Introduction

MsPASS Features

MsPASS is an acronymn that stands for Massively Parallel Analysis System for Seismologists. Some key features of MsPASS are the following:

  • As the name suggests MsPASS is a domain-specific package for seismologists. We emphasize, however, that our definition of seismologists is broad. It includes exploration geophysicists and anyone in any field that needs to work with data that match the data models described in this manual.

  • Few seismologists have strong expertise in modern information technology. Furthermore, installing a software package can prove challenging today even on a desktop on which you may have special privileges. On large high performance (HPC) systems installation is usually impossible to install special software without a long string of meetings and correspondence with system managers. For this reason MsPASS uses modern container technology to simplify installation. That allows you as the user to install MsPASS without system privileges. Containers are also essential for operating the package in a cloud system.

  • Another keyword in the acronym is “parallel”. Computer technology reached the physical limit of single CPU computers many years ago. NO existing open-source package for handling earthquake data provided consistent support for parallel processing until MsPASS.

  • A first order goal of MsPASS was building a consistent framework to allow scalability from a single desktop machine with multiple cores to a giant cloud-based system. In MsPASS we developed a simplified API for running processing in parallel. The API makes it relatively easy to prototype a workflow with a test data set before porting it to a large cluster to handle a more massive processing job.

  • MsPASS uses an integrated database management system that was known previously to perform well in a massively parallel environment. We use a stable, open-source system called MongoDB. Our database API aims to abstract interactions with MongoDB as much as possible, but some knowledge of the query language used by MongoDB will be required to use the package effectively. There are several published books on MongoDB and extensive documentation and tutorials are available online for the package. Web searches for MongoDB commands or a book at your side are an essential tool for working with MsPASS at any level.

  • MsPASS uses python as the job control language. This is in contrast to traditional programs like SAC or older seismic processing systems that use a custom command interpreter. We assert that custom languages like that in SAC or even the unix shell will become the equivalent of Latin as a language in the coming years. Python has emerged as the most common glue language used within open-source software packages. It was thus the clear choice as the driver language for MsPASS.

  • Although python has huge advantages as “glue language” and as a way to quickly develop prototypes, it also has a serious limitation. Python, like matlab’s command line language, is an interpreted language. That means it is essentially compiled on the fly. Some classes of algorithms will run orders of magnitude faster if implemented in a compiled language like C/C++ or FORTRAN compared to python. For this reason, we implemented core data objects in C++. Any python package with any hope of performance follows this same model. Obspy does this by manipulating sample data with numpy and some core C libraries. Antelope python does this as well by implementing core functions in C. We follow this model. Python functions and classes implemented in C/C++ are all found in the MsPASS hierarchy as all modules under mspasspy.ccore.

  • Errors are a universal issue in large-scale data processing. As the size of a data set increases it is becomes an increasingly difficult to guarantee all the data are “clean”. By that, we mean one or more algorithms may treat a particular data problem as an unrecoverable error. Handling errors in a large scale processing environment is problematic for a long list of reasons. We solve this problem in MsPASS by having error logs be an intrinsic part of the data. That approach is actually essential to preserve errors using Spark and DASK. It would be very difficult to sort out problems from verbose log files that would otherwise be saved and interleaved in a scratch, logs directory if we used simpler print statements. Instead in MsPASS the error messages posted to any data object are automatically saved in the database when the data object is saved with cross-references to the data with which that error was associated. In addition, following a well-established approach used is seismic reflection systems since the 1960s, MsPASS provides a integrated “kill” mechanism that allows data to be carried along but ignored. That model maps well to massively parallel scheduling because dead data are treated like live data but they just process faster.

  • To promote reproducible science MsPASS has an integrated processing history capability. The goal of that component of MsPASS is to ultimately allow publication of the processing workflow used in a scientific paper that would allow the reader to reproduce the data that paper used. At the time this user’s manuals was written that part of the system is incomplete. For that reason (and for efficiency) MsPASS processing functions make handling history optional and by default it is turned off. If the system grows as we hope that limitation will disappear.

  • The design of MsPASS has stressed leading edge but not bleeding edge open-source technologies. MsPASS was assembled from a long list of general purpose, open-source packages. Some are C/C++ libraries that are linked with code in the ccore modules (e.g. boost and the GNU Scientific Library) and some are python packages like ObsPy.

Getting Started

The first step to use MsPASS is to install a local copy that you can use for initial experimentation. Click here for installation instructions.

We have an extensive set of tutorials based on jupyter notebooks found here. For most people these are a good way to learn the package on their own.

Organization of User Manual

The titles of the sections of this manual should serve as guides for how to learn more about a topic of interest. Except for this section the manual is not intended to be read in the order of the topics posted as hypertext in the contents page. A learning model most people find most effective is to work through the tutorials and reading sections of this manual as questions arise or by following hypertext links from the tutorials.