Download Middleware in Distributed Systems: Network Coordination & Scalability and more Study Guides, Projects, Research Communication in PDF only on Docsity!
Middleware: Past and Present a Comparison
Hennadiy Pinus
ABSTRACT
The construction of distributed systems is a difficult task for programmers, which can be simplified with the use of middleware. Middleware can be seen as a layer between ap- plications and operating systems. There are four main types of middleware: transactional, message-oriented, procedural and object-oriented middleware. The first three types of middleware can be seen as a past and the last one as a present middleware. The variety of middleware types leads to a choice problem for the software developers. In order to solve it, we research the classification of the middleware, pro and contra of each middleware type. This review is based on analyzing of middleware through the following requirements, proposed by [9]: network communication, coordination, re- liability, scalability and heterogeneity. During this paper we try to find out, how good the middleware types support these requirements.
Keywords
RPC, MOM, marshalling, TP, middleware
1. INTRODUCTION
Nowadays a great number of transactions is realized via in- ternet and networks. The ”networks were introduced to connect PCs, workstations and mainframes, and there was a strong technical drive from the software and computer industry towards distributed computing.” (see [8], p.1) Dis- tributed system(DS) is ”a computer system, where compo- nents of the system are held on physically separated, au- tonomous computers.” (see [14], p.1) These computers are connected by a communication network, which enables the integration of components. The time available for an in- tegration of the components is often too short, that’s why DS seems to be a better choice than building a new sys- tem. The integrated components may include existing or old components (legacy components). However, it is con- siderably more difficult to construct a DS, than to build a centralized or completely new system. In a DS components
have to communicate with each other on the multiple points via a network, thus increasing the likelihood of errors. The solution of the component integration into a DS can be found by using middleware [9, 14]. This paper is structured in the following way. At first we discuss the definition of the middleware (Section 2). Then we define the requirements that middleware has to support (Section 3). The classification of middleware is presented as next. For each type of the middleware we analyze, how it ful- fils the requirements (Section 4). At the end of the work we make comparative conclusions concerning middleware types (Section 5).
2. WHAT IS MIDDLEWARE?
The term middleware first appeared in the late 1980s. It was used to describe network connection management software. In the middle 1990s, when network technology had achieved sufficient penetration and visibility, it found its widespread [4]. There are many definitions of middleware found in liter- ature. In common, middleware can be defined as a software layer located above the OS and networking software and below applications. (see Figure 1) Middleware enables the
Figure 1: Middleware in Distributed System, taken from [7], p. 89
interaction and communication between different applica- tions through Application Programming Interfaces (APIs) across the distributed components [7, 8, 13]. API ”is the interface by which an application program accesses operat- ing system and other services”. (see [1]) Middleware can be seen as a common development and runtime environment
that enables the connection of components written in dif- ferent languages and running on different operating systems [14]. The goal of middleware is to simplify the construc- tion of the DS, where the application engineers can ”ab- stract from the implementation of low-level details, such as concurrency control, transaction management and network communication.” (see [9], p.120) Due to middleware devel- opers can concentrate on application requirements.
3. MIDDLEWARE REQUIREMENTS
The following requirements were proposed by [9]: network communication, coordination, reliability, scalability and het- erogeneity. Network Communication: Distributed systems need a network communication, because the components are lo- cated on different hosts. The network communication is based on network protocols. The communication requires the transformation of the complex data structures into a suitable format, which can be transmitted using transport protocols. These transformations are called marshalling (com- ponent requesting service passes parameters to a message request) and unmarshalling. To enable automatical (un-) marshalling it is required, that all data involved in a request have to be described. It can be done with an Interface Def- inition Language (IDL) [4, 9]. Coordination: Coordination is required to control mul- tiple communication points, which exist in distributed sys- tems. There are several mechanisms, which can influence co- ordination of the components in distributed system. These include synchronization and activation (deactivation) poli- cies. Synchronization is useful during the communication of the concurrent components on the same host. There are sev- eral ways to achieve synchronization. The one way, where a component is blocked till the other component completes execution of a requested service, is called synchronous. If the component that asks some service from another com- ponent remains unblocked and can continue to perform its operations, then this way is called deferred synchronous (if service request was initiated by client) or asynchronous (if initiated by server). Other coordination mechanisms are activation and deactivation policies. Activation (deactiva- tion) allows to start (end) a component independently from the applications, which it executes. One of the reasons to use these policies is that ”components may be idle for long periods, thus wasting resources if they were kept in virtual memory all the time.” (see [9], p.121) The middleware needs to support mechanisms called threading policies in order to control, how the server reacts on the concurrent requests [9]. Reliability: There are several types of reliability proposed in the literature: best effort, at-most-once, at least-once and exactly-once. ”Best effort service requests do not give any assurance about the execution of the request. At-most-once requests are guaranteed to execute only once. It may hap- pen that they are not executed, but then the requester is notified about the failure. At-least-once service requests are guaranteed to be executed, possibly more than once. The highest degree of reliability is provided by exactly-once re- quests, which are guaranteed to be executed once and only once.” (see [9], p.121) The usage of the definite network protocol can also influence reliability, because different pro- tocols have different reliability levels. Distributed system implementations need to include error detection and correc- tion mechanisms to liquidate the unreliabilities caused by
