=Paper=
{{Paper
|id=Vol-2203/161
|storemode=property
|title=Framework for Distributed Computing on the Web
|pdfUrl=https://ceur-ws.org/Vol-2203/161.pdf
|volume=Vol-2203
|authors=Jakub Siller,Jaroslav Kuchar
|dblpUrl=https://dblp.org/rec/conf/itat/SillerK18
}}
==Framework for Distributed Computing on the Web==
https://ceur-ws.org/Vol-2203/161.pdf
S. Krajči (ed.): ITAT 2018 Proceedings, pp. 161–167
CEUR Workshop Proceedings Vol. 2203, ISSN 1613-0073, c 2018 Jakub Šiller and Jaroslav Kuchař
Framework for Distributed Computing on the Web
Jakub Šiller and Jaroslav Kuchař
Web Intelligence Research Group, Faculty of Information Technology
Czech Technical University, Thákurova 9, 160 00, Prague 6, Czech Republic
{sillejak|jaroslav.kuchar}@fit.cvut.cz
Abstract: This work is a brief summary of a master the- were focused on various types of computing tasks. For
sis that focuses on design and implementation of a frame- instance projects [6], [7], [8] and [9] were using web
work that uses computers of website visitors as comput- browsers for simulated evolution. Work [10] was focused
ing nodes through web browsers. It contains an analysis on image processing. Machine learning in web browsers
of the Web environment, summarization of previous ap- was the topic of [11]. Authors of [12] created web search
proaches and projects, design and implementation of the engine using web browsers. In [13] and [14] map reduce
framework. The work describes the solution of computing frameworks were presented. There were also works that
node failure, reaction to slow computing node, possibili- were focused on creating general framework: [15], [16],
ties of controlling the load of the framework on a website [17]. Architecture of most of presented works was client-
visitor’s computer, strategies for work distribution and se- server.
curity of the framework. At the end of the work, the ex- Developers of commercial project Computes are trying
periment results and proposal of improvements are listed. to create distributed decentralized supercomputer from all
kind of devices.
There are also several commercial tools for mining alt-
1 Introduction coins.
Published works are usually just proof of concept and
Nowadays, web technologies and web services are an in-
authors don’t deal with every aspect of systems. For in-
separable part of human life. We are using web browser
stance security is often omitted. The main contribution of
for communication, entertainment, shopping and a many
this work is to create complex framework that could be
other activities. A lot of web users have powerful de-
deployed.
vices but they don’t use their full power often. Thus, in
the world, there is huge computing potential that is idle.
Framework presented in this work is able to utilize this 3 Analysis
computing power.
In section 2 there is a summarization of previous works The web has several major characteristics. One of them is
and current project in the field of distributed computation diversity. Web users are using various browsers of various
in web browsers. Analysis of web environment and frame- versions. Each user has different device, different connec-
work requirements are content of section 3. Section 4 is tion speed etc.
focused on design of the framework. Finally, experiment Another strong aspect of the Web is dynamics. Web
results are presented in section 5. technologies are still evolving and continuously new tech-
nologies are emerging. Also web browsing is very dy-
namic. According to [18] users often stay on a page just
2 Related works for 10-20 seconds.
Utilizing idle power via the Internet to create distributed The Web is free and open. Everyone can join and pub-
computer is not a new idea. According to [1] the first lish and consume data.
project about this topic was Great Internet Mersenne And the last but not least aspect is its’ enormous size.
Prime Search [2] that started in 1996. A few years later [19] and [20] states that the Web has more than 3.7 billions
projects SETI@Home [3] that was focused on the Search users and this number is getting bigger every year.
for Extraterrestrial Intelligence and Folding@home [4]
that was using computers of volunteers for medical re- These aspects of the Web have impact on the frame-
search was launched. After that BOINC (Berkeley Open work requirements. Dynamic browsing will cause fre-
Infrastructure for Network Computing) [5] was created. It quent computing node failures. The framework have to be
is probably the biggest and the best known platform for able to detect failure and solve the situation. The frame-
volunteer computing. All of mentioned projects are still work also should be able to work with computing nodes
alive. In order to join a computation in these projects a of different performance. Security mechanisms should be
volunteer have to install some additional software. involved on server side and on client side as well. Frame-
Since 2007 several works related to utilizing computa- work also have to ensure correctness and reliability of
tion power using web browser have been published. They tasks results.
