework over GPU-STOMP which has the following additional features:

  • Tiling for large inputs
  • Computation in fp32, mixed fp32/fp64, or fp64 (double is recommended for most datasets, single precision will work for some)
  • fp32 version should get good performance on GeForce cards
  • AB joins (you can produce the matrix profile from 2 different time series)
  • Distributable (we use GCP but other cloud platforms can work) with verified scalability to billions of datapoints
  • More types of matrix profiles! KNN, Matrix Summary, Sum, and 1NN without index! See the Docs!
  • Extremely Efficient Implementation
  • Extensible to adding optimized versions of custom join operations.
  • CPU Support (Only enabled for double precision; does not support KNN joins yet)
  • Handles NaN input values. The matrix profile will be computed while excluding any subsequence with a NaN value
  • Python module: Use SCAMP in Python with pyscamp
  • conda-forge integration: Install pyscamp seamlessly with conda.
  • Extensive integration testing: SCAMP has thousands of input configurations tested with every PR.
  • Automatic benchmarking: Helps ensure performance doesn't slip with future updates.

Why use SCAMP?

  • It is faster than other matrix profile libraries. For example, it is 20x to 100x faster than stumpy.
  • It is very easy to install using conda and has very few dependencies.
  • It handles real data: very large inputs, missing values, and flat regions with little issue.
  • It can compute various other types of matrix profiles, including efficiently computing KNN matrix profiles, and matrix summaries (a.k.a. mplots). And can be extended to compute other types of profile efficiently.

Documentation

SCAMP's documentation can be found at readthedocs.

Python module

pyscamp is available through conda-forge:

# To install pyscamp with cpu/gpu support on Linux and Windows.
conda install -c conda-forge pyscamp-gpu

# To install pyscamp with cpu support only on Windows, Linux, or MacOS.
conda install -c conda-forge pyscamp-cpu

Note that pyscamp-gpu can be installed and used even if you don't have a GPU, it will simply fall back to using your CPU. However, pyscamp-cpu is preferrable if you don't have a GPU because it builds with a newer compiler and does not require installing the cudatoolkit depencency.

If you run into problems using GPUs with pyscamp-gpu make sure your NVIDIA drivers are up to date. This is the most common cause of issues.

Installing from source

If you want you can build pyscamp from source which will have improved performance. A source distribution for a python3 module using pybind11 is available on pypi.org to install run:

# Python 3 and a c/c++ compiler is required.
# cmake is required (if you don't have it you can pip install cmake)
pip install pyscamp

Once installed you can use SCAMP in Python as follows:

import pyscamp as mp

# Allows checking if pyscamp was built with CUDA and has GPU support.
has_gpu_support = mp.gpu_supported()

# Self join
profile, index = mp.selfjoin(a, sublen)

# AB join using 4 threads, outputting pearson correlation.
profile, index = mp.abjoin(a, b, sublen, pearson=True, threads=4)

More information and the API documentation for pyscamp is available on readthedocs

Run Using Docker

You can run SCAMP via nvidia-docker using the prebuilt image on dockerhub.

In order to expose the host GPUs nvidia-docker must be installed correctly. Please follow the directions provided on the nvidia-docker github page. The following example uses docker 19.03 functionality:

docker pull zpzim/scamp:latest
docker run --gpus all \
   --volume /path/to/host/input/data/directory:/data \
   --volume /path/to/host/output/directory:/output \
   zpzim/scamp:latest /SCAMP/build/SCAMP \
   --window=<window_size> --input_a_file_name=/data/<filename> \
   --output_a_file_name=/output/<mp_filename> \
   --output_a_index_file_name=/output/<mp_index_filename>

Distributed Operation

We have a client/server architecture built using grpc. Tested on GKE but should be possible to get working on Amazon EKS as well.

For more information on how to use the scamp client and server, please take a look at the documentation

Reference

If you use SCAMP in your work, please reference the following paper:

Zimmerman, Zachary, et al. "Matrix Profile XIV: Scaling Time Series Motif Discovery with GPUs to Break a Quintillion Pairwise Comparisons a Day and Beyond." Proceedings of the ACM Symposium on Cloud Computing. 2019.

About

The fastest way to compute matrix profiles on CPU and GPU!

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 74.3%
  • Python 8.9%
  • CMake 7.2%
  • Cuda 5.3%
  • Shell 3.4%
  • C 0.7%
  • Dockerfile 0.2%