Analysis of simple randomized buffer management for parallel I/O

Proceedings. Fifth International Conference on High Performance Computing (Cat. No. 98EX238)

We address the problems of prefetching and I/O scheduling for read-once reference strings in a parallel I/O system. Read-once reference strings, in which each block is accessed exactly once, arise naturally in applications like databases and video retrieval. Using the standard parallel disk model with £ disks and a shared I/O buffer of size ¤ , we present a novel algorithm, Red-Black Prefetching (RBP), for parallel I/O scheduling. The number of parallel I/Os performed by RBP is within O(£ ¦ ¥ § ©) of the minimum possible. Algorithm RBP is easy to implement and requires computation time linear in the length of the reference string. Through simulation experiments we validated the benefits of RBP over simple greedy prefetching.

We present a simulation study of several prefetching policies to improve the I O performance of external merging using parallel I O. In particular we consider traditional sequential prefetch, forecast-based greedy prefetching, and oblivious prefetching. In conjunction with the prefetching policies we e v aluate the bene t of two di erent data placement strategies: run-level striping and block-random placement, in the presence of data skew. We show that the I O performance is greatly improved by using forecasting techniques. This method outperforms the other policies in achieving higher disk parallelism, and scales well with increased numbers of disks and increasing data skew. Additionally, the performance of block-random data placement i s s h o wn to be uniformly good, independent of the data skew, while the performance of run-level striped placement degrades with increasing skew.

Lecture Notes in Computer Science, 1998

We address the problem of I O scheduling of read-once reference strings in a multiple-disk parallel I O system. We present a n o vel online algorithm, Red-Black Prefetching RBP, for parallel I O scheduling. In order to perform accurate prefetching RBP uses L-block l o o k ahead. The performance of RBP is analyzed in the standard parallel disk model with D independent disks and a shared I O bu er of size M. W e show that the number of parallel I Os performed by RBP is within a factot maxf p M D = L ; D 1=3 g of the number of I Os done by the optimal oline algorithm. This ratio is within a canstant factor of the best possible when L is L = OM D 1=3 .

We consider the natural extension of the well-known single disk caching problem to the parallel disk I/O model (PDM) [17]. The main challenge is to achieve as much par-allelism as possible and avoid I/O bottlenecks. We are given a fast memory (cache) of size M memory blocks along with a request sequence Σ = (b1, b2, ..., bn) where each block bi resides on one of D disks. In each parallel I/O step, at most one block from each disk can be fetched. The task is to serve Σ in the minimum number of parallel I/Os. Thus, each I/O is analogous to a page fault. The difference here is that during each page fault, up to D blocks can be brought into memory, as long as all of the new blocks entering the memory reside on different disks. The problem has a long history [18, 12, 13, 26]. Note that this problem is non-trivial even if all requests in Σ are unique. This restricted version is called read-once. Despite the progress in the offline version [13, 15] and read-once version [12], the general online problem still remained open. Here, we provide comprehensive results with a full general solution for the problem with asymptotically tight competitive ratios. To exploit parallelism, any parallel disk algorithm needs a certain amount of lookahead into future requests. To provide effective caching, an online algorithm must achieve o(D) competitive ratio. We show a lower bound that states, for lookahead L ≤ M , any online algorithm must be Ω(D)-competitive. For lookahead L greater than M (1 + 1//), where is a constant, the tight upper bound of O(p M D/L) on competitive ratio is achieved by our algorithm SKEW. The previous algorithm tLRU [26] was O((M D/L) 2/3) competitive and this was also shown to be tight [26] for an LRU-based strategy. We achieve the tight ratio using a fairly * different strategy than LRU. We also show tight results for randomized algorithms against oblivious adversary and give an algorithm achieving better bounds in the resource augmentation model.

Conference on Parallel andDistributed Information Systems, 1991

Improvements in the processing speed of multiprocessorsare outpacing improvements in the speed ofdisk hardware. Parallel disk I/O subsystems have beenproposed as one way to close the gap between processorand disk speeds. In a previous paper we showedthat prefetching and caching have the potential to deliverthe performance benefits of parallel file systems toparallel applications. In this paper we describe experimentswith practical

All server storage environments depend on disk arrays to satisfy their capacity, reliability, and availability require- ments. In order to manage these storage systems efficiently, it is necessary to understand the behavior of disk arrays and predict their performance. We develop an analytical model that estimates mean performance measures of disk arrays under a synchronous I/O workload. Synchronous I/O requests are generated by jobs that each block while their request is serviced. Upon I/O service completion, a job may use other computer resources before issuing another I/O request. Our disk array model considers the effect of workload sequentiality, read-ahead caching, write-back caching, and other complex optimizations incorporated into most disk arrays. The model is validated against a mid-range disk-array for a variety of synthetic I/O workloads. The model is computationally simple and scales easily as the number of jobs issuing requests increases, making it potentially useful ...