162 Jakub Šiller and Jaroslav Kuchař
4 Design The framework utilizes replication and majority voting in
order to ensure correctness and reliability of works results.
4.1 Computational model User can for each task specify the replication factor. In
The framework is focused on types of tasks in which server each work object server holds information that indicates
sends work to client, client processes the work and sends to how many more clients the work should be distributed
the result back to the server. There is no communication in order to reach the replication factor. This information is
between clients and no communication between client and in attribute remaining.
server regarding the computation but distributing works
and receiving results. Framework user can define divid- In order to compute a task a user sends data and id of
ing and merging functions. In case that a task has too big task prototype to the server. Server insert the task to the
data the framework will use dividing function in order to task queue. The server holds collection of works that are
automatically recursively divide the data and create partial currently being distributed to clients. If there is enough
subtasks called works. Works are then distributed to the place for more works server prepare another task from
client. Results of works from clients are then merged by the queue for distribution. Preparation of a task for dis-
framework using merging function. Described model is tribution consists mainly of dividing task’s data. When a
shown in figure 1. client sends work request the server response with several
In order to compute a tasks there will be a lot of mes- works. Works that are returned to the client have attribute
sages between clients and the server. Speed of commu- remaining greater then zero. After assigning a work to a
nication depends on connection quality. It might be time client the attribute is decremented. Number of works that
consuming. Therefore, the framework is more suitable for are returned to the client depends on strategy that is used.
computing intensive tasks rather than data intensive tasks. Strategies implemented in the framework are described in
following section. When a client receives works it starts to
4.2 Users process them. The client doesn’t wait until all works are
Users of the framework are divided into two groups. Users done but each result is sent back to the server as soon as it
from the first group are creators of task prototypes. A task is available. This is shown in figure 3.
prototype consists of definition of code that should be ex- When the server received results from all works it merge
ecuted in client and optional dividing and merging func- them to the result of the task. The task’s result is stored in
tion. This group of users should know the framework - database. When the user send request for the result to the
its’ advantages, disadvantages and some technical details server it responds with the result from the database.
in order to create effective code for the framework. When a client processes all assigned works it sends new
The second group are common users. They use frame- work request to the server.
work for computations. They don’t need to know anything
about the framework but the id of task prototype they want Strategies for work distribution There are several strate-
to use. gies in the framework for determining number of works
This division has several reasons. The first of them is that should be returned to a client.
security. We can assume that number of task prototypes The most easy one is a strategy that returns fixed num-
creators will be much smaller than number of other users. ber of works.
Therefore it might be relatively easy to make sure that cre- Another strategy is based on fixed sum of works sizes.
ators of task prototypes are trustworthy. And then we can The framework uses statistics to estimate maximum num-
assume that their code is probably trustworthy as well. An- ber of works so that sum of their sizes is less then config-
other reason for the division is quality of code. If there are ured threshold.
users who are focusing on creating code for the frame- The most complex strategy uses time elapsed from start
work there is high probability that the code will be effi- of a client session. According to [18] a distribution of ses-
cient, without bugs and there will be no task prototypes sion duration follows Weibull distribution with negative
for tasks that are not suitable for the framework. More- aging. That means that at the begging of a web page visit
over common users don’t need to know anything about the the probability that user will leave the page is very high
framework. and it decrease over time. Therefore this strategy increase
number of works that are sent to the client accordingly to
4.3 Framework in a nutshell the session duration.
Basic architecture of the framework is client-server. The
server consists of two main parts - ProgrammerServer Modes of work distribution Work distribution can run
and VolunteerServer. ProgrammerServer is responsible in two modes. In the first mode data with full work code
for communication with users of the framework. Volun- are sent to the client. In the second mode data along code
teerServer is the main part of the framework. It is respon- identifier are sent. In this mode client have to make an-
sible for communication with clients, processing tasks, other request to the server in order to get the code. How-
distributing works to clients and processing results. ever, in this mode it is possible to cache the code in the
Framework for Distributed Computing on the Web 163
(b) Receiving and merging phase
(a) Dividing and distributing phase
Figure 1: Visual example of the computation model of the framework
browser and in intermediate network devices. Therefore
the amount of data on wires can be decreased. This mode
is efficient only if a client or a group of clients are per-
forming tasks of a single or a few task prototypes. In other
cases it might be inefficient because of more requests.
