## Asynchronous Scheduling of Redundant Disk Arrays (2000)

### Download From

IEEE### Download Links

- [www.mpi-sb.mpg.de]
- [www.mpi-sb.mpg.de]
- [csis.bits-pilani.ac.in]
- [csdl.computer.org]
- DBLP

### Other Repositories/Bibliography

Venue: | In 12th ACM Symposium on Parallel Algorithms and Architectures |

Citations: | 13 - 5 self |

### BibTeX

@INPROCEEDINGS{Sanders00asynchronousscheduling,

author = {Peter Sanders},

title = {Asynchronous Scheduling of Redundant Disk Arrays},

booktitle = {In 12th ACM Symposium on Parallel Algorithms and Architectures},

year = {2000},

pages = {89--98}

}

### OpenURL

### Abstract

Allocation of data to parallel disk using redundant storage and random placement of blocks can be exploited to achieve low access delays. New algorithms are proposed which improve the previously known shortest queue algorithm by systematically exploiting that scheduling decisions can be deferred until a block access is actually started on a disk. These algorithms are also generalized for coding schemes with low redundancy. Using extensive simulations, practically important quantities are measured which have so far eluded an analytical treatment: The delay distribution when a stream of requests approaches the limit of the sytem capacity, the system efficiency for parallel disk applications with bounded prefetching buffers, and the combination of both for mixed traffic. A further step towards practice is taken by outlining the system design for a: automatically load-balanced parallel hard-disk array. Additional algorithmic measures are proposed for a that allow variable sized blocks, seek time reduction, fault tolerance, inhomogeneous systems, and flexible priorization schemes.

### Citations

1950 |
Sloane The Theory of Error-Correcting Codes
- MacWilliams, A
- 1977
(Show Context)
Citation Context ... explicitly scheduling write accesses using the same methods as described here for reading. A disadvantage of codes with w ? r + 1 is that they are computationally more expensive than parity-encoding =-=[34, 35, 36, 37, 38, 13]-=-. It makes sense to use different values for r and w concurrently on the same system for different purposes. Data which is rarely read and can be reconstructed otherwise, e.g., checkpointing data for ... |

1065 | Computer Networks - Tanenbaum - 2002 |

725 |
A Case for Redundant Arrays of Inexpensive Disks (RAID
- Patterson, Gibson, et al.
- 1988
(Show Context)
Citation Context ... the scheduling algorithms presented here. Section 8 summarizes the results and mentions some open questions. 1.3 Related Work An automatic load balancing approach widely used in practice is striping =-=[3, 4]-=-. In our terminology that means a logical block size D times larger than the physical block size, where each logical block is dispersed over all disks. This works for scanning large amounts of consecu... |

637 | LEDA: A platform for combinatorial and geometric computing - Mehlhorn, Naher - 1995 |

537 |
The input/output complexity of sorting and related problems
- Aggarwal, Vitter
- 1988
(Show Context)
Citation Context ...sible if we do not exploit both redundancy and random placement of data. The model considered here is the asynchronous pendant to the synchronous multi-head parallel disk model of Aggarwal and Vitter =-=[1]-=- which turned out to be very convenient for devising external memory algorithms [2]. 1.1 Basic Model We now introduce a simple form of our system model that suffices to understand most of this paper. ... |

533 |
Networks flows
- Ahuja, Magnanti, et al.
- 1993
(Show Context)
Citation Context ...rresponds to the optimal schedule shown in Figure 1. This matching representation can now be used to optimize schedules in several respects starting with well known algorithms for computing matchings =-=[24]-=-. 3.1 Minimizing the Maximal Delay If there is a schedule with maximal delaysL, it can be found by constraining the bipartite graph G 2 to those edges with weight at mostsL. As a side effect, this als... |

453 | Ecient dispersal of information for security, load balancing and fault tolerance
- Rabin
- 1989
(Show Context)
Citation Context ... explicitly scheduling write accesses using the same methods as described here for reading. A disadvantage of codes with w ? r + 1 is that they are computationally more expensive than parity-encoding =-=[34, 35, 36, 37, 38, 13]-=-. It makes sense to use different values for r and w concurrently on the same system for different purposes. Data which is rarely read and can be reconstructed otherwise, e.g., checkpointing data for ... |

266 |
A case for redundant arrays of inexpensive disks
- Patterson, Gibson, et al.
- 1988
(Show Context)
Citation Context ...augmenting path algorithm still computes schedules which are optimal for all request currently in the system. 1.2 Related Work An automatic load balancing approach widely used in practice is striping =-=[26, 25]-=-. In our terminology that means a logical block size D times larger than the physical block size, i.e., r = w = D. This works for scanning large amounts of consecutive data but is of little help for s... |

