%\VignetteIndexEntry{RelatedWork}
%\VignetteKeywords{Web services}
%\VignetteKeywords{Grid services}
%\VignettePackage{RWebServices}

\documentclass[]{article}

%\usepackage{Sweave}
\usepackage[colorlinks,linkcolor=blue,pagecolor=blue,urlcolor=blue,citecolor=blue]{hyperref}
\usepackage{graphicx}

\newcommand{\lang}[1]{{\texttt{#1}}}
\newcommand{\pkg}[1]{{\textit{#1}}}
\newcommand{\code}[1]{\texttt{#1}}
\newcommand{\func}[1]{{\texttt{#1}}}
\newcommand{\method}[1]{{\texttt{#1}}}
\renewcommand{\arg}[1]{{\texttt{#1}}}
\newcommand{\ret}[1]{{\texttt{#1}}}
\newcommand{\obj}[1]{{\texttt{#1}}}
\newcommand{\class}[1]{{\textit{#1}}}

\newcommand{\R}{\lang{R}}
\newcommand{\Java}{\lang{Java}}

\newcommand{\RWebServices}{\pkg{RWebServices}}
\newcommand{\TypeInfo}{\pkg{TypeInfo}}
\newcommand{\SJava}{\pkg{SJava}}

%% \newcommand{\STS}{\func{SimultaneousTypeSpecification}}
%% \newcommand{\ITS}{\func{IndependentTypeSpecification}}
%% \newcommand{\TypedSignature}{\func{TypedSignature}}
%% \newcommand{\STT}{\func{StrictIsTypeTest}}
%% \newcommand{\DTT}{\func{DynamicTypeTest}}
%% \newcommand{\ITT}{\func{InheritsTypeTest}}

\begin{document}

\title{Approaches to \R{} as web and analytic service}
\author{
  Nianhua Li, Martin T. Morgan, Seth Falcon, Robert Gentleman \\[\bigskipamount]
  Fred Hutchinson Cancer Research Center \\
  1100 Fairview Ave.\ N. \\
  PO Box 19024 Seattle, WA 98109
}
\date{15 Sept, 2006}
\maketitle

\begin{abstract}
  The aim of \RWebServices{} is to facilitate exposing \R{} functions
  as effective \Java-based web or grid services. The application uses
  existing technologies in web and grid services, XML, as well as new
  and existing inter-system interfaces between \R{} and other
  languages. We briefly review the state of art of these research
  areas here, with an emphasis on applications related to
  \RWebServices{}.
\end{abstract}

\section{Introduction}

\emph{Web services}~\cite{W3C2004, Cerami2002} refer to a set of
standards and techniques that allow applications to be described,
published, discovered and invoked over an internet protocol backbone.
Web services enable applications to communicate in a platform and
language independent way, increasing software sharing and
reuse. Multiple programming languages have been used in web services,
such as \Java{}~\cite{Sun2006, Irani2002}, \lang{Perl}~\cite{Ray2002},
\lang{C++}, \lang{C\#}, \lang{Python}, \lang{PHP}, \lang{VB}, and
\lang{Ruby}.

A \emph{grid service} is a web service specializing on a computing
environment that integrates distributed and heterogeneous resources
across various administrative organizations~\cite{Foster2001,
Foster2002}.  An important distinction between a grid service and a
general web service is the support of stateful resources, i.e.,
services can be created and destroyed during run-time. The high-level
definitions of grid service architecture and its requirements are
given by the Open Grid Services Architecture (OGSA)~\cite{Foster2005,
Treadwell2005}.  The Web Services Resource Framework
(WSRF)~\cite{Czajkowski2004} provides detailed specifications for
stateful resource management using web services technology.  There are
implementations of OGSA and WSRF in various languages. Example
implementations include the Globus Toolkit in \lang{Java} and
\lang{C++}~\cite{FosterG2005}, WSRF::Lite in
\lang{Perl}~\cite{wsrflite2005}, WSRF.NET in Microsoft \lang{.NET}
framework~\cite{wsrfnet2005}, and pyGlobus in
\lang{Python}~\cite{pyglobus2006}.
  
