Avraham Leff

Replica control in distributed systems: as asynchronous approach

An asynchronous approach is proposed for replica control in distributed systems. This approach applies an extension of serializability called epsilon-serializability (ES R), a correctness criterion which allows temporary and bounded inconsistency in replicas to be seen by queries. Moreover, users can reduce the degree of inconsistency to the desired amount. In the limit, users see strict l-copy serializability. Because the system maintains ESR correctness (1) replicas always converges to global serializability and (2) the system permits read access to object replicas before the system reaches a quiescent state. Various replica control methods that maintain ESR are described and analyzed. Because these methods do not require users to refer explicitly to ESR criteria, they can be easily encapsulated in high-level applications that use replicated data.

Web-application development using the Model/View/Controller design pattern

The Model/View/Controller design pattern is very useful for architecting interactive software systems. This design pattern is partition-independent, because it is expressed in terms of an interactive application running in a single address space. Applying the Model/View/Controller design pattern to web-applications is therefore complicated by the fact that current technologies encourage developers to partition the application as early as in the design phase. Subsequent changes to that partitioning require considerable changes to the applications implementation-despite the fact that the application logic has not changed. This paper introduces the concept of Flexible Web-Application Partitioning, a programming model and implementation infrastructure, that allows developers to apply the Model/View/Controller design pattern in a partition-independent manner Applications are developed and tested in a single address-space; they can then be deployed to various client/server architectures without changing the applications source code. In addition, partitioning decisions can be changed without modifying the application.

Replication algorithms in a remote caching architecture

Studies the cache performance in a remote caching architecture. The authors develop a set of distributed object replication policies that are designed to implement different optimization goals. Each site is responsible for local cache decisions, and modifies cache contents in response to decisions made by other sites. The authors use the optimal and greedy policies as upper and lower bounds, respectively, for performance in this environment. Critical system parameters are identified, and their effect on system performance studied. Performance of the distributed algorithms is found to be close to optimal, while that of the greedy algorithms is far from optimal. >

Service-level agreements and commercial grids

Service-level agreements impose unique requirements on commercial grid infrastructures - specifically, they emphasize the need for a dynamic offload infrastructure.

LAN and I/O convergence: a survey of the issues

Local area networks and computer I/O are both interconnects that move information from one location to another. Despite this shared purpose, LANs have traditionally connected independent and widely separated computers. In contrast, computer I/O has traditionally connected a host to peripheral devices such as terminals, disks, and tape drives. Because these connection tasks were different, the architectures developed for one task were not suitable for the other. Consequently, the technologies used to implement one architecture could not address the issues faced by the other, and the technologies were seen as fundamentally different. However, an examination of the architectural requirements of modern I/O and LANs shows that the differences between the two technologies are now disappearing. We believe that LAN and I/O architectures are in fact converging, and that this convergence reflects significant changes in how and where computing resources are used. To illustrate this convergence and its implications, this article examines several modern LANs and channels.<<ETX>>

Policies for efficient memory utilization in a remote caching architecture

The high performance networks in modern distributed systems enable a site to obtain cached objects from the main memory of other sites more quickly than the time needed to access local disks. In this environment, efficient mechanisms can be devised to support rapid request/response exchanges for objects that reside on remote sites. As a result, it becomes possible to use remote memory as an additional layer in the memory hierarchy between local memory and disks. The paper studies the performance potential of remote memory caching. A key issue is managing the of object replicas maintained by different sites. Although exploiting remote memory improves performance over a wide range of cache sizes as compared to a distributed client/server caching architecture, efficient use of remote memory can be difficult to achieve. The effect of several system parameters, and several cache policies are examined; their effect on both overall and local system performance is described.<<ETX>>

Autonomous transaction execution with epsilon serializability

The authors study the feasibility of autonomous transaction execution in systems with asynchronous transaction processing based on epsilon serializability (ESR). The abstract correctness criteria defined by ESR are implemented by techniques such as asynchronous divergence control and asynchronous consistency restoration. Concrete application examples in a distributed environment, such as banking, illustrate the advantages of using ESR to support execution autonomy. The ability for asynchronous transaction processing also supports efficient concurrent transaction and query processing.<<ETX>>

Efficient LRU-based buffering in a LAN remote caching architecture

The possibility of fast access to the main memory of remote sites has been advanced as a potential performance improvement in distributed systems. Even if a page is not available in local memory, sites need not do a disk access. Instead, the sites can use efficient mechanisms that support rapid request/response exchanges in order to access pages that are currently buffered at a remote site. Hardware and software support in such a remote caching architecture must also include algorithms that determine which pages should be buffered at what sites. When each site uses the classic LRU replacement algorithm, performance can be much worse than optimal in many system configurations. Because sites do not coordinate individual decisions, overall system buffering/caching decisions yield very inefficient global configurations. This paper proposes an easily implementable modification of the LRU replacement algorithm for LAN environments that reduces replication. The algorithm substantially improves hit-ratios-and thus performance-over a wide range of parameters. The relatively simple LAN topology implies that much less state information need be available for good replacement decisions compared to general network topologies. Two implications of two variations of the algorithm are explored. In an environment where the network is not a performance bottleneck, and where performance is memory-limited, performance of the proposed replacement algorithm is shown to be close to optimal.

A comparison of fibre channel and 802 MAC services

As the differences between channels and networks narrow, the services and capabilities of channels need to be examined relative to those of networks. A comparison is made between the Fibre Channel protocol and networking protocols. Specific features of the Fibre Channel protocol are compared with those of the Fiber Distributed Data Interface (FDDI) LAN protocol and 802.2-type media access controllers (MACs). Fibre Channel may, in the future, fill a role where it serves as a single communication medium providing both channel and network services.

On robust transaction routing and load sharing

In this paper we examine the issue of robust transaction routing in a locally distributed database environment where transaction characteristics such as reference locality imply that certain processing systems can be identified as being more suitable than others for a given transaction class. A response time based routing strategy can strike a balance between indiscriminate sharing of the load and routing based only on transaction affinity. Since response time estimates depend on workload and system parameters that may not be readily available, it is important to examine the robustness of routing decisions to information accuracy. We find that a strategy which strictly tries to minimize the response time of incoming transactions is sensitive to the accuracy of certain parameter values. On the other hand, naive strategies, that simply ignore the parameters in making routing decisions, have even worse performance. Three alternative strategies are therefore examined: threshold, discriminatory, and adaptive. Instead of just optimizing an incoming transactions response time, the first two strategies pursue a strategy that is somewhat more oriented towards global optimization. This is achieved by being more restrictive on either the condition or the candidate for balancing the load. The third strategy, while trying to minimize the response time of individual incoming transactions, employs a feedback process to adaptively adjust future response time estimates. It monitors the discrepancy between the actual and estimated response times and introduces a correction factor based on regression analysis. All three strategies are shown to be robust with respect to the accuracy of workload and system parameters used in the response time estimation.