249 | Balanced allocations
- Azar, Broder, et al.
- 1999
(Show Context)
Citation Context ...ent has been proposed for the parallel file system RAMA [6]. Combining random placement and redundancy has first been considered in parallel computing for PRAM emulation [7] and online load balancing =-=[8]-=-. For scheduling disk accesses, these techniques have been used for multimedia applications [9, 10, 11, 12, 13, 14]. These papers use shortest queue, do not specify the scheduling algorithm, or schedu... |

235 |
Computers and intractability
- Garey, Johnson
- 1979
(Show Context)
Citation Context ... 2, random allocation, and identical release times for all requests, finding a currently optimal schedule without allowing preemption is NP-hard. This problem is equivalent to the partition problem =-=[39]-=-. Since many hard disks do not allow us to preempt a request; once started, we seem to be stuck. However, a good compromise is to split all requests into subrequests of estimated equal access time. Th... |

199 | The power of two choices in randomized load balancing
- Mitzenmacher
- 2001
(Show Context)
Citation Context ...ll known technique, shortest queue, maintains a FIFO queue of committed requests. A newly arrived request fi; jg that could be served on two disks i and j is committed to the disk with shortest queue =-=[6, 31, 23]-=-. We show that executing the requests locally in FIFO order is indeed optimal. However, immediately committing a request when it arrives is unnecessary. Perhaps the simplest more flexible strategy, la... |

172 |
Disk Striping
- Salem, Garcia-Molina
- 1986
(Show Context)
Citation Context ...augmenting path algorithm still computes schedules which are optimal for all request currently in the system. 1.2 Related Work An automatic load balancing approach widely used in practice is striping =-=[26, 25]-=-. In our terminology that means a logical block size D times larger than the physical block size, i.e., r w D. This works for scanning large amounts of consecutive data but is of little help for small... |

164 | Probability and Statistics with Reliability, Queueing, and Computer Science Applications - Trivedi - 1982 |

100 |
The Art of Computer Programming: Sorting and Searching, volume 3
- Knuth
- 1973
(Show Context)
Citation Context ...with large I/O demands are often designed in such a way that many requests can be submitted to the system in parallel. A prototypical example is external memory sorting via multiway merge sort, (e.g. =-=[18]-=-). Using an appropriate data structure, it can be predicted which data blocks will be needed in the near future. These blocks can then be prefetched asynchronously so that they are available when need... |

85 |
auf der Heide, Efficient PRAM Simulation on a Distributed Memory Machine
- Karp, Luby, et al.
- 1994
(Show Context)
Citation Context ...o. Nonredundant random placement has been proposed for the parallel file system RAMA [22]. Combining random placement and redundancy has first been considered in parallel computing for PRAM emulation =-=[17]-=- and online load balancing [6]. For scheduling disk accesses, these techniques have been used for multimedia applications [29, 30, 19, 24, 8, 28]. These papers use shortest queue, do not specify the s... |

81 |
A Case for Redundant Arrays
- Patterson, Gibson, et al.
- 1988
(Show Context)
Citation Context ...scheduling algorithms presented here. Section 8 summarizes the results and mentions some open questions. 1.3 Related Work An automatic load balancing approach widely used in practice is striping [3], =-=[4]-=-. In our terminology, that means a logical block size D times larger than the physical block size, where each logical block is dispersed over all disks. This works for scanning large amounts of consec... |

78 | Coding techniques for handling failures in large disk arrays. Algorithmica 12
- Hellerstein, Gibson, et al.
- 1994
(Show Context)
Citation Context ... explicitly scheduling write accesses using the same methods as described here for reading. A disadvantage of codes with w ? r + 1 is that they are computationally more expensive than parity-encoding =-=[34, 35, 36, 37, 38, 13]-=-. It makes sense to use different values for r and w concurrently on the same system for different purposes. Data which is rarely read and can be reconstructed otherwise, e.g., checkpointing data for ... |

75 | Adversarial queuing theory
- Borodin, Kleinberg, et al.
(Show Context)
Citation Context ...tu We could now generalize the above results to arbitrary distributions with finite variance using queuing theory. However, we choose the simpler and more powerful model of adversarial queuing theory =-=[28]-=-, [29]. This model makes it not hard to justify assumptions like independence and is very simple: Any sequence of event arrivals is allowed, as long as, within any time window of extent W, at most bWD... |

