UserGuideChapter1

What is OPeNDAP?

The OPeNDAP provides a way for ocean researchers to access oceanographic data anywhere on the Internet from a wide variety of new and existing programs. By developing network versions of commonly used data access Application Program Interface (API) libraries, such as NetCDF , HDF , JGOFS , and others, the OPeNDAP project can capitalize on years of development of data analysis and display packages that use those APIs, allowing users to continue to use programs with which they are already familiar.

The OPeNDAP architecture uses a client/server model, with a {\em

{client}} that sends requests for data out onto the network to some "server", that answers with the requested data. This is exactly the model used by the World Wide Web where client programs called browsers submit requests to web servers for the data that make up web pages. Of course, OPeNDAP clients can do much more than browse this data. Using flexible data types suitable for many uses, including scientific data, the OPeNDAP servers deliver real data directly to the client program in the format needed by that client.

In fact, the network communication model used by OPeNDAP uses URL addresses and web servers ("httpd") to deliver data to the researcher. This is done by using the OPeNDAP software to convert a researcher's data analysis software into a sophisticated (though specialized) web browser. In addition to providing network-compatible versions of popular data access APIs, the OPeNDAP project also provides a software client and server toolkit to help other developers create network-compatible OPeNDAP versions of other APIs.

To expand the universe of data available to a user, OPeNDAP incorporates a powerful data translation facility, so that data may be stored in data structures and formats defined by the data provider, but may be accessed by the user in a manner identical to the access of local data files on the user's own system. Though there are limitations on the types of data that may be translated (See ( data,trans)), the facility is flexible and general enough to handle many of the possible translation. There are two important results:

• A user may not need to know that data from one set are stored in a format different from data in another set. Further, it may be possible that "neither" data set is stored in a format readable by the original (i.e. without OPeNDAP) version of the data analysis and display program he or she uses.
• No segment of OPeNDAP users will be effectively cut off from accessing data because of its storage format. A scientist who wishes to make his or her data available to other OPeNDAP users may do so while keeping that data in what may actually be a highly idiosyncratic storage format. Of course, it doesn't have to be in a highly idiosyncratic format. The point is that OPeNDAP can handle a wide variety of possible cases.

The combination of the OPeNDAP network communication model and the data translation facility make OPeNDAP a powerful tool for the retrieval, sampling, and display of large distributed datasets. Though OPeNDAP was developed by oceanographers, its application is not constrained to oceanographic data. The organizing principles and algorithms may be applied to many other fields where data can be stored on computers.

The population of people who may be interested in a system such as OPeNDAP may be divided into data consumers and data providers. Though it was an important observation to the development of OPeNDAP that the two roles are often assumed by the same scientists, the division is a useful one for the introduction of the system. The following two sections provide a broad introduction to the roles of data consumer and data provider. The remainder of this guide is organized around this distinction between classes of users.

Why Use OPeNDAP to Read Data?

A scientist wishing to examine and sample some dataset will typically be comfortable using a relatively small number of data analysis and display programs or packages. Some of these packages will use one of the popular data access APIs currently available. However, few data access APIs provide direct access to distributed data

refers to datasets that reside on different computers which are linked by a network such as the Internet. The computers may or may not be physically remote from each other. The main point is that the computers manage their data resources independently. In this guide the terms "remote\/} and {\em distributed\/" are used to imply independently managed resources.}, so this access must be made with network tools, such as web browsers or "ftp". While relatively straightforward in principle, this process can nonetheless become time-consuming and somewhat challenging in practice.

The following example illustrates some of the differences between accessing distributed data with the tools currently in widespread use, and the same operation using OPeNDAP.

An Example: Using ftp

The advent of the WWW has made possible simple data browsers that allow sophisticated interactive sampling of on-line datasets. Using a web browser and "ftp", a user can sample any of several large oceanographic datasets available on the Internet. However, there are several problems with these data search engines that may only become apparent when a user actually tries to use the data.

