A few years ago, distributed
computing was a cutting-edge technology used only by early adopters or IT groups with very
specific technological needs. Today, due to the increasing demand for internet computing
and enormous complexity of enterprise-wide computing infrastructure, distributed computing
has become very much an accepted solution. Distributed computing is, however, still in its
infancy when it comes to the technology that supports it. This status is evident in the
lack of mature standards and uniform approaches to common issues.
In terms of distributed software components, the competing standards are the Object
Management Group's Common Object Request Broker Architecture (CORBA) 2.1, Microsoft's
ActiveX and Sun's JavaBeans. No other standards for distributed components are likely to
be successful in today's market. Previous technology proposals such as OpenDoc, Taligent
Frameworks and NeXT Objects have come and gone.
Until the 1990s, multi-user computing consisted primarily of a single large processor
(generally a mainframe or a midrange system) supporting multiple users through a
time-sharing scheme. Users interacted with the processor through a terminal and a
keyboard. Terminals were called dumb terminals then because they performed no computing.
Instead, they simply displayed character output generated by activities performed on the
main processor.
This style of computing was known as host-based computing, but it frequently is referred
to as a single-tier computing because computation occurs in only one location: the main
processor.
Initially, networking between PCs followed a peer-to-peer model: PCs could access data and
documents on other PCs, butt there was as yet no centralized repository for all a
company's data. Although this process avoided some of the problems of data duplication, it
did not solve them, and it created unwanted complexity in managing data that was
distributed between large numbers of PCs att an IT site.
Eventually, the model that emerged was to place all data on a central server and have the
PC clients access the data as needed. Processing still would bee performed locally on the
PCs, but all documents and data would rely on a central file server. Eventually, this
file-sharing model evolved to a system in which the file servers (now known as database
servers) also performed computational work. This model came to be known as client-server
computing and constituted the advent of two-tier architectures. The client tier consisted
of PCs, and the server tier included servers of all types.
Eventually, client server computing evolved to perform many data management tasks on the
server itself. This transition conserved network bandwidth and, for certain applications,
freed client PCs to do other work while waiting for the results of a query or a report.
Predictably, this approach encountered the problem of overloading servers, which now were
saddled with thee task of providing database access and performing much of the enterprise
computation. (Interestingly, the now-freed resources of the PC were quickly put to use by
new versions of the operating system and by business applications such as PowerBuilder
that required significant processing power from the client.) Eventually, a third tier
appeared in enterprise computing: the application server. In this scheme, the processing
logic for business applications was performed on an application server that interacted
with a separate database server: meanwhile, the client handled data presentation and some
data validation. There were now three distinct tiers: client PCs, application servers, and
database servers. This model proved successful and frequently is found today at businesses
of all sizes.
Three-tier computing quickly, lost its conceptual simplicity, however. The two driving
causes of this change were the increasing heterogeneity of computing in IT departments and
the advent of large, multi-site enterprise computing architectures.
To address the difficulties of developing applications for such a widely disparate
architecture and to handle the management of so many tiers, two different technologies
evolved. The older technology is messaging middleware, which arose from the success of TP
monitors. The newer technology, which is seen as the next wave of enterprise computing, is
distributed objects.
Middleware originally was designed as a solution to the problem of developing software for
heterogeneous and n-tier computing infrastructures. Specifically, the question was how to
write software that could access databases such as Microsoft SQL server, Oracle, and
Sybase while supporting clients running Windows, MacOS, and Unix and while interfacing
with application servers that ran on Unix boxes. Beyond the problem of writing the
application for such as heterogeneous environment was the question of maintaining the
application if a new database server was added or a new client operating system was
introduced.
The solution middleware offers is to provide developers with a uniform interface through
which their programs can access enterprise resources. The developers interact with only
the middleware layer, and the middleware then performs the necessary translations for the
respective databases, operating systems, and clients. In this sense, the middleware
presents the developer with a single, consistent interface that masks the complex
computing infrastructure. The middleware package-generally a large, intricate and
expensive piece of software-is installed in the enterprise between the application servers
and the clients on one side and the database servers on the other.
