QMachine: commodity supercomputing in web browsers

API server

The API server is implemented as a simple Node.js [39] program that loads and executes all of its application logic from QM’s own publicly available module, “qm”, using the Node Package Manager (NPM) [40]. The module supports five different databases as targets for persistent storage: Apache CouchDB [41], MongoDB [42], PostgreSQL [43], Redis [44], and SQLite [45]. These five open-source databases were chosen for support based on their high performance and popularity, and their differences in design help to guide the development of QM as an HPC solution for a heterogeneous database landscape. The relative merits of the alternative implementations to the default use of MongoDB are as follows. CouchDB and MongoDB are both document-centric NoSQL databases with MapReduce APIs that understand JS, but their designs are very different. CouchDB is more than just a database – it is nearly sufficient to implement QM by itself because it bundles a web server and an HTTP-accessible API. MongoDB, by way of contrast, has an API that mimics the traditional relational style used by PostgreSQL and SQLite, with a much stronger focus on clustering and “sharding” (horizontal partitioning) across nodes. PostgreSQL represents relational database management systems (RDBMS), the workhorses that traditionally power enterprise applications and data warehouses, while SQLite represents embedded (serverless) databases. Redis is an in-memory key-value store that is often referred to as a “data structure server” because its keys can contain strings, hashes, lists, sets, and sorted sets. The ability to map QM’s persistent representation layer across such a wide variety of storage systems simplifies deployment and maintenance significantly. The service that backs this report’s illustrative examples, available at https://v1.qmachine.org, uses MongoDB.

QM’s API server supports Cross-Origin Resource Sharing (CORS) [46] so that any webpage can embed QM to distribute workflows across web browsers without violating the Same-Origin Policy [47]. There is wide support for CORS in web browsers [48].

Web server

The web server, like the API server, is implemented as a Node.js program, and its logic is contained in the same NPM module, “qm”. That is, the installation of all of QMachine’s base libraries can be achieved simply by running Node’s built-in module management system: npm install qm. It is worth recalling the minimal role played by the server-side components of QM (see Figure 1 in Results). The web server exists only to provide the presentation/analytical layer’s resources to client machines. Because these resources are static, the web server can be replaced by off-the-shelf web servers like Apache [49] and Nginx [50].

Website

The website functions as the presentation/analytical layer of QM. It was developed as a browser client for the QM API, and thus it is implemented in HTML5, CSS, and JS, as can be verified by inspecting its source code at https://github.com/wilkinson/qmachine. The website consists of a single webpage that communicates with the API server periodically via XMLHttpRequest (XHR) using a technique known as Asynchronous JavaScript and XML (AJAX). The “AJAX” name is a bit misleading, however, because XHR is not limited to handling XML data; all of the browser client’s communications with the API server use JavaScript Object Notation (JSON).

When a browser loads the webpage, it initially loads only the presentation layer, comprised of the HTML, CSS, and JS resources necessary to render the graphical user interface (GUI). Immediately after the GUI loads, the browser retrieves QM’s analytical layer, which is written entirely in JS. This design improves the user experience by loading the GUI faster, and it isolates the presentation layer’s code from the analytical layer’s code. Thus, third parties can embed QM’s analytical layer and thereby use QM’s persistent representation layer without loading QM’s presentation layer, as shown by the examples at https://v1.qmachine.org/barebones.html and http://q.cgr.googlecode.com/hg/index.html.

QM’s browser client models a workflow as a set of transforms that should be applied to input data in a specific order to produce output data. A “task description” is an object that contains the transform f, the data x, and any information needed to prepare the environment prior to execution.