4.4 Computing node failure
In order to detect computing node failure the framework
holds in each session object time of last access. It is up-
dated with each request from the session. During compu-
tation a client code running in a session periodically sends
empty request to the server - a heartbeat - so the server is
informed the session is still alive as is shown in figure 3.
Figure 2: Base architecture of the framework. A server module DeadSessionCollector periodically
collects dead sessions. A session is considered to be dead
if time elapsed from the last access is higher than a con-
figured threshold. When a session die the framework in-
crements attribute remaining of all unprocessed works that
were assigned to the session. So the work will be assigned
again to some session.
4.5 Slow computing node
Because of different performance of web user devices, dif-
ferent connection speed and different utilization of the de-
vice by it’s user it may happen that computation of a work
would last on one client much longer than computation of
the same work on other client. Thus, when some client is
processing some work too long it may be efficient to assign
Figure 3: Simplified process of distributing works to the
the work again to another client.
client and subsequent sending results back to the server.
The framework’s module LongRunningSessionCollec-
During computation of a work heartbeats are sent to the
tor periodically checks whether there is a slow client for
server.
some work. A client is considered too slow for a work if
164 Jakub Šiller and Jaroslav Kuchař
the time elapsed since the work was assigned to the client 4.8 Security
is several times longer than average time of computation
of the work on other clients. If a client is marked as slow The framework contains several security mechanisms. The
for a work the framework increment attribute remaining of server accepts user request only with valid API keys. List
the work. of API keys is in configuration of the framework. If a re-
quest doesn’t contain valid API key the server responds
with HTTP code 403 forbidden.
4.6 Client side The framework executes user code on server-side and
also on client-side. The framework must ensure that the
At client side the framework is using Web Worker technol- code will cause no harm to server or client’s device. There-
ogy [21]. A web browser create new thread for each Web fore, user code is executed in a sandbox. At the server-side
Worker object. Therefore experience of browsing a web VM2 module [25] is used. At the client-side a sandbox is
page should not be affected by the framework. The frame- implemented as a white-list of allowed function and ob-
work creates several Web Worker objects in which works jects. Every other function or object is disallowed.
are processed. When a client receive works it put them in Communication between a client and the server is en-
a queue. The Web Worker objects are taking objects from crypted. This decrease probability that content of mes-
the queue and processing them. As soon as a Web Worker sages is changed by malicious third-party.
object computes a work the result is sent to the server. This As was mentioned earlier in order to ensure correctness
is shown in figure 4. an reliability of results majority voting is involved. A user
can specify replication factor but he or she should be aware
that probability that data are correct is never 100%.
Securing data is complicated topic. The purpose of web
browsers is to serve data to its user so the framework can
not hide data from the user of the browser it is running in.
Therefore the only way how to secure the data is comput-
ing on encrypted data. There are mathematical models that
enables it. Some of them are described in [26]. There is
no need to change the framework in order to compute on
encrypted data - it is responsibility of task prototypes im-
plementation. Computing on encrypted data can also en-
sure 100% probability that result is correct excluding bugs
Figure 4: Illustration of processing works assigned to a
in the task prototype.
client on multiple Web Worker objects
For computation of a work the framework is able to 5 Experiments
use new client-side web technologies asm.js [22] and We-
bAsembly [23] instead of JavaScript [24]. The creator of
Experiments were performed on 60 computers in class-
task prototype can implement working function in C++
rooms at FIT CTU. Configuration of computers is in table
which is then at the server compiled into asm.js and We-
1. Test task was naive algorithm for computing determi-
bAssembly modules.
nant of a matrix. Maximum size of a matrix that was sent
to clients was 11 × 11. If the matrix was bigger it was
recursively divided to matrices of size 11 × 11.