Web and grid service support for \R{} is still in its early stages.
An effective way to connect \R{} with web or grid services is to
leverage existing efforts in other language domains, in particular
using \Java{} as a conduit to provide \R{} web and grid services. We
chose \Java{} rather than other languages because web and grid
services support is well-established in the \Java{} community, and
because communication with third-party \Java{} components is a feature
of the \emph{caBIG} project.

This document discusses systems integrating \R{} and \Java{}, and
provides a brief review of existing web and grid services support in
the \R{} community. The document concludes with a brief synopsis of
formats for statistical data exchange.

\section{Bridging \R{} and \Java{}}

Two different approaches have been used to achieve interoperability
between \R{} and \Java{}: message streams and shared libraries.

\subsection{The message stream approach} 

In this approach, \R{} and \Java{} are run as separate processes. Data
is transported as message streams with an encoding that can be
understood by both parties. Message streams can be text-based or
binary.  Transportation media can be socket connections, web services
over the internet, or even file systems (standard I/O).  Open
Statistical Services (\pkg{OSS})~\cite{mcconnell2005} and
\pkg{Rserve}~\cite{urbanek2003} are two examples of the message
streams approach.

This approach is useful if the \R{} component may potentially talk
with other language systems in the future. New language systems can be
easily included, since the only requirement is serialization and
deserialization between the message encoding and language specific
data structures. Moreover, it is known that \R{} is not thread-safe,
whereas \Java{} supports multi-threading.  Because \R{} and \Java{}
processes are independent from each other in this approach, it is
possible to have multiple \Java{} threads communicating with \R{}
concurrently.

There are several disadvantages to a message streaming approach. First
is efficiency: data is serialized and deserialized at every stage
during transportation. Second, stream-based transportation is slower
than direct programming calls, no matter what encoding or
transportation mechanism is used. The third concern is related to data
persistence and error handling.  Because \R{} and \Java{} belong to
separate processes, extra implementation is required to ensure the
ability to deliver error messages even after after a server or data
transportation failure. Security is also a challenge in server-client
architectures. The final challenge is session management: imagine that
a \Java{} process invokes an \R{} function which uses objects of the
caller (the \Java{} process). Both parties need server support, and
the transportation layer needs to support session management.

Open Statistical Services (\pkg{OSS}) integrates \Java{} and \R{} via
\lang{StatML}, an XML-based markup language for encoding statistical
data. \pkg{OSS} contains two major components: the \emph{Java bridge}
and the \emph{R bridge}. Both components support data serializations
between \lang{StatML} and data structures in \R{} (the R bridge)
and \Java{} (the Java bridge).  The message transportation
channel can be either standard I/O, or internet-based web services. A
message of \R{} method invocation from \Java{} includes two parts: the
function name as a character string, and input parameters in
\lang{StatML} encoding. Basic \R{} data types including numeric,
integer, logical, and character vectors, as well as list and array,
are covered by the \lang{StatML} encoding. The \lang{StatML} encoding
is text-based for scalars, and 64-bit binary for lists and arrays.
Invocation of \Java{} methods from \R{} is performed in a similar
manner.

\pkg{Rserve} is a TCP/IP server that allows \R{} to be invoked from
other applications. Session management is supported. For example, a
client can first store data on \pkg{Rserve}, then modify the \R{}
environment on \pkg{Rserve} through several method invocations, with
intermediate results kept on the server, and finally retrieve results.
Method invocation follows normal \R{} syntax, and is sent as a series
of text messages. Data is sent in binary forms using an in-house
serialization protocol.  The current implementation supports basic
\R{} data types including integer, numeric and character vectors.
Complex data types such as S3 and S4 objects can also be encoded using
a low-level interface, but the decoding part has not been implemented
yet.  The unique part of \pkg{Rserve} is that it forks a new \R{}
process for each client.  Therefore multiple clients can be supported
simultaneously. This cannot be done with standard \R{}.  \pkg{Rserve}
also provides error handling and security support. There are
\lang{C}/\lang{C++} and \Java{} clients available for \pkg{Rserve}.