Of the many categories of middleware, modern TP monitors come closest to realizing the
original goal of the solution. TP monitors initially were devised to ensure the integrity
of transactions in high-volume, on-line transaction processing (OLTP) environments. Since
then, they have been extended to be effective tools for masking client and database
heterogeneity. However, TP monitors are primarily transaction-oriented and rarely are used
as a pure communications medium across different processes in the enterprise. This latter
goal-the ability to communicate data across disparate platforms, multiple sites, and
through or within firewalls-has become the main driver for generic middleware. Pure
middleware products now relegate themselves to the primary task of providing the
communications infrastructure for data processing within the enterprise.
Middleware originally was built on a technology called Remote Procedure Calls (RPCs), but
it has evolved to a message-oriented approach, which is the most common implementation
today.
Remote procedure calls
RPCs allow a program running one computer to call a procedure (similar to a function or
subroutine) that executes on a remote machine and performs a single, discrete task.
Generally, the RPC sends a request in the form of some data to the Unix operating system,
and today they function in many high-end client server operating systems such as Windows
NT.
This approach is conceptually simple and can be learned by any experienced programmer.
However, it suffers from several drawbacks. First, the program containing the RPC call
must wait for a response to the call before continuing. This approach (known as
request-reply) depends on synchronous operation: The request must be responded to before
processing can continue. This process requires guaranteed communication and suffers from
difficulties if the
remote machine is busy, offline, or if it cannot return a response quickly.
Second, RPCs are very tedious to write. Due to their conceptual simplicity, numerous RPCs
must be generated for even straightforward work to be done between computers. Each of
these RPCs must be written individually, requiring considerable effort on the part of
developers. Finally, RPCs allow point-to-point communication between machines but do very
little to mask thee details of implementation from the developers. Therefore, although
RPCs are effective communication devices, they are not suited to serve as the principal
form of middleware.
RPCs are generally not sold as stand-alone packages. Unix provides one model of RPC,
Windows NT another (although similar) model. The primary third-party RPC implementation
comes from The Open Group as part of the Distributed Computing Environment (DCE). DCE is
massive, enterprise-wide package that is supposed to provide a comprehensive set of
services for distributed computing specifically optimized for business processing. It
includes modules for security, clock synchronization, distributed file systems, and many
other purposes. However, the lack of commercial successes of DCE has thwarted attempts to
assess its capabilities to deliver on these claims. A notable exception to the absence of
third parties is Noblenet, whose product RC 3.0 provides an RPC infrastructure and
numerous software development tools that ease programming for RPCs.
Message-oriented middleware
Message-Oriented Middleware (MOM) uses a different approach from RPCs. Messages are sent
asynchronously to a destination. that is, they are sent there, but the sending program
does not require a reply before continuing operations. This approach generally is
implemented using one of two mechanisms: publish-and-subscribe or message queuing.
In publish-and-subscribe, the message is sent (or published) and the target destination
awaits receipt of messages specifically addressed to it or to a destination in which it
has expressed interest (subscribed). In this model, many different targets could subscribe
to the same data. For example, several programs could be waiting for real-time data to
come in from the factory floor. Rather than having the real-time data collection process
figure out who is interested in its data, the program simply publishes the data to a
predetermined target, whereupon all subscribers would receive the data. Another example
would be sending the same HTML page to numerous intranet users.
Publish-and subscribe does not work well in 'store-and-forward' situations, where data
must pass over several network segments and be stored safely at each segment before being
sent on. Under the message queuing approach, the message is sent to the target's queue.
Then the target, on its own schedule, can check the queue for messages it needs. This
approach is suited to communication with devices that might be off-line, such as computers
of mobile users. The messages pile up in the queue and await connection with the target to
begin delivery.
In both cases, the middleware package manages the message, making sure they are delivered
to the correct queue or target in real time and that they are sent only once.
How does the original program obtain an answer if its computation is predicated on a
return message? It checks its own queue for messages from the original destination system.
If it cannot proceed without the specific reply, it will continue checking the queue until
the reply is received (or until some preset interval has been established, whereupon it
views the delay as an error condition). In this manner, if thee program needs the
reliability of a request-respond model as is provided by RPCs, it can implement it with
MOM, but the program has the additional asynchronous capability of sending message that do
not require a response.