4.7 Controlling of client utilization
Classroom CPU Model size of RAM
There are two ways how the framework is able to control
Intel® Core™ i5-6500
stressing of client’s device. The first one is controlling the T9-350 16 GB
CPU @ 3.20GHz
number of Web Workers that are used. The second option
Intel® Core™ i5-3470
is throttling. JavaScript code can not control stressing of T9-351 8 GB
CPU @ 3.20GHz
CPU at particular time but it is possible to control stressing
Intel® Core™ i5-3470
from long-term view. For instance if x seconds is processor T9-303 8 GB
CPU @ 3.20GHz
fully stressed by the framework and then x seconds is idle
we can say that the framework stressed CPU by 50% dur- Intel® Core™ i5-4570S
T9-349 8 GB
ing 2x seconds. The configuration attribute throttleFactor CPU @ 2.90GHz
holds information how many times the duration of the last Intel® Core™i5-2410M
server 6 GB
computation the server has to wait before it could again as- CPU @ 2.30GHz
sign a work to a client. In other words it is inverted value
to the desired CPU usage. Table 1: Configuration of test computers
Framework for Distributed Computing on the Web 165
Tests were performed in the mode in which just iden-
tifier of a code is sent to the client. Replication factor
was set to 3. The framework was using all CPU cores
of the client devices. Test environment refreshed test web
page at a client after randomly chosen time from interval
[0, maxDuration] where maxDuration is parameter. Test
Settings of earlier mentioned parameters are in table 2.
Parameter Value
deadSessionCollector:
interval 2 sec
deadTime 10 sec Figure 6: Influence of number of clients to average number
LongRunningSessionCollector: of sessions to which one work was assigned
interval 5 sec
factor 5
Other framework settings:
throttleFactor 0
heartBeatInterval 5 sec
test environment settings:
maxDuration 960 sec
Table 2: Test setting of mentioned parameters
The purpose of the first experiment was to test how
would change computation time of the task from user point
of view with increasing number of clients. In figure 5 we
Figure 7: Statistics of a computation time of one matrix at
can see that with increasing number of clients computa-
client-side. For each number of client on axis X mean and
tion time is decreasing. During tests in classrooms T9-
mean ± standard deviation of computation time at client-
350, T9-351 and T9-303 speeding up stops around number
side is shown
30. That probably happened because computers in T9-350
were more powerful that others (shown in figure 7) so the
framework considered the others to be slow and started to implemented in C++ using OpenMP [27] in order to uti-
assign one work to more clients that was necessary. This lize all CPU cores of a computer. Local computation was
is shown in figure 6. Also there was a mistake in the con- executed on the server’s computer. In figure 8 we can see
figuration. Interval of sending heartbeat was equal to the that for small matrices the local computation was more ef-
time after which a session was considered as dead. After ficient but for big matrices it was more efficient to use the
some changes in configuration the computation becomes framework.
again a little bit faster.
Figure 8: Influence of matrix size to computation time
Figure 5: Influence of number of clients to computation from user point of view.
time from user point of view.
The third experiment tests influence of throttleFactor on
The second experiment tests how size of a matrix would the computation time from user point of view. In figure
affect computation time from user point of view. It also 9 we can see that there is linear dependency. This test
compares computation time of the task using framework prove that influence of throttleFactor is expected and pre-
and computation time using local computation that was dictable.
166 Jakub Šiller and Jaroslav Kuchař
A web page could profit from involving it’s visitors to the
framework. So the web page could offer a discount of their
services to visitors that allow joining to the framework.
Another use case could be in a company that have a lot
of computers for it’s employees. In this case a proxy could
inject web pages with framework’s client-side code and so
create company’s big distributed computer with minimum
additional costs.