As described above, the client-side application that is distributed when a browser is pointed to https://v1.qmachine.org was developed using only web technologies: HTML5, JS, and CSS. In order to stay within the core JS syntax that is supported by all browsers and all platforms – including mobile devices – code development was assisted by JSLint [51]. JSLint is also used directly within the analytical layer as a static analysis tool to identify tasks which can be serialized faithfully into JSON for distribution to volunteer machines. A generic library, Quanah [52], was also developed to solve the numerous concurrency challenges faced in asynchronous data transfer by QM; it is therefore a key component of the prototype described here and is accordingly also made publicly available with open source. The presentation layer uses jQuery [53] and Twitter Bootstrap [54] to ensure consistent look-and-feel across a variety of mobile and desktop browsers. The GUI additionally attempts to support outdated browsers through optional integration with Google Chrome Frame [55], HTML5 Shiv [56], and json2.js [57], but it does so only as a courtesy.

Demonstration program

The demonstration program is written in pure JavaScript so that it can be run in an ordinary web browser without dependencies on any native applications, plugins, or add-ons. It extends a demonstration from a previous study [36] in which a MapReduce decomposition of a sequence analysis procedure was used to find the longest similar segment between a user-given sequence and a full bacterial genome. The demonstration in this study will not only reproduce the previous results using remote execution on another machine, but it will do so in parallel for all of the twenty strains of Streptococcus pneumoniae that are currently available from the National Center for Biotechnology Information (NCBI). It uses an updated version of the same Universal Sequence Maps (USM) [58] library used by the previous study, as referenced directly by exact version from its online repository.

Loading the client-side library

QM’s analytical layer is provided by a JS library that can be loaded by any web browser automatically as part of any webpage that contains the following code:

Once loaded, the JS environment will contain a global object named QM with convenient high-level methods that can be used to reproduce the results of the four examples that follow.

(1) Simple algebraic operation

For the first illustrative example, let f be a function that increments a given number x by 2, and let x = 2. To compute the result, f (x), on a volunteer machine, we could use the QM.submit method:

As in the rest of the illustrative series, this example is described and demonstrated in the accompanying screencast (http://goo.gl/tnpMiQ). Note also that this simple operation is easily expressed in other languages such as CoffeeScript [38]):

As discussed in “Methods”, QM’s architecture does not impose the use of a specific programming language, as long as a compiler to JS, the web’s “assembler language” [35], is distributed with the remote call. To support this claim, the QM client library delegates to a compiler – written in JS – for the CoffeeScript language. For a list of compilers that can translate programs written in other languages into JS so that they can be interpreted by volunteering browsers, see http://bit.ly/altjsorg.

(2) Simple distributed map

Because each QM.submit operation is an asynchronous call, multiple calls can run simultaneously. Thus, it is straightforward to distribute the execution of a “map” function, a higher-order functional pattern that applies the same operation to each element of an array. This pattern is so ubiquitous in scientific computing that it warrants a dedicated method, QM.map, that can be used as follows:

(3) Simple distributed MapReduce

Just as in the “map” function shown above, it is straightforward to distribute the execution of a “reduce” function, a higher-order functional pattern which combines elements of an array two at a time until only one value remains. As recently surveyed by Zou et al. [59], the MapReduce programming template is at the very core of modern computationally intensive bioinformatics applications. This third illustration demonstrates the MapReduce pattern as an extension of the second example by subsequently summing the results of the distributed “map” using a “reduce” also distributes across QM’s volunteers:

(4) Real-world genomic analysis

The fourth illustrative example assesses QM’s ability to scale the asynchronous operations demonstrated above for use in a real-world bioinformatics workflow. The example is a Fractal MapReduce decomposition of sequence alignment [36] which distributes both the processing and networking loads across QM’s volunteers, as described in Figure 1. It also demonstrates that libraries of any complexity or elaboration can be distributed to the volunteers along with the commands that invoke those libraries. Specifically, both the data and the library encoding for the sequence analysis procedure are invoked by QM but entirely resolved and executed by the volunteer browsers. It also illustrates the ability of a volunteer node to call code and data from multiple sources which are independently developed and maintained.