Among the problems that can arise are those that appear when a user tries to use the results of one dataset to search a second dataset. Suppose that a user wishes to choose a sea-surface temperature image from the NOAA/NASA Pathfinder AVHRR archive at:

http://podaac-www.jpl.nasa.gov/mcsst/mcsst_subset.html

using the results of a time-series generated from the COADS Climatology archive at:

http://ferret.wrc.noaa.gov/fbin/climate_server

The steps are theoretically straightforward:

1. Create the time series from the COADS Climatology archive. This is done by answering the menu of options on the COADS web page.
2. Import the time series from step 1 to the user's local data analysis system. Note that this step may itself require several steps:
   1. The data must be down-loaded, using "ftp" or a similar program.
   2. Once down-loaded, the data may have to be converted into a format that can be read by the data analysis program.

3. Examine the data and formulate a request to the AVHRR archive. This is again done by answering the menu of option on the AVHRR Web page. Note that the COADS and AVHRR pages are not completely compatible in this respect. For example, the date formats of the two pages are different.
4. Import the result of step 3 to the user's local data display system. This may also require several steps:
   1. The data must be down-loaded again.
   2. And again, once down-loaded, the data may have to be converted into a format that can be read by the data analysis program. Note that the set of available formats on the COADS page are distinct from the available options from the AVHRR archive.

5. Think about the results.

Though the procedure is straightforward and the web servers designed to make sampling the datasets a simple task, upon close examination, the combination of the steps may create unforeseen difficulties. For example, a request to the COADS server will return either a spreadsheet suitable for use on a PC, a netCDF format file, or a file in one of a selection of simple ASCII formats. If the user is fortunate, the returned file will already be in a format compatible with the desired analysis package. But not all users will be so fortunate. Often this file must be converted to some other file format before it can be imported to the user's analysis program. This may or may not be a simple task.

Even a file format for which a user is properly equipped may be used in an unfamiliar manner. For example, the independent and dependent variables might be in a different order or an ASCII data file may use tabs instead of spaces.

Assuming the import of the COADS data has been accomplished and boundaries for the AVHRR search identified, the task of selecting from the second archive may begin. Unfortunately, the request to the AVHRR archive will return either a GIF picture, an HDF format file, or a raw (binary) data file. Again, importing this output into the user's analysis program may or may not be simple, but it will not be the same procedure as the one used for the first data request.

Other problems are also apparent. The COADS Climatology sampling program requests the user supply dates (month and day), whereas the AVHRR archive asks for the "Julian day" (an integer between 1 and 365 or 366). One server will accept "S" and "W" to indicate South latitudes and West longitudes, while the other requires that these be indicated with negative coordinate values. The sampling of the COADS dataset, while flexible, may not allow sampling in the manner the user needs. It cannot, for example, provide a section except along a line of constant latitude or longitude. If a user wanted to see a section along a NE-SW line, it would be a challenging and time-consuming task to assemble one from many small data requests.

Further, it might be desirable to use the results of sampling these two databases to construct a time series. This could conceivably mean repeating the entire procedure many times.

An Example: Using OPeNDAP

To produce the same data selection using OPeNDAP, a user would follow essentially the same steps. However, the steps themselves would be performed differently. Once the user's data analysis package has been converted to an OPeNDAP client (( opd-client,link)), the \tbd{add xref to install GUI

clients} accesses to the remote datasets are made through the analysis package itself. Instead of specifying a data file by a pathname reference to some local disk file, the user specifies a URL, which may point to either a local or a remote dataset. Here is a re cap of the same operation, outlined as they would be performed by an OPeNDAP application program:

1. Create the time series from the COADS Climatology archive. This is done by using the sampling facilities of whatever data analysis program a scientist is familiar with. If desired, OPeNDAP constraint expressions may be used to reduce the network load, or to provide a sampling scheme not supported by the data analysis program.
2. The data need not be imported to the user's data analysis program, since it was down-loaded and converted automatically in step 1.
3. Examine the data and formulate a request to the AVHRR archive. This is again done through the sampling facilities of whatever data analysis program the user is using, and OPeNDAP constraint expressions. Note that, whatever their actual format, both COADS and AVHRR archives appear to the OPeNDAP client to be stored in identical formats.
4. The data need not be imported to the user's data analysis program, since it was down-loaded and converted automatically in step 3.
5. Think about the results.

