Elizabeth Varki

Issues and challenges in the performance analysis of real disk arrays

The performance modeling and analysis of disk arrays is challenging due to the presence of multiple disks, large array caches, and sophisticated array controllers. Moreover, storage manufacturers may not reveal the internal algorithms implemented in their devices, so real disk arrays are effectively black-boxes. We use standard performance techniques to develop an integrated performance model that incorporates some of the complexities of real disk arrays. We show how measurement data and baseline performance models can be used to extract information about the various features implemented in a disk array. In this process, we identify areas for future research in the performance analysis of real disk arrays.

Mean value technique for closed fork-join networks

Elizabeth Varki Department of Computer Science University of New Hampshire email: varki@cs.unh.edu A simple technique for computing mean performance measures of closed single-class fork-join networks with exponential service time distribution is given here. This technique is similar to the mean value analysis technique for closed product-form networks and iterates on the number of customers in the network. Mean performance measures like the mean response times, queue lengths, and throughput of closed fork-join networks can be computed recursively without calculating the steady-state distribution of the network. The technique is based on the mean value equation for fork-join networks which relates the response time of a network to the mean service times at the service centers and the mean queue length of the system with one customer less. Unlike product-form networks, the mean value equation for fork-join networks is an approximation and the technique computes lower performance bound values for the fork-join network. However, it is a good approximation since the mean value equation is derived from an equation that exactly relates the response time of parallel systems to the degree of parallelism and the mean arrival queue length. Using simulation, it is shown that the relative error in the approximation is less than 5% in most cases. The error does not increase with each iteration. Permlsslon to make digital or hard copies of all or part of this work for personal or classroom use is granted without tee provided that copa are not made or distributed for pro10 01 commercial advantage and that copies bear this notice and the full cNa110n on the first Page. To copy otherwise, to republish. to post on servers Or t0 redwribute to lists. requires pnor specific permission and/or a fee. SIGMETRICS ‘99 5/99 Atlanta, Georgia. USA

Response time analysis of parallel computer and storage systems

Fork-join structures have gained increased importance in recent years as a means of modeling parallelism in computer and storage systems. The basic fork-join model is one in which a job arriving at a parallel system splits into K independent tasks that are assigned to K unique, homogeneous servers. In the paper, a simple response time approximation is derived for parallel systems with exponential service time distributions. The approximation holds for networks modeling several devices, both parallel and nonparallel. (In the case of closed networks containing a stand-alone parallel system, a mean response time bound is derived.) In addition, the response time approximation is extended to cover the more realistic case wherein a job splits into an arbitrary number of tasks upon arrival at a parallel system. Simulation results for closed networks with stand-alone parallel subsystems and exponential service time distributions indicate that the response time approximation is, on average, within 3 percent of the seeded response times. Similarly, simulation results with nonexponential distributions also indicate that the response time approximation is close to the seeded values. Potential applications of our results include the modeling of data placement in disk arrays and the execution of parallel programs in multiprocessor and distributed systems.

An integrated performance model of disk arrays

All enterprise storage systems depend on disk arrays to satisfy their capacity, reliability, and availability requirements. Performance models of disk arrays are useful in understanding the behavior of these storage systems and predicting their performance. We extend prior disk array modeling work by developing an analytical disk array model that incorporates the effects of workload sequentiality, read-ahead caching, write-back caching, and other complex optimizations incorporated into most disk arrays. The model is computationally simple and scales easily, making it potentially useful to performance engineers.