errors. If some components in the system are not available, then the reliability of the system suffers. The increase of the reliability in such situation is possible, if using a replica com- ponent. Therefore, replication of the components increases reliability. The programmers, however, need to know that increase of performance decreases reliability. It is a trade-off problem [9]. Scalability: Scalability defines, how well a hardware or a software system can adapt to increased demands [2]. The limited scalability in centralized systems can be overcome by DSs. The main task is to provide the changes in DS with- out changing its architecture or design. In order to achieve this task, it is desired that middleware respects different dimensions of transparency specified by the ISO Open Dis- tributed Processing (ODP) reference model. For example, access transparency means that a component can access the services of the remote component, as if it were local. Loca- tion transparency means that the components are not aware of the physical location of the interacted components. Mi- gration transparency allows components to change their lo- cation, which is not seen for a client requesting these com- ponents. Replication transparency means that requesting components don’t care about the place of the needed ser- vice. It can either be main component or its replica. Load balancing, where requests are forwarded to a replica in order to release a loaded server, can use replication mechanisms. Hence, ”components whose services are in high demand may have to exist in multiple copies.” (see [9], p.122) A scalable system middleware needs to support access, location, migra- tion and replication transparency [9]. Heterogeneity: Heterogeneity is the quality of being di- verse and not comparable in kind [1]. The components of distributed systems can be of different types (legacy or new components) and written in different languages. For exam- ple, legacy components ”tend to be written in imperative languages, such as COBOL, PL/I or C, newer components are often implemented using object-oriented programming languages.” (see [9], p.122) There are different dimensions of heterogeneity in DSs: ”hardware and operating system platforms, programming languages and indeed the middle- ware itself.” (see [9], p.122) These differences need to be resolved by the middleware. In the next section we’ll see, how good the different middle- ware types support the defined requirements.
4. MIDDLEWARE CLASSIFICATION
In the literature one can found four main types of mid- dleware. These are: transactional, procedural, message- oriented and object-oriented middleware [5, 6, 9, 14]. The first three types of middleware can be seen as a past and the last one as a present middleware.
4.1 Transactional Middleware (TM)
Transactional middleware(TM) or transaction processing (TP) monitors were designed in order to support distributed syn- chronous transactions. The main function of a TP monitor is a coordination of requests between clients and servers that can process this requests. Request is a ”message that asks the system to execute a transaction.” (see [16], p.94) TM uses clustering (”a grouping of a number of similar things” [1]) of the service requests into transactions. ”A transac- tion must support ACID properties - Atomic, Consistent, Isolated, and Durable. Atomic means ”all or nothing” qual-
sistent queues are to choose when the reliability is more im- portant than performance, like in banking fund transfer. To increase reliability message queuing supports different Qual- ity of Service (QoS) [4, 5, 9]. These QoS are defined in [4] as: ”1. Reliable message delivery - during exchange of mes- sages no network packets are lost. 2. Guaranteed message delivery - messages are delivered to the destination node ei- ther immediately (with no latency - network is available), or eventually (with latency - when the network is unavailable). In the latter case, middleware guarantees that messages are delivered as long as the network becomes available within a specified time period. 3. Assured, non-duplicate message delivery - if the messages are delivered, they are delivered only once.” (see [4], p.8) Scalability: ”The publish-subscribe communications model provides location transparency, allowing a program to send the message with a subject as the destination property while the middleware routes the message to all programs that have subscribed to that subject.” (see [5], p.7) Although location transparency is supported, MOMs have a limited support for access transparency. It happens, because queues are used for remote and not for local communication. If ac- cess transparency doesn’t exist, it leads to a lack of migra- tion and replication transparency, thus complicating scal- ability. ”Moreover, queues need to be set up by adminis- trators and the use of queues is hard-coded in both client and server components, which leads to rather inflexible and poorly adaptable architectures.” (see [9], p.123) Heterogeneity: The support of data heterogeneity is rather limited, because marshalling is not automatically generated, and needs to be implemented by programmers [9]. Advantages: 1. MOM supports group communication, which is atomic. Either all clients receive a delivery or none. That’s why a process doesn’t have to worry about what to do, if some clients don’t receive a message. 2. The use of persistent queues increases reliability in MOM products. 3. Support for transactional message queues is included in most MOM products, meaning that advanced delivery guarantees are supported. 4. MOM supports more network protocols than RPCs do [10, 16]. 5. MOM can send the message exactly-once due to QoS, thus increasing it’s reliability. Disadvantages: 1. MOM has limited scalability and het- erogeneity support [9]. 2. There is a bad portability sup- port, because MOM products don’t support any standards. Applications that are made for one MOM product are not compatible to another MOM product [4]. Where to use: MOM can be used in the applications, where the network or all-components availability is not war- ranted [4].