\subsection{The shared library approach}

\R{} and \Java{} can also be integrated into a single process, with
one embedded into the other. There is no direct programming interface
between \R{} and \Java{}. However, the interface between \Java{} and
\lang{C} is well-developed in the Java Native Interface (\lang{JNI}),
and the internals of \R{} itself are largely written in \lang{C}.  The
technique of bridging \R{} and \Java{} via \lang{C} has been used by
applications such as \pkg{SJava}~\cite{lang2000},
\pkg{arji}~\cite{carey2006}, and \pkg{JGR}~\cite{Helbig2004} and
related tools (\pkg{JRI} and \pkg{rJava}).

The shared library approach has two advantages. First, the interface
(\lang{C}) is a programming language, which is more powerful than any
encoding mechanism. It therefore gives more flexibility to integrate
languages.  Secondly, both \R{} and \Java{} are in a single process,
avoiding challenges in a typical client-server architecture such as
session management, data persistence and security.  Error handling is
also easy inside one process, because both \lang{JNI} and the \R{}
interface to \lang{C} have error handling mechanism. Some applications
also provide advanced features such as callback and object references.
These concepts are easier to implement than when using the message
stream approach.

There are also downsides to the shared library approach. Because \R{}
is not thread-safe, parallel processing requires additional layers
even though \Java{} supports multi-threading. Also, the learning curve
for low-level application development is steep, because interactions
between at least three languages are involved.

\pkg{SJava} can be used to embed \R{} in \Java{}, or to embed \Java{}
in \R{}.  Interesting features of \pkg{SJava} includes callbacks,
object references, and customized data conversion.  For example, the
\R{} instance embedded in \Java{} can call back to the original
\Java{} virtual machine. Object references allow data stored in one
language format to be referred to in another language. Thus a \Java{}
data structure (e.g. \lang{Vector}) can contain a reference to \R{}
data.  This avoids unnecessary between-language data transformation
and evaluation.  Data conversion in \pkg{SJava} is highly configurable
yet easy to use.  Built-in converters map \R{} character, integer,
numeric and logical vectors to \Java{} primary type arrays or existing
classes in \Java{}. Users can provide additional converters written in
\lang{C} or \R{} to extend or overwrite the default data
transformation behavior. These convert functions can be
``high-level'', with details of the data conversion handled by
\pkg{SJava}.  Moreover, the registration of user-level converters in
\pkg{SJava} is performed during run time via a well-defined interface.
Therefore, the transformation of arbitrary or complex data types is
achieved with minimal efforts.
 
\pkg{JGR} is an integrated development environment for \R{} written in
\Java{}. Its \Java{}-based graphical user interface communicates with
\R{} through two facilities: \pkg{JRI} (Java/R Interface) for \Java{}
to \R{} and \pkg{rJava} for the reverse task.  Both provide low-level
\R{}-\Java{} interfaces using JNI. Yet another example of the shared
library approach is \pkg{arji}, a Bioconductor package similar to
\pkg{rJava}.

\section{Bridging \R{} and web or grid services}

Web and grid services support in the \R{} community is receiving
increasing attention. At least two applications have been developed in
the last year to transform \R{} and \lang{Bioconductor} packages into
web services.

\pkg{RProteomics} uses Open Statistical Services to integrate \R{}
with \Java{}. The \Java{} application is further developed as a
\pkg{caGrid} analytical service, which offers semantic
interoperability and security control in addition to classical web
service capabilities. \pkg{RProteomics} focuses on proteomics data
analysis.

The Comprehensive R-based Microarray Analysis web service
(\pkg{CARMAweb})~\cite{Rainer2006} contains a \lang{J2EE} front end
and an \pkg{Rserve} back end.  It provides a web interface and a web
service interface.  For a typical service request in \pkg{CARMAweb},
the method being invoked and tuning parameters are specified as
primary data types (e.g. character, integer). Input data sets are
provided by files. The \lang{J2EE} front end organizes input
information into a \LaTeX{} document, and transfers it to \R{} via
\pkg{Rserve}. Data analysis in R is wrapped by \pkg{Sweave}, so that
\pkg{Sweave} extracts \R{} commands from \LaTeX{} and packs results
back to a document (\LaTeX, converted to PDF). Results are returned as
PDF files, and targeted for human interpretation.  This is in contrast
to typical web services where service responses are expected to be
interpreted by other applications.
  
