Technical Architecture¶

The technical diagram above represents the basic architecture of Arvados.

At the base layer is a "cloud operating system." Currently the platform is deployed on AWS (using EBS as POSIX volumes and EC2 instances as VMs) in addition to an on-site installation based on commodity hardware with Xen VMs. The project roadmap includes OpenStack Integration and potentially other cloud OS.

Key Components¶

Data Manager - The Data Manager helps to orchestrate interactions with data storage. This includes managing rules about permissions, replication, archiving, etc.

Content Addressable Object File Store ("Keep") - Arvados stores files in Keep. Keep is an object file store that has been optimized for big files and write once read many (WORM) scenarios. Keep chunks files into 64MB chunks and distributes them across physical drives or virtual volumes. Keep stores its chunks on any POSIX filesystem. Keep is also a content addressable store (CAS). When a file is stored, each 64MB chunk gets and MD5 hash. Then a "collection" (a text file containing data block hashes) is used to represent the complete set of files. Each collection also has an MD5 hash, which becomes the canonical reference to the set of files.

Pipeline Manager - The Pipeline Manager orchestrates execution of pipelines. It finds jobs suitable for satisfying each step of a pipeline, queues new jobs as needed, tracks job progress, and keeps the metadata database up-to-date with pipeline progress.

MapReduce Engine - The Job Manager executes the distributed processing of the data across cores using the MapReduce system. The Job Manager is optimized for Map steps, and it moves processing to cores that are physically close to where Keep has stored the data. In private clouds where drives and CPUs are on the same node this eliminates disk I/O constraints. (The Job Manager has been optimized for these problems. Another MapReduce engine such as Hadoop could also fulfill this purpose, although no work has been done to enable this.)

In-Memory Compact Genome Database ("Lightning") - Lightning uses a scale-out, open source in-memory database to store genomic data in a compact genome format. VCF files are not suitable for efficient look-ups so we are developing a format to represent variants and other key data for tertiary analysis. Putting this in in a scale-out, in-memory database will make it possible to do very fast queries of these data. (This part of the project is in the design stage.)

API Service - This component provides OAuth2-authenticated REST APIs to Arvados subsystems (metadata database, jobs, etc.) with the notable exception of Keep (which requires direct access to avoid network performance bottlenecks) and VMs and git (which use the SSH protocol and public key authentication).

Workbench - Workbench is a set of visual tools for using the underlying Arvados services from a browser. This is especially helpful for querying and browsing data, visualizing provenance, and monitoring jobs and pipelines. Workbench has a modular architecture designed for seamless integration with other Arvados applications.

Command Line Tools - The CLI tools provide convenient access to the Arvados API from the command line.

SDKs - Arvados provides native language SDKs for Python, Perl, Ruby, R, and Java to make it easier to work with the REST APIs in common development environments. (Some SDKs have not yet been implemented.)

Documentation - In addition to the contributors' wiki on the project site, the Arvados source tree includes a documentation project with four sections:

• User Guide - Introductory and tutorial materials for developers building analysis or web applications using Arvados.

• API Reference - Details of REST API methods and resources, the MapReduce job execution environment, permission model, etc.

• Admin Guide - Instructions to system administrators for maintaining an Arvados installation.

• Install Guide - How to install and configure Arvados on the cloud management platform of your choice.