The absence of an immediate response does not create a reliability issue for asynchronous
operations based on MOM. If for some reason the MOM cannot deliver the message, the MOM
will perform some predetermined action. The two most popular options involve use of a
callback: Either an error function is called by the MOM to inform the sending program of
the error, or the connection uses multiple threads for communication, one of which is used
to signal the error. Therefore, a built-in mechanism ensures notification if a message
cannot be delivered or if some other error occurs.
In practice, MOM has not been able to fulfill the original mission of providing a uniform
interface across heterogeneous clients, operating systems, and databases. Thee main reason
for this failure has been the near-exponential growth in complexity of most enterprise
computing infrastructures. Instead, MOM has emerged as a messaging infrastructure; It
provides enterprise-wide communication of data, but third-party products then perform the
translation to the individual databases or applications as needed. An example of this
arrangement is TSI International's Mercator translator package that plugs into IBM's
MQSeries MOM. Many sites choose to write their own custom interfaces to the MOM.
This problem highlights one key drawback of messaging middleware: No standards for the
message formats exist. Every vendor has its own format and design for messages. As a
result, a key criterion when assessing a specific middleware package is the extent of
available third party software that can plug directly into the MOM.
Transaction processing monitors
Modern TP monitors provide the same transactional integrity across numerous client
platforms and servers and interface with most leading databases. In addition to verifying
the transaction, they enforce numerous other data-processing policies, such as security,
transaction and error logging, data replication, and so on.
Traditionally, TP monitors have used synchronous communication: For every action they
perform, they require verification that the action has occurred correctly. This
requirement is inherent in the need to provide data and transactional integrity. Although
this function is suitable for OLTP, the response delay built into this model means the TP
monitors have not functioned well as pure message-passing tools (such as for real-time
data feeds) in the same manner that MOM currently is used. Recently, however, TP monitors
have added asynchronous communications to their feature list and now provide message
queuing similar to most MOM. This functionality is provided separately from the TP
monitoring activity. However, this development underscores the increasing convergence of
MOM and TP monitors.
TP monitors have retained a singular commitment to making transaction integrity in a
heterogeneous environment as easy as possible for programmers to develop. Most TP
monitors, despite supporting a wide variety of clients and servers, use only 30 or 40
application programming interfaces (APIs), so most developers can avail themselves easily
of TP monitors' capabilities. In addition, The Open Group has published a set of standard
APIs for TP monitors, making their appeal to developers even stronger.
Distributed computing
Although RPCs, messaging middleware, and TP monitors provide a communications mechanism by
which business processes can be performed across a wide variety of platforms, they share a
common drawback in that the middleware layer typically is proprietary. This layer is
defined and maintained by one vendor, and all tools and applications must integrate with
the layer to be able to derive its benefits.
This model may work well within a single enterprise, where a uniform environment can be
created; however, it begins to be difficult if internet computing and substantial remote
access are necessary. Dial-in customers, extranets, and Web site visitors may need to
perform business transactions with a company, but they cannot be expected to have chosen
the necessary software to plug into the messaging middleware used by the vendor site.
Therefore, companies with these needs increasingly are looking to distributed objects to
allow global access to resources while hiding platform and database heterogeneity.
A method gaining increasing popularity in this context is based on ORBs. These software
products allow software components to interact across the enterprise. To understand how
ORBs work requires examining the nature of objects and components.
How components work
In their simplest form, components are parts of a program that can be called by the main
program to perform certain specific tasks. Some of the earliest component-like
technologies were Dynamic Link Libraries (DLLs). These libraries, common in Unix for more
than 10 years and in the Windows family of operating systems since the early 1990s, work
in the following manner. A DLL will consist of a series of related functions, for example,
functions that will render three-dimensional (3-D) graphics on the screen. This DLL is
known to exist on the system at some predefined location (on Unix, this location is almost
always/etc/lib; on Windows systems, the location can change, but it must be somewhere on
the execution path as specified in the user's environment). When a function in the DLL is
called, the DLL is read into memory and linked into the main program as if the DLL always
had been part of the original program. Once linked this way, any and all functions in the
DLL can be called and executed without further steps.
In this context, components can be defined as being a group of related functions
encapsulated with data into an object that can be called by a running program to perform a
specific task. The calling program can be on the same machine, on a different machine
within the enterprise, or on a completely different continent connecting to the calling
program across the internet. Examples of objects typically found on the same machine are
ActiveX components and JavaBeans; objects found across the enterprise would be CORBA
objects plus the previous two; objects found across the Internet would be CORBA objects
and Java applets. These are just they typical settings, however. Any one of these
component types can be found anywhere in the enterprise.