8 Conclusion
Figure 9: Influence of attribute throttleFactor to computa- In this work a framework for distributed computation us-
tion time from user point of view. ing web browsers is presented. It contains mechanisms
that solve computing node failure and reaction to a slow
computing node. It describes strategies for distributing
The last experiments tests how computation time from works within clients and modes in which the distribution
user point of view is changing when maximal session du- can be done. The work deals with controlling of client’s
ration is changing. In figure 10 we can see that with in- device stressing. Security mechanisms are described as
creasing session duration computation time is decreasing well. Experiment results that are presented have shown
as it was expected. that the framework is useful and deployable. The future
of the framework may be a computational platform with
market of task prototypes and computation power.
Acknowledgments
This research was supported by Faculty of Informatics,
Czech Technical University in Prague.
References
[1] Miller, D. M.: The Online Community Grid Volunteer
Figure 10: Influence of maximal session duration (at- Grid Computing with the Web Browser. Georgia Institute
tribute maxDuration of test environment) to computation of Technology, 2008. Available on: https://smartech.
time from user point of view. gatech.edu/handle/1853/33477
[2] Mersenne Research, Inc.: Great Internet Mersenne Prime
Search. [cit. 30.4.2018]. Available on: https://www.
mersenne.org
6 Future work [3] University of California : About SETI@home. [cit.
30.4.2018]. Available on: https://setiathome.
berkeley.edu/sah_about.php
Experiments have shown that framework is useful and it
may be deployed. However there are still some limits and [4] Pande Lab: Folding@home. [cit. 30.4.2018]. Available on:
http://folding.stanford.edu
drawbacks of the implementation that should be solved.
For instance database queries can be optimized or new [5] Anderson, D. P.: BOINC: a system for public-resource
computing and storage. In Fifth IEEE/ACM International
strategies for distributing works can be implemented.
Workshop on Grid Computing, Nov 2004, ISSN 1550-
The framework can be also extended in several ways. 5510, s. 4–10, doi:10.1109/GRID.2004.14.
For instance GUI and monitoring extension would be very
[6] Merelo, J. J.; García, A. M.; Laredo, J. L. J.; aj.: Browser-
practical. There might be client, user and web sites ad- based Distributed Evolutionary Computation: Performance
ministration. Extensions of the framework may lead to and Scaling Behavior. In Proceedings of the 9th An-
computing platform with market that would act similar as nual Conference Companion on Genetic and Evolution-
application markets. ary Computation, GECCO ’07, New York, NY, USA:
ACM, 2007, ISBN 978-1-59593-698-1, s. 2851–2858, doi:
10.1145/1274000.1274083. Available on: http://doi.
7 Motivation for joining computations acm.org/10.1145/1274000.1274083
[7] Klein, J.; Spector, L.: Unwitting Distributed Genetic Pro-
There can be several reasons for a web user to join the gramming via Asynchronous JavaScript and XML. In Pro-
framework but probably the most likely situation is this: ceedings of the 9th Annual Conference on Genetic and
Framework for Distributed Computing on the Web 167
Evolutionary Computation, GECCO ’07, New York, NY, [19] Liedke, L.: 100+ Internet Stats and Facts
USA: ACM, 2007, ISBN 978-1-59593-697-4, s. 1628– for 2018. [cit. 30.4.2018]. Available on:
1635, doi:10.1145/1276958.1277282. Available on: http: https://www.websitehostingrating.com/
//doi.acm.org/10.1145/1276958.1277282 internet-statistics-facts-2018/
[8] Merelo-Guervos, J. J.; Castillo, P. A.; Laredo, J. L. J.; [20] Stevens, J.: Internet Stats & Facts for 2017. [cit.
aj.: Asynchronous distributed genetic algorithms with 30.4.2018]. Available on: https://hostingfacts.com/
Javascript and JSON. In 2008 IEEE Congress on Evolution- internet-facts-stats-2016/
ary Computation (IEEE World Congress on Computational [21] Mozilla and individual contributors: MDN web docs:
Intelligence), June 2008, ISSN 1089-778X, s. 1372–1379, Using Web Workers. [cit. 30.4.2018]. Available on:
doi:10.1109/CEC.2008.4630973. https://developer.mozilla.org/en-US/docs/
[9] Duda, J.; Dlubacz, W.: Distributed Evolutionary Web/API/Web_Workers_API/Using_web_workers
Computing System Based on Web Browsers with [22] Herman, D.; Wagner, L.; Zakai, A.: asm.js: Working
Javascript. In Proceedings of the 11th International Draft. [cit. 30.4.2018]. Available on: http://asmjs.
