Daniel M. Dias

A scalable and highly available web server

We describe a prototype scalable and highly available web server, built on an IBM SP-2 system, and analyze its scalability. The system architecture consists of a set of logical front-end or network nodes and a set of back-end or data nodes connected by a switch, and a load balancing component. A combination of TCP routing and Domain Name Server (DNS) techniques are used to balance the load across the Front-end nodes that run the Web (httpd) daemons. The scalability achieved is quantified and compared with that of the known DNS technique. The load on the back-end nodes is balanced by striping the data objects across the back-end nodes and disks. High availability is provided by detecting node or daemon failures and reconfiguring the system appropriately. The scalable and highly available web server is combined with parallel databases, and other back-end servers, to provide integrated scalable and highly available solutions.

Analysis of task assignment policies in scalable distributed web-server systems

A distributed multiserver Web site can provide the scalability necessary to keep up with growing client demand at popular sites. Load balancing of these distributed Web-server systems, consisting of multiple, homogeneous Web servers for document retrieval and a Domain Name Server (DNS) for address resolution, opens interesting new problems. In this paper, we investigate the effects of using a more active DNS which, as an atypical centralized scheduler, applies some scheduling strategy in routing the requests to the most suitable Web server. Unlike traditional parallel/distributed systems in which a centralized scheduler has full control of the system, the DNS controls only a very small fraction of the requests reaching the multiserver Web site. This peculiarity, especially in the presence of highly skewed load, makes it very difficult to achieve acceptable load balancing and avoid overloading some Web servers. This paper adapts traditional scheduling algorithms to the DNS, proposes new policies, and examines their impact under different scenarios. Extensive simulation results show the advantage of strategies that make scheduling decisions on the basis of the domain that originates the client requests and limited server state information (e.g., whether a server is overloaded or not). An initially unexpected result is that using detailed server information, especially based on history, does not seem useful in predicting the future load and can often lead to degraded performance.

SP2 system architecture

Scalable parallel systems are increasingly being used today to address existing and emerging application areas that require performance levels significantly beyond what symmetric multiprocessors are capable of providing. These areas include traditional technical computing applications, commercial computing applications such as decision support and transaction processing, and emerging areas such as “grand challenge” applications, digital libraries, and video production and distribution. The IBM SP2™ is a general-purpose scalable parallel system designed to address a wide range of these applications. This paper gives an overview of the architecture and structure of SP2, discusses the rationale for the significant system design decisions that were made, indicates the extent to which key objectives were met, and identifies future system challenges and advanced technology development areas.

Buffering and caching in large-scale video servers

Video-on-demand servers are characterized by stringent real-time constraints, as each stream requires isochronous data playout. The capacity of the system depends on the acceptable jitter per stream (the number of data blocks that do not meet their real-time constraints). Per-stream read-ahead buffering avoids the disruption in playback caused by variations in disk access time and queuing delays. With heavily skewed access patterns to the stored video data, the system is often disk arm-bound. In such cases, serving video streams from a memory cache can result in a substantial reduction in server cost. In this paper, we study the cost-performance trade-offs of various buffering and caching strategies that can be used in a large-scale video server. We first study the cost impact of varying the buffer size, disk utilization and the disk characteristics on the overall capacity of the system. Subsequently, we study the cost-effectiveness of a technique for memory caching across streams that exploits temporal locality and workload fluctuations.

High performance Web site design techniques

This article presents techniques for designing Web sites that need to handle large request volumes and provide high availability. The authors present new techniques they developed for keeping cached dynamic data current and synchronizing caches with underlying databases. Many of these techniques were deployed at the official Web site for the 1998 Olympic Winter Games.

A modeling study of the TPC-C benchmark

The TPC-C benchmark is a new benchmark approved by the TPC council intended for comparing database platforms running a medium complexity transaction processing workload. Some key aspects in which this new benchmark differs from the TPC-A benchmark are in having several transaction types, some of which are more complex than that in TPC-A, and in having data access skew. In this paper we present results from a modelling study of the TPC-C benchmark for both single node and distributed database management systems. We simulate the TPC-C workload to determine expected buffer miss rates assuming an LRU buffer management policy. These miss rates are then used as inputs to a throughput model. From these models we show the following: (i) We quantify the data access skew as specified in the benchmark and show what fraction of the accesses go to what fraction of the data. (ii) We quantify the resulting buffer hit ratios for each relation as a function of buffer size. (iii) We show that close to linear scale-up (about 3% from the ideal) can be achieved in a distributed system, assuming replication of a read-only table. (iv) We examine the effect of packing hot tuples into pages and show that significant price/performance benefit can be thus achieved. (v) Finally, by coupling the buffer simulations with the throughput model, we examine typical disk/memory configurations that maximize the overall price/performance.

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.

Analysis of replication in distributed database systems

The authors develop an approximate analytical model to study the tradeoffs of replicating data in a distributed database environment. Several concurrency control protocols are considered, including pessimistic, optimistic, and semi-optimistic protocols. The approximate analysis captures the effect of the protocol on hardware resource contention and data contention. The accuracy of the approximation is validated through detailed simulations. It is found that the benefit of replicating data and the optimal number of replicates are sensitive to the concurrency control protocol. Under the optimistic and semi-optimistic protocols, replications can significantly improve response time with an additional MIPS (million instructions per second) requirement to maintain consistency among the replicates. The optimal degree of replication is further affected by the transaction mix (e.g. the fraction of read-only transactions), the communications delay and overhead, the number of distributed sites, and the available MIPS. Sensitivity analyses have been carried out to examine how the optimal degree of replication changes with respect to these factors. >

Scheduling algorithms for distributed Web servers

A distributed Web system, consisting of multiple servers for data retrieval and a Domain Name Server (DNS) for address resolution, can provide the scalability necessary to keep up with growing client demand at popular sites. However, balancing the requests among these atypical distributed servers opens interesting new challenges. Unlike traditional distributed systems in which a centralized scheduler has full control of the system, the DNS controls only a small fraction of the requests reaching the Web site. This makes it very difficult to avoid overloading situations among the multiple Web servers. We adapt traditional scheduling algorithms to the DNS, propose new policies, and examine their impact. Extensive simulation results show the advantage of using strategies that schedule requests on the basis of the origin of the clients and very limited state information, such as whether a server is overloaded or not. Conversely, algorithms that use detailed state information often exhibit the worst performance.

On coupling multi-systems through data sharing

The demand for larger transaction rates and the inability of single-system-based transaction processors to keep up with demand have resulted in the growth of multi-processor-based database systems. The focus here is on coupling in a locally distributed system through multi-system data sharing in which all systems have direct access to the data. This paper addresses the following questions; i) How does a workload running on a single system today perform if migrated to a multi-system? ii) What are the multi-system locking design issues that limit multi-system performance and what is the maximum number of systems that may be effectively coupled? iii) Can alternative locking designs increase the number of systems that may be effectively coupled? Our analysis is based on traces from large mainframe systems running IBMs IMS database management system. We have developed a hierarchical modeling methodology that starts by synthesizing a multi-system IMS lock trace and a reference trace from single-system traces. The multisystem traces are used in trace-driven simulations to predict lock contention and database I/O increase in multi-system environment and to generate workload parameters. These parameters are used in event-driven simulation models to examine the overall performance under different system structures. Performance results are presented for realistic system parameters to determine the performance impact of various design parameters. Lock contention is found to be the critical factor in determining the coupling effectiveness and the effect of alternative locking design to reduce lock contention is studied. The limit on coupling is explored and the analysis indicates that, for this workload, on the order of 6 to 12 systems may be effectively coupled through data sharing, depending on system structure and locking design.