DIMACS Series in Discrete Mathematics and Theoretical Computer Science, 1999

In this work we address the problems of prefetchingand I O scheduling for read-once reference strings in a parallel I O system. We use the standard parallel disk model with D disks a shared I O bu er of size M. W e design an on-line algorithm ASP Adaptive Segmented Prefetching with M L-block lookahead, L 1, and compare its performance to the best on-line algorithm with the same lookahead. We show that for any reference string the numberof I Os done by ASP is with a factor C, C = minf p L; D 1=3 g, of the number of I Os done by the optimal algorithm with the same amount o f l o o k ahead.

2007

Current disk prefetch policies in major operating systems track access patterns at the level of the file abstraction. While this is useful for exploiting application-level access patterns, file-level prefetching cannot realize the full performance improvements achievable by prefetching. There are two reasons for this. First, certain prefetch opportunities can only be detected by knowing the data layout on disk, such as the contiguous layout of file metadata or data from multiple files. Second, non-sequential access of disk data (requiring disk head movement) is much slower than sequential access, and the penalty for mis-prefetching a 'random' block, relative to that of a sequential block, is correspondingly more costly.

2015

Abstract: A general law is proposed that states "a large numbers of slower processors may be better than a small number of faster processors for I/O-dominant applications". The need for such a guideline is demonstrated because simple linear sums of individual processor performances do not provide an accurate estimation of I/O performance for a parallel computer. Furthermore, the law was formulated to allow better cost estimations when choosing the number and type processor for a Massively Parallel (MP) I/O application. The law is confirmed with a simple proof, analytical model, simulation, and benchmarking. A Distributed Cache Subsystem (DCS) technique is proposed to further improve the performance of the MP computers running I/O-dominant applications. Using simulations and benchmarks the DCS technique has shown the potential to achieve very high performance using standard sequential file systems. The general conclusion proposed is that a MP computer containing a small num...

Journal of Systems and Software, 2009

Anticipatory scheduling (AS) of I/O requests has become a viable choice for block-device schedulers in open-source OS-kernels as prior work has established its superiority over traditional disk-scheduling policies. An AS-scheduler selectively stalls the block-device right after servicing a request in hope that a new request for a nearby sector will be soon posted. Clearly, this decision may introduce delays if the anticipated I/O does not arrive on time. In this paper, we build on the success of the AS and propose an approach that minimizes the overhead of unsuccessful anticipations. Our suggested approach termed workload-dependent anticipation scheduling (WAS), determines the length of every anticipation period in an on-line fashion in order to reduce penalties by taking into account the evolving spatio-temporal characteristics of running processes as well as properties of the underlying computing system. We harvest the spatio-temporal features of individual processes and employ a system-wide process classification scheme that is re-calibrated on the fly. The resulting classification enables the disk scheduler to make informed decisions and vary the anticipation interval accordingly, on a per-process basis. We have implemented and incorporated WAS into the current Linux kernel. Through experimentation with a wide range of diverse workloads, we demonstrate WAS benefits and establish reduction of penalties over other AS-scheduler implementations.

Distributed and Parallel Databases, 1993

Improvements in the processing speed of multiprocessors are outpacing improvements in the speed of disk hardware. Parallel disk I/O subsystems have been proposed as one way to dose the gap between processor and disk speeds. In a previous paper we showed that prefetching and caching have the potentT"al to deliver the performance benefits of parallel file systems to parallel applications. In this paper we describe experiments with practical prefetching policies that base decisions only on on-line reference history, and that can be implemented efficiently. We also test the ability of those policies across a range of architectural parameters.

2001

Motivation t SSP goal: develop analytic models of storage devices to predict workload performance -Finding: even moderately sophisticated models give insufficient accuracy. Why?

1995

A general law is proposed that states "a large numbers of slower processors may be better than a small number of faster processors for I/O-dominant applications". The need for such a guideline is demonstrated because simple linear sums of individual processor performances do not provide an accurate estimation of I/O performance for a parallel computer. Furthermore, the law was formulated to allow better cost estimations when choosing the number and type processor for a Massively Parallel (MP) I/O application. The law is confirmed with a simple proof, analytical model, simulation, and benchmarking. A Distributed Cache Subsystem (DCS) technique is proposed to further improve the performance of the MP computers running I/O-dominant applications. Using simulations and benchmarks the DCS technique has shown the potential to achieve very high performance using standard sequential file systems. The general conclusion proposed is that a MP computer containing a small number of hig...

International Journal of Communications, Network and System Sciences, 2015

In this paper, we present a comparative study between informed and predictive prefetching mechanisms that were presented to leverage the performance gap between I/O storage systems and CPU. In particular, we will focus on transparent informed prefetching (TIP) and predictive prefetching using probability graph approach (PG). Our main objective is to show the main features, motivations, and implementation overview of each mechanism. We also conducted a performance evaluation discussion that shows a comparison between both mechanisms performance when using different cache size values.

2001