Consider, as in the first example, that we have some x that we wish to transform via some function f, so that x is now an array of URLs that reference FASTA files hosted by NCBI:

We want to perform a particular sequence analysis on each FASTA file, namely a Fractal MapReduce decomposition of the Chaos Game Representation [36]. Thus, we define a function f for use with the QM.map method that will take a URL as input and return the results of the sequence analysis as output:

There is a key challenge, however, in that our function f depends on a usm function that exists only after an external library has been loaded. Therefore, to specify the task completely, we will need either to include usm as part of f or else to pass a reference to the library in the form of a URL. We chose the latter strategy in this case so that the library can be downloaded in parallel by each volunteer simultaneously without burdening the API server. Each external function may have multiple dependencies, and thus QM.map accepts an optional env parameter so that the dependencies for each external function can be specified as an array of URLs to be loaded sequentially:

Finally, we will specify the box parameter explicitly for demonstration purposes. The box parameter takes the place of an API key and allows volunteers to execute tasks in a particular queue. This mechanism allows submitters to direct tasks into different queues and further enables the use of abstractions like MapReduce:

Putting these definitions together, we now launch twenty individual genomic sequence analyses for simultaneous execution via

A full version of these examples can be found online at http://q.cgr.googlecode.com/hg/index.html. The version there includes the full URLs to all twenty Streptococcus pneumoniae genomes and also to the versioned libraries specified by env. An accompanying screencast for these examples is also provided in that page.

Usage statistics

The dissemination of browser-based tools in social coding environments like GitHub [19] is characterized by the same expansive dynamics as social media. For example, although this is our first report describing it, QM can be – and has been – discovered by the community at large. During the 12 months period beginning in April 2013, QM received more than 2.2 million API calls from 2,100 IP addresses in 87 countries to more than 1,800 QM “boxes” (the code and results exchange domains defined by token), with 98 boxes receiving more than 1,000 calls each and 16 boxes receiving 10,000 calls or more. The statistics of QM usage are described in Figure 3, and the geographic distribution of its users is described in Figure 4. It is unclear exactly how much of QM’s usage is associated with the distributed computational genomics web apps that motivated its development, but the wide geographic distribution of its users suggests an appeal driven by a more general interest in distributed computing. This interpretation is reinforced by unsolicited reports about QM in HPC media such as HPCwire (article at http://goo.gl/9H5W03) and insideHPC (http://goo.gl/bDkJZL). Finally, as noted in Methods, all of the server- and client-side software are open-source and permissively licensed. The browser client requires nothing more than script tag loading to be included in a web app, and the server is just as accessible through NPM [40]. It is therefore conceivable that other QM deployments are in use at other addresses, perhaps even within the firewall of medical centers, as was the specific intention of QM’s development.

The server load associated with orchestrating this initial heavy use of QM is very modest because of the reliance on code distribution rather than on code execution. In fact, the deployment supporting the usage statistics described above (the server behind https://api.qmachine.org) was never overwhelmed by traffic spikes even though it was running on a shared-tenant virtual machine with just 512 MB of RAM, 2 × 512 MB MongoDB databases, and no hard drive. Furthermore, the authors do not incur any maintenance costs for the public tool dissemination, either from GitHub or from NPM’s package repository. We are therefore committed not to collect any data beyond the broad statistics described in Figures 3 and 4 for the reference deployment discussed here. Particularly relevant for the biomedical usage scenario that motivated this work, we are also committed not to collect any data at all from private deployments of QM; in other words, no part of QM’s software ever sends data back to our server(s) from other deployments. This design allows administrators to deploy their own QM servers through NPM and fully configure their own security as needed for clinical and/or biomedical research usage. These assurances can, of course, be verified through inspection of QM’s source code.

MapReduce

Many researchers with access to large-scale computational resources still find those resources inaccessible because “everyday” workflows often require more than just fast computers – they require programming skills that are harder to acquire. Bioinformatics workflows increasingly rely on MapReduce as an abstraction, but available MapReduce resources still expose researchers to programming environments with strict procedural requirements and steep learning curves. QM is much simpler to set up and operate than Apache Hadoop [61], for example. It allows users to run MapReduce jobs on multiple physical machines and to crowdsource elastic computing resources with the simplicity of writing and loading a webpage – skills performed every day by millions of people worldwide. We argue that using the web computing architecture explored by QM – that is, without installing a dedicated application – is a natural evolution of current cloud-based MapReduce services, just as Hadoop was a step up from one-off compile-and-run workflows.