67 |
Queueing System with Selection of the Shortest of Two Queues: an Asymptotic Approach
- Vvedenskaya, Dobrushin, et al.
- 1996
(Show Context)
Citation Context ....1 we will see that one can do much better. Even the simplest scheduling heuristics for RDA are quite difficult to treat analytically for asynchronous request arrivals and small e. Vvedenskaya et al. =-=[15]-=- and Mitzenmacher [16] analyze the shortest queue heuristics as D !sfor Poisson arrivals with fixed arrival rate l = D=(1 + e) and exponentially distributed service times. There are theoretical result... |

61 | Simple randomized mergesort on parallel disks
- Barve, Grove, et al.
- 1997
(Show Context)
Citation Context ... that it can take time W i T log(N=D) loglog(N=D) j to complete a computation on N blocks and performing TD block accesses. Load balancing by random placement of data is a well known technique (e.g., =-=[7, 22]-=-). In this situation, delays for read requests can be investigated analytically using generating function techniques similar to those used in [27] for write buffering. Average delays behave like Q(1=e... |

60 |
EVENODD: An optimal scheme for tolerating double disk failures in RAID architectures
- Blaum, Brady, et al.
- 1994
(Show Context)
Citation Context |

59 | Comparing random data allocation and data striping in multimedia servers
- Santos, Muntz, et al.
- 2000
(Show Context)
Citation Context ...undancy has first been considered in parallel computing for PRAM emulation [7] and online load balancing [8]. For scheduling disk accesses, these techniques have been used for multimedia applications =-=[9, 10, 11, 12, 13, 14]-=-. These papers use shortest queue, do not specify the scheduling algorithm, or schedule large batches in a synchronous fashion. Some RAID arrays use load balancing techniques to spread read requests o... |

57 | Balanced allocations: The heavily loaded case
- Berenbrink, Czumaj, et al.
(Show Context)
Citation Context ...curate by significant constant factors when the number of requests in the system is much larger then D. (A recent analysis of shortest queue for the case of high loads only applies to a batched model =-=[18]-=-.) Adler et al. [19] consider an algorithm similar to the lazy queue algorithm based on synchronized rounds of allocation and job consumption. For small arrival rates 2 (1+es6e ? 16) the expected maxi... |

56 | Adaptive packet routing for bursty adversarial traffic
- Aiello, Kushilevitz, et al.
- 1998
(Show Context)
Citation Context ...could now generalize the above results to arbitrary distributions with finite variance using queuing theory. However, we choose the simpler and more powerful model of adversarial queuing theory [28], =-=[29]-=-. This model makes it not hard to justify assumptions like independence and is very simple: Any sequence of event arrivals is allowed, as long as, within any time window of extent W, at most bWD=ð1þ "... |

50 | Fast concurrent access to parallel disks
- Sanders, Egner, et al.
(Show Context)
Citation Context ...considered here is the asynchronous pendant to the synchronous multi-head parallel disk model of Aggarwal and Vitter [1] which turned out to be very convenient for devising external memory algorithms =-=[2]-=-. 1.1 Basic Model We now introduce a simple form of our system model that suffices to understand most of this paper. Generalizations to this model are introduced later when appropriate. The overview i... |

48 | Tolerating Multiple Failures in RAID Architectures with Optimal Storage and Uniform Declustering
- Alvarez, Burkhard, et al.
- 1997
(Show Context)
Citation Context |

38 |
A comparison of highavailability media recovery techniques
- Copeland, Keller
- 1989
(Show Context)
Citation Context ...tegy. Even if two copies go to different disks, all the work might have to be done by two disks. Ring allocation: A block stored on disk i is also stored on disk i +1 mod D where i is chosen randomly =-=[30, 31]-=-. Aerts et al. [31] propose this scheme because they can give an efficient algorithm for 15 finding optimal schedules for batches of requests. Although this scheme is better than plain mirroring, it i... |

30 |
Random duplicate assignment: An alternative to striping in video servers
- Korst
- 1997
(Show Context)
Citation Context ...undancy has first been considered in parallel computing for PRAM emulation [7] and online load balancing [8]. For scheduling disk accesses, these techniques have been used for multimedia applications =-=[9, 10, 11, 12, 13, 14]-=-. These papers use shortest queue, do not specify the scheduling algorithm, or schedule large batches in a synchronous fashion. Some RAID arrays use load balancing techniques to spread read requests o... |

27 | RAMA: An easy-to-use, highperformance parallel file system
- Miller, Katz
- 1997
(Show Context)
Citation Context ... that it can take time W i T log(N=D) loglog(N=D) j to complete a computation on N blocks and performing TD block accesses. Load balancing by random placement of data is a well known technique (e.g., =-=[6]-=-). In this situation, 4 delays for read requests can be investigated analytically using generating function techniques similar to those used in [2] for write buffering. Average delays behave like Q(1=... |