CORBA and ORBs
For components to work successfully across the enterprise, many vendors felt it was
necessary to band together and form standards by which components could be defined to a
central authority and invoked with the knowledge that any standard-compliant component
could be found and activated successfully. A group of vendors came together in the early
1990s as the Object Management Group (OMG) to fashion a set of guidelines for distributed
computing with components.
Unfortunately, the group chose the use the term "object" rather than
"component," forever causing confusion. OMG-defined objects are really
components. Components tend to be larger than simple objects-often, they are built from
objects-and they tend to lack some characteristics peculiar to objects. In addition,
components need not be written with object-oriented programming languages. Nonetheless,
because the term "object" is used by the OMG to refer to components, discussion
of distributed computing often uses these two terms interchangeably.
The first document produced by the OMG was CORBA VERSION 1.0. Since its first release, the
document has been updated several times. CORBA 2.1 was released in later 1997.
The CORBA models uses ORBs to handle requests from programs to components and from
components to each other. ORBs are programs that serve as traffic cops: They handled the
requests, fine the requested components, call them, and return the results. To enable
these operations, CORBA defines a standard way for components to specify their interfaces;
the kind of data they expect to be handed and the kind of data they expect to return. The
standard way of defining interfaces is known as the CORBA Interface Definition Language
(IDL). Using IDL, a programmer can identify that a component will accept an integer (for
example, the diameter in centimeter) and return an integer (the circumference, rounded to
closest centimeter). These interfaces are in IDL (the syntax of which looks similar to
C++). A translator program then generates code for the component: client stub code and
server skeleton code. This code is the basis on which objects can speak to each other.
Once the IDL definition is written, the interface definition can be placed in the
Interface Repository, a database where all object interfaces that a given ORB can
recognize are stored. For an ORB to interact with a component, the component's interface
information (technically, its metadata) must first exist in the Interface Repository.
IIOP
Introduced with the CORBA 2.0 specification in 1996, the Internet Inter-ORB Protocol
(IIOP) is a key extension to the CORBA environment. It solves the problem of how ORBs
communicate with each other. Prior to CORBA 2.0, the OMG had begun designing a protocol
for inter-ORB communication. This protocol came to be known as the General Inter-ORB
Protocol (GIOP). It specified the format of seven key data messages ORBs, would need to
share data, and it detailed how data should be sent between ORBs. Specifically, the data
needed to follow the formatting rules of the Common Representation (CDR). CDR rules
compensate for the fact that different processors use numbers in different formats (for
example, on Intel machines, a two-byte number is stored with the low-value byte first; on
RISC machines, the high-value byte generally is stored first).
Java-based component architecture
JavaBeans is a specification for components that can be deployed on any Java platform.
Like Java applets, JavaBeans components (Beans) can be used to give Web pages or other
applications interactive capabilities such as computing interest rates or varying page
content based on user or browser characteristics. JavaBeans provides a component
development environment to extend Java's cross-platform "write once, run
anywhere" capability to the development of enterprise-class applications with the
features of component reusability. JavaBeans has emerged as a component definition to
create distributable, reusable objects within the Java environment. The Java-based
component model has followed a slightly different evolutionary path than Microsoft's
Distributed COM (DCOM), although both had their foundations in client-side GUI component
development.
The Java model did not emerge from an RPC-based component definition that grew to include
distribution and object interoperation. Instead, the JVM concept and associated Java
programming language created an environment where applets could be written and then
fetched across the internet and run anywhere. As the applet concept grew and Java began to
take hold as a cross-platform development language, JavaSoft (a division of Sun) sought to
offer a component model by providing JavaBeans as a component definition. The actual
component architecture model is a compilation of facilities from both SunSoft (a
subsidiary of Sun) and JavaSoft, the key to which is Java. Fundamental to a working
Java-based component architecture model are the Java based APIs that provide the operating
services on top of the Java language environment. The collective component model includes
Java, Java-based APIs and JavaBeans.