4.3 Procedural Middleware (PM)
Remote Procedure Calls (RPCs) were developed by Sun Mi- crosystems in the early 1980s. RPCs are represented on different operating systems, including most Unix and MS Windows systems. ”Windows NT, for example, supports lightweight RPCs across processes and, with DCOM, full RPCs.” (see [15], p.172) Typical products: Open Software Foundation’s Distributed Computing Environment DCE, Microsoft RPC Facility [9]. Network Communication: RPCs define server compo- nents as RPC programs. An RPC program contains param- eterized procedures. Remote clients can invoke these proce- dures across the network using the network protocols. These
protocols are low-level, such as Transmission Control Proto- col/Internet Protocol (TCP/IP) or User Datagram Protocol (UDP). The communication happens in the following way. If a client wants to receive some services, then it makes a re- quest to a server. This request consists of a message, which includes the marshalled parameters. On the other side, the server receives this message, unmarshalls the parameters, executes the requested service and sends the result back to the client [9, 13]. ”In order to connect the client and server components of a distributed application using RPCs as the middleware link, it is required that every function that a client application can call should be represented by a stub, i.e. a placeholder, for the real function on the server.” (see [4], p.5) We can note, that RPC is the category of mid- dleware, where marshalling and unmarshalling are imple- mented automatically by the IDL compiler in stubs, hence making the life easier for developers [9, 13, 14]. Coordination: ”Typical RPC-based communication is syn- chronous, i.e., an RPC client is blocked until the remote pro- cedure has been executed or an error occurs.” (see [10], p.19) RPCs don’t support asynchronous communication. Proce- dural middleware provides different forms of activation poli- cies. These include activation of RPC on demand and ”RPC is always available”. Reliability: Procedural middleware possesses an at-most once reliability. If RPC fails, then an exception is returned. RPCs’ communication, based on TCP/IP protocol, can be seen as reliable, because TCP/IP provides reliability [9, 12, 14]. ”An application that uses TCP knows that data it sends is received at the other end, and that it is received correctly.” (see [12], p.16) Scalability: The scalability of RPCs remains rather lim- ited, because RPCs lack replication mechanisms. As men- tioned above the communication is based on stubs, which provide location transparency of the requested service to the client. It means, that client can invoke a remote procedure as if it were local [10]. As we know from the requirements section location transparency is a prerequisite to scalability [9]. Heterogeneity: The heterogeneity in RPCs is realized via IDL. It can define interfaces that represent relations between servers and clients. IDL is programming language indepen- dent, which means that client doesn’t need to know the lan- guage that server supports, if IDL compiler can translate the client’s request to a server [4]. Advantages: 1. RPC has a good heterogeneity support, because ”RPC has bindings for multiple operating systems and programming languages.” (see [14], p.3) 2. Marshalling and unmarshalling are automatically generated, thus sim- plifying the development of DSs [14]. Disadvantages: 1. RPCs don’t support group communica- tion [16]. 2. They have no direct support for asynchronous communication, replication and load balancing, therefore leading to a limited scalability. 3. Fault tolerance is worse than by other middleware types, because ”many possible faults have to be caught and dealt with in the program.” (see [14], p.4) Where to use: According to [5], RPCs could be used in small, simple applications with primarily point-to-point communication. ”RPCs are not a good choice to use as the building blocks for enterprise-wide applications where high performance and high reliability are needed.” (see [5], p.6)
4.4 Object-oriented Middleware (OOM)