Conference on Applied Parallel and Scientific Com- org/spec/latest/
puting, PARA’12, Berlin, Heidelberg: Springer-Verlag, [23] WebAssembly. [cit. 30.4.2018]. Available on: http://
2013, ISBN 978-3-642-36802-8, s. 183–191, doi:10.1007/ webassembly.org/
978-3-642-36803-5_13. Available on: http://dx.doi.
[24] Mozilla and individual contributors: MDN web docs:
org/10.1007/978-3-642-36803-5_13
JavaScript reference. [cit. 30.4.2018]. Available on:
[10] Zorrilla, M.; Martin, A.; Tamayo, I.; aj.: Web Browser- https://developer.mozilla.org/en-US/docs/
Based Social Distributed Computing Platform Applied Web/JavaScript/Reference
to Image Analysis. In 2013 International Conference on
[25] Šimek, P.: VM2. [cit. 30.4.2018]. Available on: https:
Cloud and Green Computing, Sept 2013, s. 389–396, doi:
//www.npmjs.com/package/vm2
10.1109/CGC.2013.68.
[26] Vaikuntanathan, V.: How to Compute on Encrypted Data.
[11] Meeds, E.; Hendriks, R.; al Faraby, S.; aj.: ML-
In Progress in Cryptology - INDOCRYPT 2012, editace
itB: Machine Learning in the Browser. CoRR, ročník
S. Galbraith; M. Nandi, Berlin, Heidelberg: Springer
abs/1412.2432, 2014, 1412.2432. Available on: http:
Berlin Heidelberg, 2012, ISBN 978-3-642-34931-7, s. 1–
//arxiv.org/abs/1412.2432
15.
[12] Turek, W.; Nawarecki, E.; Dobrowolski, G.; aj.: WEB
[27] OpenMP ARB: OpenMP. [cit. 30.4.2018]. Available on:
PAGES CONTENT ANALYSIS USING BROWSER-
http://www.openmp.org/
BASED VOLUNTEER COMPUTING. Computer Science,
ročník 14, č. 2, 2013: str. 215, ISSN 2300-7036. Available
on: https://journals.agh.edu.pl/csci/article/
view/278
[13] Pan, Y.; White, J.; Sun, Y.; aj.: Gray Computing: An Anal-
ysis of Computing with Background JavaScript Tasks. In
2015 IEEE/ACM 37th IEEE International Conference on
Software Engineering, ročník 1, May 2015, ISSN 0270-
5257, s. 167–177, doi:10.1109/ICSE.2015.38.
[14] Ryza, S.; Wall, T.: MRJS: A JavaScript MapRe-
duce Framework for Web Browsers. 2010. Avail-
able on: http://static.cs.brown.edu/courses/
csci2950-u/f11/papers/mrjs.pdf
[15] Cushing, R.; Putra, G. H. H.; Koulouzis, S.; aj.: Distributed
Computing on an Ensemble of Browsers. IEEE Internet
Computing, ročník 17, č. 5, Sept 2013: s. 54–61, ISSN
1089-7801, doi:10.1109/MIC.2013.3.
[16] Wilkinson, S. R.; Almeida, J. S.: QMachine: commod-
ity supercomputing in web browsers. BMC Bioinformatics,
ročník 15, č. 1, Jun 2014: str. 176, ISSN 1471-2105, doi:
10.1186/1471-2105-15-176. Available on: https://doi.
org/10.1186/1471-2105-15-176
[17] Fabisiak, T.; Danilecki, A.: Browser-based Harness-
ing of Voluntary Computational Power. ročník 42,
03 2017. Available on: https://www.degruyter.
com/downloadpdf/j/fcds.2017.42.issue-1/
fcds-2017-0001/fcds-2017-0001.pdf
[18] Nielsen, J.: How Long Do Users Stay on
Web Pages? [cit. 30.4.2018]. Available
on: https://www.nngroup.com/articles/
how-long-do-users-stay-on-web-pages/