In essence, JavaBeans is a component specification for the Java language. Initially
intended for the Sun- originated Java and aimed at NEO (Sun's first commercial ORB, which
Sun abandoned in January 1998), it has developed greater compliance to CORBA standards and
broadened its interface capability to become a true, standalone component framework.
JavaBeans
A Bean is a piece of Java-based software that can range from a small visual control such
as a progress indicator to full-fledged application such as spreadsheets. A Bean is a
specialized, primarily visual component Java class that can be added to any application
development project and manipulated by the Java builder tool. Beans can be combined and
interrelated to build Java applets or applications and as such must be executed within a
larger application (the hosting application is called a container). Beans also can be e
used to create new, more comprehensive, or specialized Beans.
Beans can be flexibly manipulated, but not executed, without the requirement for a
dedicated object container. The properties of a Bean are attributes, which describe the
state of the Bean such as color, position, and title, and methods, which are the actual
functions to be performed by the Bean. Bean methods include functions that can be
performed on the Bean itself to manipulate its internal state or to interact externally
with other objects. Methods also can respond to events, which are actions performed when
special internal or external signals occur such as a mouse click. Beans also can implement
conceptual and non-visual objects such as business rules. In these regards, JavaBeans do
not differ much from other components, except for their requirement of a Java
implementation.
Enterprise JavaBeans
EJBs are specialized, non-visual JavaBeans that run on a server. The EJB 1.0 specification
defines a component model to support multi-tier, distributed object applications. EJBs
extend JavaBeans from its origins as a client-side GUI component ÿset to a server-side
componentÿmodel, thus simplifying the process of moving application logic to the server
by automating services to manage components.
Server components execute within a component execution system, which provides a set of
run-time services for the server components, such as threading, transactions state
management, and resource sharing. This requirement means that EJBs do not need the
standard JavaBeans container context to execute.
EJB provide hooks for enterprise-class application developers to build reusable software
components that can be combined together to create full-function applications. The hooks
include interfaces to common development environments and richer class libraries geared
toward the enterprise.
The EJB model supports implicit transactions. To participate in distributed transactions,
individual EJBs do not need special code to distinguish individual transactions and their
operations. The EJB execution environment automatically manages the start, commit, and
rollback of transactions on behalf of the EJBs. Transaction policies can be defined during
the deployment process using declarative statements, Optionally, transactions can be
controlled by the client application.
Microsoft's COM
Microsoft's Component Architecture is based on Microsoft's COM, which was originally
designed to provide interoperability between Windows applications (Microsoft Office in
particular). Microsoft's COM supports development of components in the form of ActiveX
controls, compound document services and frameworks in the form of DocObjects, and service
components in the form of COM objects (OLE) and their automation.
The COM model is extended by Distributed COM (DCOM), allowing clients to access components
across a network and supporting client-to-server and server-to-server connections between
Windows 95 and Windows NT systems. The COM model is extended further by Component Object
Model Services (COM+), which supports platform-independent development of componentized
applications over standard protocols such as TCP/IP.
Microsoft bundles COM/DCOM along with a series of other framework facilities and
development tools into Windows NT Enterprise, taking the view that component models and
frameworks are integral elements of the operating system. Like many component
architectures, COM/DCOM is dependent on object communications brokering. However, unlike
other component architecture implementations, Microsoft's approach must communicate across
three networking protocols (NetBIOS, Novell IPX/SPX, and TCP/IP) which reflects of the
mixed heritage of PC component technologies and the lack of industry standards for PC
desktops. Microsoft is using Microsoft Transaction Service (MTS) as an abstract layer
behind which the three protocols can be bridged. Although not explicitly part of DCOM, MTS
is an important element in the Microsoft view of distributed applications and comes
bundled in Windows NT Server, Enterprise Edition.
COM and ActiveX
COM principally provides the architectural abstraction for OLE objects. COM supports an
implementation typing mechanism centered around the concept of a COM class. A COM class
has a well-defined identity, and a repository (the system registry) that maps
implementations (identified by class IDs) to specific executable code in the form of
ActiveX controls. (ActiveX controls were formerly known as OLE controls, or OCXs). COM
automation allow applications such as Microsoft Internet Explorer, Microsoft Office, and
development tools such as Visual Basic to access COM objects.