It is important to note that "any" data analysis package that can handle one of the DODS-supported data access APIs can be converted into an OPeNDAP client program capable of reading data stored by "all" of the DODS-supported data access APIs. (There are some limitations on translation. See ( intro,opd-client) and ( data,trans) for more information.) Therefore, assuming the user has some analysis package capable of doing the required sampling and analysis on local data, all the steps would be performed from within that package, just as if the user were operating on local files. The result is a simpler procedure, even though the same essential steps are followed.

The OPeNDAP scenario has, among others, the following advantages:

• The user need not learn about any of the archival formats, since the OPeNDAP server and client cooperate to deliver the data in the format in which the analysis package expects to see it. Whereas the user of the ftp server has to worry about importing the data into the analysis program, the OPeNDAP client program imports it transparently.
• The user can sample the distant datasets in any fashion supported by his or her own (local) analysis package. Unnecessary data need not be sent over the Internet.
• By appending a "constraint expression" to the URLs given to the analysis program, the user can sample data using techniques that their analysis program cannot do.\footnote{For example, suppose a user wishes to access the NODC XBT database using a program that uses the netCDF API. A program that can process the arrays that netCDF manipulates are largely unsuitable for XBT station data. However, a user can define constraint expressions in the URL to sample the data and deliver it in a form the netCDF API can use. For more information about constraint expressions, see Section~(opd-client,constraint). For more information about data models and translation, see Chapter~(data).}\tbd{Use a different example in the footnote}
• A substantial amount of the searching and sampling is performed on the server machines. This reduces Internet traffic, as well as decreasing the load on the local machine.

The OPeNDAP Client

OPeNDAP uses a client/server model. As mentioned, the OPeNDAP servers are simply "httpd} web servers, equipped to interpret an OPeNDAP URL sent to them. (See \chapterref{opd-server".) The OPeNDAP client program can be any program that uses one of the supported APIs, such as JGOFS or netCDF.\footnote{Or a program specially developed to read data from OPeNDAP servers.}

Without OPeNDAP, an application program that uses one of the common data access APIs such as netCDF will operate as shown in File:Intro,fig,unlinked. The user makes a request for data from the application program. The program in turn uses procedures defined by the data access API to access the data, which is stored locally on the host machine. Some APIs are somewhat more sophisticated than this, of course, but their general operation is similar to this outline.

\figureplace{The Architecture of a Data Analysis Package.}{htbp} {intro,fig,unlinked}{unlinked.ps}{unlinked.gif}{}

The operation of an OPeNDAP client is illustrated in File:Intro,fig,linked. Here, the same application program that was used in File:Intro,fig,unlinked has been linked with an OPeNDAP version of the data access API. Now, in addition to being able to use local data as before, the application program is able to access data from OPeNDAP server anywhere on the Internet in the same manner as the local data.

To make some program into an OPeNDAP client, it must only be re-linked with the OPeNDAP implementation of the supported API library. This is a simple process, generally requiring only a few minutes. The process will create a program that accepts URLs, specifying a location for the data somewhere on the Internet, in addition to file pathnames which only specify a location on the local platform's file system. (See ( opd-client,link).)

\figureplace{The Architecture of a Data Analysis Package Using OPeNDAP.}{htbp} {intro,fig,linked}{linked.ps}{linked.gif}{}

OPeNDAP also provides a data translation facility. Data from the original data file is translated by the OPeNDAP server into an OPeNDAP data model for transmission to the client. Upon receiving the data, the client translates the data into the data model it understands. (See ( data) for more information about the OPeNDAP data model.) Because the data transmitted from an OPeNDAP server to the client travel in the OPeNDAP format, the data set's original storage format is completely irrelevant to the user of an OPeNDAP client. If the client was originally designed to read netCDF format files, the data returned by the OPeNDAP-netCDF library will appear to have been read from a netCDF file, whatever the actual format of the files from which the data were read\footnote{Note that there is a limit to what can be translated. An API meant to support two-dimensional arrays may be able to handle one-dimensional vector data, but a program designed to process one-dimensional vector data will not know what to do with a two-dimensional array. The set of data access APIs supported by OPeNDAP contain several such mismatches. See Section~(data,trans) for more information.}. If the program expects JGOFS data, the DODS-JGOFS library will return data that seem to have come from a JGOFS dataset, again, no matter what the actual input file format.