Finally, it is worth mentioning efforts to expose \R{} functionality
through a web interface. This approach allows end users to take
advantage of the statistical computing power of \R{} without tedious
command line programming.  \lang{Perl} is often used as a bridge
between \R{} and the web. Examples include
\pkg{MIDAW}~\cite{Romualdi2005} (MIcroarray Data Analysis Web tool),
\pkg{RACE}~\cite{Psarros2005} (Remote Analysis Computation for gene
Expression data), and \pkg{webbioc}~\cite{Smith2005}.

\subsection{Statistical data exchange formats}

Statistical computations use specialized concepts (e.g., \code{NA},
corresponding to `missing' data observations) that do not exist in
general-purpose programming languages. Bridging \R{} with other web or
grid services requires either that these data structures be described
in a way that facilitates data exchange, or the services exposed by
\R{} make use of general purpose data representations that allow
interoperability without sacrificing too much expressivity.

There are three efforts that propose data exchange formats for
statistical data: \lang{StatDataML}~\cite{statdataml},
\lang{CDFML}~\cite{cdfml} and \lang{StatML}.  All of these formats are
XML-based markup languages aiming to facilitate the exchange of
statistical data across platforms.  They have been used in
communication between \R{} and other systems such as \lang{MATLAB} and
\lang{Java}.  McConnell~\cite{mcconnell2005} provides a review of
these data exchange formats.

There are several limitations to employing a formalized data exchange
model such as those outlined in the previous paragraph. First,
effective interoperability requires implementation of serialization
mechanisms in each target language. Second, data exchange formats
require integration into web or grid service standards like SOAP or
WSDL.  Third, no clear consensus on data exchange format is available,
leaving little hope for interoperability amongst services implementing
a single data model. Finally, from the perspective of communication
between \R{} and the web- or grid-services front end, the `shared
library' approach is much less dependent on formalized serialization
models than is the `message stream' approach. 

The limitations sketched in the previous paragraph point toward use of
light weight and \emph{ad hoc} data structures that capture key
features of the statistical domain, without requiring complex
serialization mechanisms or data structures in non-statistical
languages. This approach provides a short-term solution that enables
exchange of statistical data amongst a diversity of programming
languages, while sacrificing the expressivity of a carefully
structured exchange format and the long-term benefits of
interoperability such formats provide. This approach does not preclude
the implementation of structured exchange formats once patterns of
usage become well-established.

\begin{thebibliography}{25}

\bibitem{W3C2004}
D. Booth, H. Haas, F. McCabe, E. Newcommer, M. Champion, C. Ferris, and 
D. Orchard.\newblock Web Services Architecture.
\newblock \textit{W3C Working Group Note}, 2004.

\bibitem{Cerami2002}
E. Cerami. \newblock Web Services Essentials, Distributed Applications
with XML-RPC, SOAP, UDDI \& WSDL.
\newblock O\'Reilly Media, 1st edition, 2002.

\bibitem{Sun2006}
Java Technology and Web Services.
\newblock \url{http://java.sun.com/webservices/}

\bibitem{Irani2002}
R. Irani, S. Basha.\newblock AXIS: Next Generation Java SOAP.
\newblock Peer Information, 1st edition, 2002.

\bibitem{Ray2002}
R. Ray and P. Kulchenko.\newblock Programming Web Services with Perl.
\newblock O\'Reilly Media, 1st edition, 2002.

\bibitem{Foster2005}
I. Foster, H. Kishimoto, A. Savva, D. Berry, A. Djaoui, A. Grimshaw, B. Horn, F. Maciel, F. Siebenlist, R. Subramaniam, J. Treadwell, J. Von Reich.
\newblock The Open Grid Services Architecture, Version 1.0.
\newblock \url{http://forge.gridforum.org/projects/ogsa-wg}

\bibitem{Foster2001}
I. Foster, C. Kesselman, S. Tuecke.
\newblock The Anatomy of the Grid: Enabling Scalable Virtual Organizations.
\newblock \textit{International Journal of Supercomputer Applications}, 15(3):200-222, 2001.