22 | Design and performance tradeoffs in clustered video servers
- Tewari, Mukherjee, et al.
- 1996
(Show Context)
Citation Context ...undancy has first been considered in parallel computing for PRAM emulation [7] and online load balancing [8]. For scheduling disk accesses, these techniques have been used for multimedia applications =-=[9, 10, 11, 12, 13, 14]-=-. These papers use shortest queue, do not specify the scheduling algorithm, or schedule large batches in a synchronous fashion. Some RAID arrays use load balancing techniques to spread read requests o... |

22 | A parallel disk storage system for real-time multimedia applications
- Muntz, Santos, et al.
- 1998
(Show Context)
Citation Context |

19 | Optimal read-once parallel disk scheduling
- Kallahalla, Varman
- 1999
(Show Context)
Citation Context ...rform on resource limited traffic, we report simulations based on a simple traffic pattern that produces all these effects and actually appears in applications based on sorting or offline prefetching =-=[32, 33]-=-. Note that sorting is the most I/O intensive part of many data base operations like index construction or certain joins. Consider m block buffers and a sequence of requests S to be consumed by the ap... |

18 | E.: On balls and bins with deletions
- Cole, Frieze, et al.
- 1998
(Show Context)
Citation Context ...arrivals with fixed arrival rate l = D=(1 + e) and exponentially distributed service times. There are theoretical results on system models which keep the number of requests in the system fixed (e.g., =-=[17]-=-). But so far none of these models seems to be able to approximate the behavior of disk servers. Also, many of the techniques for analyzing RDA seem to be inaccurate by significant constant factors wh... |

18 | Mixed-media file systems
- Bosch
- 1999
(Show Context)
Citation Context ... discuss how data should be allocated to the disks to allow both fine grained and coarse grained access to the same file. A good compromise is to use large physical blocks, e.g., currently around 1MB =-=[40]-=- to allow high bandwidth access with large requests. Later we are free to read smaller blocks. For general r out of w encoding this requires some further explanations. The obvious way --- chopping a l... |

17 | Duality between prefetching and queued writing with parallel disks
- Hutchinson, Sanders, et al.
- 2001
(Show Context)
Citation Context ...rform on resource limited traffic, we report simulations based on a simple traffic pattern that produces all these effects and actually appears in applications based on sorting or offline prefetching =-=[32, 33]-=-. Note that sorting is the most I/O intensive part of many data base operations like index construction or certain joins. Consider m block buffers and a sequence of requests S to be consumed by the ap... |

16 | Reconciling simplicity and realism in parallel disk models
- Sanders
- 2002
(Show Context)
Citation Context ...lizations for batched scheduling including variable block sizes, disk failures, 2 This restriction is improved in [20] for D ! . 5 and communication overheads can be obtained using similar techniques =-=[21]-=-. Batched scheduling algorithms can be converted into asynchronous scheduling algorithms by pipelining batched schedules. While one batch is executed on the disks, the newly arriving requests are reta... |

14 |
The Power of Two Choices
- Mitzenmacher
- 1996
(Show Context)
Citation Context ...e can do much better. Even the simplest scheduling heuristics for RDA are quite difficult to treat analytically for asynchronous request arrivals and small e. Vvedenskaya et al. [15] and Mitzenmacher =-=[16]-=- analyze the shortest queue heuristics as D !sfor Poisson arrivals with fixed arrival rate l = D=(1 + e) and exponentially distributed service times. There are theoretical results on system models whi... |

12 | Analyzing an infinite parallel job allocation process
- Adler, Berenbrink, et al.
- 1998
(Show Context)
Citation Context ...t constant factors when the number of requests in the system is much larger then D. (A recent analysis of shortest queue for the case of high loads only applies to a batched model [18].) Adler et al. =-=[19]-=- consider an algorithm similar to the lazy queue algorithm based on synchronized rounds of allocation and job consumption. For small arrival rates 2 (1+es6e ? 16) the expected maximal delay is shown t... |

10 | Block allocation in video servers for availability and throughput
- Tetzlaff, Flynn
- 1996
(Show Context)
Citation Context |

5 | Random duplicate storage strategies for load balancing in multimedia servers
- Aerts, Korst, et al.
- 2000
(Show Context)
Citation Context ...d read requests over primary and mirror disks equally. This approximately halves the observed delays. In Section 5.1 we will see that one can do much better. Copeland and Keller [15] and Aerts et al. =-=[2]-=- consider a technique seemingly similar to mirroring. The primary copy of a logical block is allocated on a random disk d and a redundant copy is placed on disk d + 1 mod D. In Section 5.1 we will see... |