Distributed computing

QM’s web service provides a message passing interface for distributed computing. This statement may at first sound paradoxical, but JS’s single-threaded programming model does not limit JS programs to single-threaded execution; external execution contexts can be used to support concurrency via event-driven programming. QM leverages browsers’ asynchronous (non-blocking) network communications layers to connect multiple machines’ execution contexts, but browsers that support Web Workers [62] can execute concurrent programs within the same physical machine.

Cloud browsers

An interesting new twist in the development of web computing architectures is the emergence of the “cloud browser” [63]. In these systems, a mobile browser behaves as a thin client when a webpage’s scripts demand heavy computation. Cloud browsers therefore demonstrate browser scaling in the vertical direction, whereas QM demonstrates browser scaling in the horizontal direction. Because QM makes no assumptions about its volunteers’ underlying resources, cloud browsers can volunteer for QM alongside ordinary browsers without loss of generality. In other words, cloud browsers represent enhancements of present-day browsers, while QM presents a solution for HPC that advances the underlying architecture of the Web towards that of a Global Computer [64, 65].

Biomedical applications

In clinical environments, it can be difficult or even impossible to distribute workflows due to privacy concerns that prevent sensitive data from leaving the hospital environment, where conventional HPC is typically absent. QM satisfies this preoccupation without requiring additional resources. As shown in Figure 5, the median computational power of the Top500 HPC in November 2013 (http://goo.gl/XIUIDP) was roughly 2,600 times faster than our lab’s standard-issue desktop machine. This is a much smaller factor than the number of machines in a typical medical center. Thus, even if restricted to a single hospital environment, volunteer computing can still rival the total capacity of very substantial HPC resources.

QM can also be used to power workflows inside of a single workstation. In such a scenario, the workstation would run QM’s API server locally and use multiple browser tabs to execute the workflow in parallel. Such a workflow might also incorporate existing bioinformatics tools such as the Basic Local Alignment Search Tool (BLAST) [66] by using traditional server-side scripting languages like Perl [67] or Python [68] to connect to QM’s API or even directly to the persistent storage layer.

Security

The security of workflows that use QM is handled orthogonally to QM by the selection of volunteers and by access control to code and data. A number of considerations should nevertheless be made to assist in the configuration of its distributed operation. It is important to recall that the web browser executes JS within a sandboxed environment, which, among other protections, prevents programmatic access to the filesystem of the volunteer machine. As a result, QM’s security is configured around two firewalls.

The first and most basic protection is associated with the uniqueness of the “box” (token) issued by the submitter, which should be shared only with trusted volunteers. An additional layer of security can be added through the use of open authentication such as OAuth 2.0 [69] to verify that only trusted volunteers are involved. This second layer of protection is particularly useful in creating audit trails. These two mechanisms can be combined in many ways, as appropriate for a particular workflow. For example, different steps of a workflow could be assigned to distinct cohorts of volunteers depending on the sensitivity of the code and data and/or the trustworthiness of the volunteers. The resulting granularity could also be used to build redundancy – and therefore robustness – into the distributed QM operation.

In short, the weakest link in QM’s architecture – and where the opportunities for abuse lie – derive from the sharing of the “box” by members of a group of volunteers. In this regard, the key feature of QM’s design is that the abuse can target the submitters but not the volunteers, because QM’s operations take place within the sandbox of the web browser.

Source code

All source code for this paper is version-controlled and open-sourced. The primary source for QMachine’s code is located in a Git [70] repository at https://github.com/wilkinson/qmachine. The code and data for the illustrative examples shown in the Results section are available in a Mercurial [71] repository at http://q.cgr.googlecode.com/hg/. Quanah’s source code repository is available at https://github.com/wilkinson/quanah, and the USM repository is available at https://github.com/usm/usm.github.com.