\bibitem{Foster2002}
I. Foster, C. Kesselman, J. Nick, and S. Tuecke.
\newblock The Physiology of the Grid: An Open Grid Services Architecture for
Distributed Systems Integration.
\newblock \textit{Globus Project}, 2002.
\newblock \url{http://www.globus.org/research/papers/ogsa.pdf}

\bibitem{Treadwell2005}
J. Treadwell. 
\newblock Open Grid Services Architecture Glossary of Terms.
\newblock \textit{Global Grid Forum OGSA-WG}. GFD-I.044, 2005.
\newblock \url{http://www.ggf.org/documents/GWD-I-E/GFD-I.044.pdf}

\bibitem{Czajkowski2004}
K. Czajkowski, F. D. Ferguson, I. Foster, J. Frey, S. Graham, I. Sedukhin, D. Snelling, S. Tuecke, and W. Vambenepe.
\newblock The WS-REsource Framework, Version 1.0.
\newblock 2004.
\newblock \url{http://www.oasis-open.org/committes/download.php/6796/ws-wsrf.pdf}

\bibitem{FosterG2005}
I. Foster.
\newblock Globus Tookkit Version 4: Software for Service-Oriented Systems.
\newblock \textit{IFIP International Conference on Network and Parallel Computing},
\newblock Springer-Verlag LNCS 3779:2-13, 2005.

\bibitem{wsrflite2005}
WSRF::Lite - An Implementation of the Web Services Resource Framework.
\newblock \url{http://www.sve.man.ac.uk/Research/AtoZ/ILCT}

\bibitem{wsrfnet2005}
M. Humphrey and G. Wasson.
\newblock Architecture Foundations of WSRF.NET.
\newblock \textit{International Journal of Web Services Research}, 2(2):83-97, 2005.

\bibitem{pyglobus2006}
Python Globus (pyGlobus).
\newblock \url{http://dsd.lbl.gov/gtg/projects/pyGlobus/}

\bibitem{mcconnell2005}
P. McConnell, R. Haney, S. Mungal, M. Peedin, and S. Lin. 
\newblock Open Statistical Services. 
\newblock caBIG Documentation, 2005.

\bibitem{urbanek2003}
S. Urbanek.\newblock Rserve-A fast Way to Provide R Functionality to Applications.
\newblock In \textit{Proc. of the 3rd International Workshop on Distributed Statistical Computing (DSC 2003)}, 2003.
\newblock ISSN 1609-395X, Eds.: K. Hornik, F. Leisch, and A. Zeileis.

\bibitem{lang2000}
D. Temple Lang.\newblock The Omegahat environment: New possibilities for 
statistical computing. \newblock \textit{JCGS}, 9(3), 2000.

\bibitem{carey2006}
V. Carey. 
\newblock arji - another R-Java interface. 
\url{http://www.bioconductor.org/packages/1.9/bioc/html/arji.html}

\bibitem{Helbig2004}
M. Helbig, S. Urbanek, M. Theus. \newblock JGR, A unified interface to R.
\newblock \textit{useR!}, 2004.

\bibitem{Rainer2006}
J. Rainer, F. Sanchez-Cabo, G. Stocker, A. Sturn, and Z. Trajanoski.
\newblock CARMAweb: comprehensive R- and bioconductor- based web service
for microarray data analysis.
\newblock \textit{Nucleic Acids Research}, 43:W498-W503, 2006.

\bibitem{Romualdi2005}
C. Romualdi, N. Vitulo, M. D. Favero, and G. Lanfranchi.
\newblock MIDAW: a web tool for statistical analysis of microarray data.
\newblock \textit{Nucleic Acids Research}, 33:W644-W649, 2005.

\bibitem{Psarros2005}
M. Psarros, S. Heber, M. Sick, G. Thoppae, K. Harshman, and B. Sick.
\newblock RACE: Remote Analysis Computation for gene Expression data.
\newblock \textit{Nucleic Acids Research}, 33:W638-W643, 2005.

\bibitem{Smith2005}
C. Smith. 
\newblock Textual Description of webbioc.
\url{http://www.bioconductor.org/repository/devel/vignette/webbioc.pdf}.
 
\bibitem{statdataml}
StatDataML 
\newblock \url{http://www.omegahat.org/StatDataML/}

\bibitem{cdfml}
CDFML
\newblock \url{http://cdf.gsfc.nasa.gov/html/cdf_xml.html}

\end{thebibliography}
\end{document}