From a COM point of view, an object typically is a subcomponent of an application that
provides a public interface to other parts of an application or to other applications. OLE
objects usually are document-oriented and are tied to some visual presentation metaphor.
The typical domain of a COM object is a single-user, multitasking, GUI-based desktop
running Microsoft Windows. The primary goal of COM and OLE is to expedite collaboration
and information sharing among application suing the same desktop, as in the case of the
Microsoft Office suite. This view of COM is different from the CORBA perspective in which
an object is an independent component providing services to client applications.
DCOM
To extend COM's reach into distributed environments, Microsoft extended the COM model with
DCOM, allowing clients to access components across a network. To access a COM object on
some other machine, the client relies on DCOM. DCOM transparently transfers a local object
request to the remote object running on a different system.
DCOM provides functions similar to the naming, security, persistence, and versioning
services found in CORBA. For directory services, DCOM relies on Active Directory,
essentially a combination of Domain Name System (DNS) and Lightweight Directory Access
Protocol (LDAP).
Because DCOM uses a proprietary implementation of the DCE RPC for its component
architecture, COM/DCOM is limited to use within a homogeneous Microsoft environment. In
fact, a major stumbling block in distributed computing is the absence of real
interoperability between COM/DCOM objects and CORBA components. Even getting these two
component models to communicate with each other has proven difficult. Limited
communication between these component type is possible through some third-party products.
The notable exception to this rule is Object Bridge from Visual Edge Software, which
provides true interoperability between CORBA and COM/DCOM (most other packages simply
allow CORBA-to-OLE Automation interoperability). The difference is that the latter
approach requires use of CORBA's dynamic invocation function, which is significantly
slower than the more common static invocation used among CORBA components.
Microsoft Transaction Server
MTS is a component-based TP system for developing and deploying transaction-oriented
processing on Windows NT server. MTS also defines an application programming model for
developing distributed, component-based applications, and it provides a runtime
infrastructure for deploying and managing these applications. MTS was written using COM
and uses DCOM to communicate with resource managers and other MTS servers on a network.
When COM components are installed into MTS and client and server components that are
distributed across more than once node and that must use DCOM to communicate are managed
automatically.
MTS is bundled into Windows NT Enterprise Edition and acts as a container for COM service
components. It defines classic services such as multithreading, concurrency control, and
shared resource utilization, and relies on native Windows NT services to provide security,
naming, and systems management services. Other MTS services include pooling, queuing,
connection, and security, all architected within the COM/DCOM model.
The MTS protocol is designed to accept automatically COM/DCOM- based applications with
some additional configuration and DLL packaging. Therefore, it acts as the primary
bridging protocol for applications written directly to WinSock, Remote Automation, direct
MS-RPC, or COM/DCOM specifications. The MTS protocol specifically allows asynchronous
application operation, using MSMQ technology as an enabling protocol.
MTS is enabled by MSMQ-an operating system service providing store-and-forward
connectivity between two application programs. MSMQ enables two programs to communicate
with one another using queues and includes a queue manager responsible for routing
messages to their destination. It also supports transactions, namely transaction delivery
using the Distributed Transaction Coordinator (DTC) as the transaction manager.
DTC is a general-purpose transaction manager architected originally to supply two-phase
commit to Microsoft's SQL Server relational database. DTC has been extended to provide
general-purpose interfaces capable of providing transaction management for
program-to-program interactions under MTS. Operating as runtime middleware environment
within DCOM, DTC hosts application program modules and endows them with transaction
semantics.
COM+ and Active Server
COM+ supports the development of platform-independent advanced applications over standard
protocols such as TCP/IP. The primary goal of COM+ is to provide additional tools to the
COM/DCOM infrastructure to ease management of object technology development, further
advancing the COM/DCOM model into a model more generically aligned to other software
component models. COM+ provides security and memory management as well as a data-binding
feature that allows binding between object fields and specific database fields. COM+
enables components to redirect their service requests dynamically by calling various
services at runtime rather than being bound to a single implementation of service. In
addition to receiving and processing events related to instance creation, calls and
returns, and errors and instance deletion, interceptors also provide the mechanism to
implement transactions and system monitoring, further intertwining the COM/DCOM and MTS
architecture.
Excerpted with permission from
Technology Forecast: 1998.
Courtesy: PriceWaterhouse Coopers.