Sequential Prefetch Cache Sizing for Maximal Hit Rate

We propose a prefetch cache sizing module for use with any sequential prefetching scheme and evaluate its impact on the hit rate. Disk array caches perform sequential prefetching by loading data contiguous to I/O request data into the array cache. If the I/O workload has sequential locality, then data prefetched in response to sequential accesses in the workload will receive hits. Different schemes prefetch different data, so the prefetch cache size requirement varies. Moreover, the proportion of sequential and random requests in the workload and their interleaving pattern affects the size requirement. If the cache is too small, then prefetched data would get evicted from the cache before a request for the data arrives, thus lowering the hit rate. If the cache is too large, then valuable cache space is wasted. We present a simple sizing module that can be added to any prefetching scheme to ensure that the prefetch cache size is adequately matched to the requirement of the prefetching scheme on a dynamic workload comprising multiple streams. We analytically compute the maximal hit rate achievable by popular prefetching schemes and through simulations, show that our sizing module maintains the prefetch cache at a size that nearly achieves this maximal hit rate.

Improve Prefetch Performance by Splitting the Cache Replacement Queue

The performance of a prefetch cache is dependent on both the prefetch technique and the cache replacement policy. Both these algorithms execute independently of each other, but they share a data structure - the cache replacement queue. This paper shows that even with a simple prefetch technique, there is an increase in hit rate when the LRU replacement queue is split into two equal sized queues. A more significant performance improvement is possible with a sophisticated prefetch technique and by splitting the queue unequally.

The Feasibility of Moving Terabyte Files between Campus and Cloud

Cloud/Grid computing is envisioned to be a predominant computing model of the future. The movement of files between cloud and client is intrinsic to this model. With the creation of ever expanding data sets, the sizes of files have increased dramatically. Consequently, terabyte file transfers are expected to be the “next big” Internet application. This application is different from other Internet applications in that it requires extensive bandwidth, orders of magnitude larger than the bandwidth requirements of existing applications. It is essential to determine whether or not existing network infrastructures can handle the augmented workload that terabyte transfers would create. This is particularly critical for academic campus networks that are already under strain from high user demand. The paper evaluates the system level challenges of incorporating terabyte transfers into an existing campus network. The evaluation finds that large file transfers can be handled by the current campus network without making major changes to the infrastructure. It is vital to employ a system level service that schedules and monitors the terabyte transfers on users’ behalf. By removing control from users, the service is able to leverage low demand periods and dynamically repurpose unused bandwidth.

Co-allocation in data grids: a global, multi-user perspective

Several recent studies suggest that co-allocation techniques can improve user performance for distributed data retrieval in replicated grid systems. These studies demonstrate that co-allocation techniques can improve network bandwidth and network transfer times by concurrently utilizing as many data grid replicas as possible. However, these prior studies evaluate their techniques from a single users perspective and overlook evaluations of system wide performance when multiple users are using co-allocation techniques. In our study, we provide multi-user evaluations of a co-allocation technique for replicated data in a controlled grid environment. We find that co-allocation works well under low-load conditions when there are only a few users using co-allocation. However, co-allocation works very poorly for medium and high-load conditions since the response time for co-allocating users grows rapidly as the number of grid users increases. The decreased performance for co-allocating users can be directly attributed to the increased workload that their greedy retrieval technique places on the replicas in the grid. Overall, we determine that uninformed, blind utilization of greedy co-allocation techniques by multiple users is detrimental to global system performance.

GPSonflow: Geographic Positioning of Storage for Optimal Nice Flow

This article evaluates the maximum data flow from a sender to a receiver via the internet when all transmissions are scheduled for early morning hours. The significance of early morning hours is that internet congestion is low while users sleep. When the sender and receiver lie in proximal time zones, a direct transmission from sender to receiver can be scheduled for early morning hours. When the sender and receiver are separated by several time zones such that their sleep times are non-overlapping, data can still be transmitted during early morning hours with an indirect store-and-forward transfer. The data are transmitted from the sender to intermediate end networks or data centers that serve as storage hops en route to receiver. The storage hops are placed in zones that are time proximal to the sender or the receiver so that all transmissions to and from storage hops occur during low-congestion early morning hours. This article finds the optimal locations and bandwidth distributions of storage hops for maximum nice internet flow from a sender to a receiver.

Where review goes wrong

Examining professional misconduct among academic publication examiners.