3 |
auf der Heide, "Efficient PRAM simulation on a distributed memory machine
- Karp, Luby, et al.
- 1992
(Show Context)
Citation Context ...=D. Nonredundant random placement has been proposed for the parallel file system RAMA [6]. Combining random placement and redundancy has first been considered in parallel computing for PRAM emulation =-=[7]-=- and online load balancing [8]. For scheduling disk accesses, these techniques have been used for multimedia applications [9, 10, 11, 12, 13, 14]. These papers use shortest queue, do not specify the s... |

3 |
Adaptiv packet routing for bursty adversarial traffic
- Aiello, Kushilevitz, et al.
- 1998
(Show Context)
Citation Context ...d. We could now generalize the above results to arbitrary distributions with finite variance using queuing theory. However, we choose the simpler and more powerful model of adversarial queuing theory =-=[28, 29]-=-. This model makes no hard to justify assumptions like independence and is very simple: Any sequence of event arrivals is allowed, as long as within any time window of extent W , at most bWD=(1+ e)c r... |

3 |
Random duplicate storage for load balancing in multimedia servers
- Aerts, Korst, et al.
- 2000
(Show Context)
Citation Context ...tegy. Even if two copies go to different disks, all the work might have to be done by two disks. Ring allocation: A block stored on disk i is also stored on disk i +1 mod D where i is chosen randomly =-=[30, 31]-=-. Aerts et al. [31] propose this scheme because they can give an efficient algorithm for 15 finding optimal schedules for batches of requests. Although this scheme is better than plain mirroring, it i... |

2 |
Bounds on the separation of two parallel disk models
- Armen
- 1996
(Show Context)
Citation Context ...arily large delays for very small arrival rates of just above one request per time unit. Even redundancy is of limited help as long as the allocation strategy is deterministic. A lower bound by Armen =-=[5]-=- shows that it can take time W i T log(N=D) loglog(N=D) j to complete a computation on N blocks and performing TD block accesses. Load balancing by random placement of data is a well known technique (... |

2 | Design of the PRESTO multimedia storage network
- Berenbrink, Brinkmann, et al.
(Show Context)
Citation Context |

2 | Simple competitive request scheduling strategies
- Berenbrink, Riedel, et al.
- 1999
(Show Context)
Citation Context ...d old requests. Old request should be preferred to limit the frequency of large delays. This observation was the motivation for developing algorithms which explicitly handle delays. Berenbrink et al. =-=[23]-=- also propose a scheduling algorithm for RDA based on bipartite matching. They analyze the competitiveness of several online scheduling algorithms compared to an optimal offline schedule. However, the... |

2 | Incremental scheduling of mixed workloads in multimedia information servers
- Nerjes, Muth, et al.
- 2000
(Show Context)
Citation Context ...flexibility as large requests. 7.2 Scan Scheduling To achieve high throughput in the case of high system load one should deviate from the strict FIFO discipline to reduce the overall seek time (e.g., =-=[41]-=-). Sorting the requests by track number is a good approach for a single disk if we do not care about large delays for some requests. Good compromises between low delays and high performance are an act... |

1 |
Asynchronous control of disks in video servers
- Korst, Coumans
- 1998
(Show Context)
Citation Context ...heduled in a batch and assigned to the disks next. This strategy, applying the bounds from [2], yields an algorithm with maximal delay independent of D. However, the average delay is quite high (e.g. =-=[22]-=-). For practically interesting D, even the maximal delay is higher than for the asynchronous shortest queue heuristics. Originally, we thought that this problem could be solved using essentially the s... |

1 |
Adversarial queing theory
- Borodin, Kleinberg, et al.
- 1996
(Show Context)
Citation Context ...d. We could now generalize the above results to arbitrary distributions with finite variance using queuing theory. However, we choose the simpler and more powerful model of adversarial queuing theory =-=[28, 29]-=-. This model makes no hard to justify assumptions like independence and is very simple: Any sequence of event arrivals is allowed, as long as within any time window of extent W , at most bWD=(1+ e)c r... |

1 |
Analyzing an Infinite Parallel Job Allocation
- Adler, Berenbrink, et al.
- 1998
(Show Context)
Citation Context ...ed that the first requests are committed from left to right. is much larger then D. (A recent analysis of shortest queue for the case of high loads only applies to a batched model [18].) Adler et al. =-=[19]-=- consider an algorithm similar to the lazy queue algorithm based on synchronized rounds of allocation and job consumption. For small arrival rates 2 (1 þ " 6e >16), the expected maximal delay is shown... |