Object middleware, evolved from RPCs, extends them by adding object-oriented concepts. These concepts are: inher- itance, object references and exceptions. OOM allows ref- erencing of remote objects and can call operations on them [9, 10, 14]. Typical products: OMG’s CORBA, Microsoft COM, Java RMI and Enterprise Java Beans [5, 9, 10]. Network Communication: Object middleware supports distributed object requests. It is possible for a client to re- quest an operation on a server object on the other host. The one thing required for a client is a reference to the server ob- ject. Marshalling of the parameters for the network request is made automatically in the client and server stubs [9]. Coordination: OOM generally supports a synchronous communication. The client object remains blocked, wait- ing for the server object response. It doesn’t mean that other synchronization types are not supported. ”CORBA 3.0, for example, supports both deferred synchronous and asynchronous object requests.” (see [9], p.125) OOM sup- ports different activation and threading policies. In OOM server objects can be active all the time or started on de- mand. CORBA Concurrency Service enables threading and coordinates the concurrent access to shared resources. It also guarantees the consistency of the object, if it is ac- cessed by multiple clients [4, 9, 13]. One should mention that CORBA ”supports group communication through its Event and Notification services.” (see [9], p.125) Reliability: At most once reliability is set by default in OOM. To handle failures during component requests OOM uses exceptions. Noticeable is the fact that due to CORBA messaging exactly-once reliability can be achieved. Fault- Tolerant CORBA, for example, provides extra reliability in DS [9, 10, 14]. Scalability: Scalability remains limited. Some CORBA implementations support load-balancing. Enterprise Java Beans have a replication support, which increases the scal- ability [9]. Heterogeneity: OOM has a wide support for heterogene- ity. For example, ”CORBA and COM both have multiple programming language bindings so that client and server objects do not need to be written in the same programming language.” (see [9], p.123) Java/RMI resolves the hetero- geneity with its Java Virtual Machine [9]. Advantages: 1. Marshalling and unmarshalling are gener- ated automatically in client and server stubs. 2. OOM sup- ports both synchronous and asynchronous communication.
- Most OOM products have a support for messaging and transactions. Hence, OOM can replace other three types of middleware in many different aspects. It makes OOM to a powerful and flexible middleware type [10, 14]. Disadvantages: 1. Lack of scalability. Where to use: OOM ”should be considered for applica- tions where immediate scalability requirements are some- what limited. These applications should be part of a long- term strategy towards object orientation.” (see [4], p.22)
5. CONCLUSIONS
As we have seen, the variety of middleware types can be used in different situations (see ”Where to use” subsections). Some of them can be combined to achieve the needed re- quirements. We can note, that all types of middleware have a limited support for scalability. The most reliable
are TM and OOM. TP monitors perform better than MOM and RPCs. RPCs and OOM support an automatical mar- shalling, thus simplifying the network communication. Un- like MOM and OOM, RPC and TM don’t support group communication. The use of IDL in RPCs and OOM makes them very heterogenous. Noticeable, that all types of mid- dleware support the activation on demand, thus saving re- sources. OOM integrates the benefits of TM, MOM and RPCs and tends to replace them. At last, the choice of the middleware depends on the task to realize.
6. REFERENCES
[1] http://www.hyperdictionary.com.
[2] http://www.webopedia.com.
[3] http://www.cs.panam.edu/∼meng/Course/CS6334/ Note/master/node88.html.
[4] Middleware white paper, 1997. http://web.cefriel. it/∼alfonso/WebBook/Documents/isgmidware.pdf.
[5] Talarian: Everything you need to know about middleware, 2000. http://searchwebservices.techtarget.com/ searchWebServices/downloads/Talarian.pdf.
[6] D. E. Bakken. Middleware. http://www.eecs.wsu. edu/∼bakken/middleware-article-bakken.pdf.
[7] P. A. Bernstein. Middleware: a model for distributed system services. Communications of the ACM, pages 86 – 98, 1996.
[8] R. P. Bob Hulsebosch, Wouter Teeuw. Middleware tintel state-of-the-art deliverable. 1999.
[9] W. Emmerich. Software engineering and middleware: A roadmap. Communications of the ACM, pages 117 – 129, 2000.
[10] C. Hartwich. A middleware architecture for transactional, object-oriented applications, 2003. http://page.mi.fu-berlin.de/∼hartwich/ diss-final.pdf. [11] D. S. Linthicum. Application servers an eai. eAI Journal, July/August 2000.
[12] J. Moss. Understanding tcp/ip. PC Network Advisor,
[13] J. M. Myerson. The Complete Book of Middleware. Auerbach Publications, 2002.
[14] R. Nunn. Distributed software architectures using middleware. http://www.cs.ucl.ac.uk/staff/W. Emmerich/lectures/3C05-02-03/aswe18-essay.pdf.
[15] C. Szyperski. Component Software. Addison Wesley,
[16] A. S. Tanenbaum. Distributed Operating Systems. Prentice Hall, 1995.
[17] T. C. E. A. Workgroup. Middleware architecture report:, 2001. http://www.vita.virginia.gov/docs/ ea/MiddlewareArchitectureV1-0-051801.pdf.
All URLs, mentioned in references, were last accessed on the 20.06.04.