OPeNDAP does not pretend to remove all the overhead of data searches. A user will still have to keep track of the URLs of interesting data sets in the same way a user must now keep track of the names of files containing interesting data. an OPeNDAP \new{catalog service} is in the process of being constructed that will help users scan the available datasets.

Providing Data with OPeNDAP

The OPeNDAP data provider is the person or organization willing to make their digital datasets available to the community with an OPeNDAP server.

The designers of OPeNDAP recognized that many of the data users are also the data providers, and OPeNDAP was built with a recognition that providing the data should be as simple and as straightforward as possible. In many cases, once a local web server is equipped to become an OPeNDAP server, a scientist need do very little beyond what must be done simply to make the data available locally. (i.e., Put the data into a file format that can be read by the locally used data analysis and display programs.) The tasks of a data provider can be separated into three parts:

• Install and configure the OPeNDAP server.

(( opd-server,install).)

• Create whatever ancillary data files are needed by the data set (if any). (( intro,ancillary).)  %
• Register the data set with the master directory (optional).  %
• Create the data catalog.

The OPeNDAP Server

The OPeNDAP data server is simply made up of a regular httpd server equipped with CGI programs (or filters) that will respond to requests for dataset structure, data attributes, and data itself. (See ( data,dap) for a description of the data returned by these requests and see ( opd-client,url) for a description of the OPeNDAP URL syntax used to send these requests.) Most of the task of a data provider consists of configuring this server. While perhaps not a trivial task, it potentially represents far less effort than packaging a dataset for submission to some central data archive. Furthermore, modifying a server's configuration to accommodate new data will be an almost trivial task, involving the simple editing of a configuration file.

Ancillary Data

In order for an OPeNDAP client to accept data from an OPeNDAP server, it must be able to allocate the data structures and arrange internal labels to organize the incoming data. The information the client library needs to do this organizing is called the ancillary data\footnote{It is also referred to as

the Data Descriptor Structure and the Data Attribute Structure. See

Chapter~(data) for more details about these structures.}. For many APIs, the ancillary data is inherent in the data files themselves, and the OPeNDAP server can glean that information by scanning the data files. For large data archives, where scanning the data files is impractical, and that might not change often, OPeNDAP can cache the ancillary data to speed access times. When a client requests the ancillary data, the OPeNDAP server can check this data cache first before scanning the data files.

This feature is useful in other cases because not all data file formats are self-describing. For example, a data set might contain several files of time vs. temperature data; the header information describing which numbers are temperature and which time may be in a different file or may simply be understood by the user of the local data analysis program equipped to look at this data. As an example, data accessed by OPeNDAP servers using the FreeForm data access API require provider-created ancillary data files.

Administration and Centralization of Data

Under OPeNDAP, there is no central archive of data. Data under OPeNDAP is organized in a manner similar to the World Wide Web itself. That is, all one need do to make one's data available is to start up a properly configured "httpd" server on an Internet node that has access to the data to be served. Each data provider is free to join and to leave the system when it is convenient, just as any proprietor of a web page is free to delete it or add to it as whimsy demands.

Of course, as can also be seen on the World Wide Web, there are some disadvantages to the lack of central authority. If no one knows about a web site, no one will visit it. Similarly, listing a dataset in a central data catalog, such as the Global Change Master Directory (http://gcmd.gsfc.nasa.gov/),can make data available to other researchers in a way that simply configuring an OPeNDAP server does not. OPeNDAP provided a facility for registering a data set with the GCMD catalog, which makes the data set known to the OPeNDAP data location service.

The remainder of this book will be divided into three major sections: instructions on the building and operating of OPeNDAP clients; a tutorial and reference on running OPeNDAP servers and making data available to OPeNDAP clients; and technical documentation describing the implementation details (and the motivation behind many of the design decisions) of the OPeNDAP software.