Holland, M. On-line Data Reconstruction in Redundant Disk Arrays, Carnegie Mellon University, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Write operations are substantially more complicated, as a result of the fundamental need to keep the parity information current. In systems that correct a single erasure, two methods of writing have been extensively studied, based on an observed difference between small and large writes; see [12], for example. Again, we consider only the portion of a write which lies within a single stripe. In the simplest case, suppose that we are writing new contents to an information disk f 0 , and there is a check disk c 0 whose contents are defined by c 0 = L 1 i=0 f i . Let b c 0 and b f 0 ....

....replacing s 1 disk reads by disk writes. However, read modify write requires s 1 additional reads and s 1 additional writes, for a total of 2s 2 disk accesses. Thus when 2s 2 1, we can expect read modify write to involve more disk accesses than does the stripe write. See, for example, [12, 14]. For the stripe write and the read modify write, it is crucial that a method be available for caching or buffering the contents of the disks accessed within the stripe. We assume throughout the presence of a buffer which is capable of holding the contents of the stripe across all disks (see ....

[Article contains additional citation context not shown here]

Mark C. Holland. On-Line Data Reconstruction In Redundant Disk Arrays. PhD thesis, Carnegie Mellon University, 1994.

....is unnecessarily expensive when writing to a small fraction of disks in an array. For large disk arrays with multiple erasure protection, when the number of disks to be written is relatively small compared with the number of information disks, all writes are implemented as a read modify write (see [5, 9]) To write to a disk, we only need to calculate the change to its check disks, before updating them. Therefore we read all of the check and information disks involved. Then the change to the parity is computed, the check disks are written and the new data is written. For a write involving only ....

Mark C. Holland. On-Line Data Reconstruction In Redundant Disk Arrays. PhD thesis, Carnegie Mellon University, 1994.

....or string failure doubles the utilization of each remaining disk in an affected reliability group. Even if the write portion of the RAID load is substantial, reconstruction of data objects lost due to the failure increases the load at disks in the same reliability group considerably. Declustering [2] (also called clustering [3] limits the utilization increase due to disk failure by organizing objects into smaller reliability groups, that are distributed over the entire RAID. Holland [2] achieves a distribution through mathematical block designs. Below we introduce the ACATS approach to ....

....due to the failure increases the load at disks in the same reliability group considerably. Declustering [2] also called clustering [3] limits the utilization increase due to disk failure by organizing objects into smaller reliability groups, that are distributed over the entire RAID. Holland [2] achieves a distribution through mathematical block designs. Below we introduce the ACATS approach to distribution that utilizes random permutations. III. MDS RAID Organization Maximum Distance Separable (MDS) RAIDs, derived from well known linear codes, provide additional redundancy to ....

[Article contains additional citation context not shown here]

M.C. Holland, "On-line data reconstruction in redundant disk arrays," Ph.D. thesis, Carnegie Mellon University, 1994.

....codes for disk arrays. Prior to examining the combinatorial problems posed, we explore the disk array application. As processor speeds have increased rapidly in recent years, one method of bridging the Input Output (I O) performance gap has been to use redundant arrays of independent disks (RAID) [10]. Individual data reads and writes are striped across multiple disks, thereby creating I O parallelism. Disk striping intercepts each individual I O request, divides it into a series of n smaller striping units and distributes these over n disks. The logical and physical location of each I O ....

....information onto additional disks allows reconstruction of lost information in the presence of disk failures. This creates disk arrays with high throughput and good reliability. However, an array of disks has a substantially greater probability of a disk failure than does an individual disk [9, 10]. Indeed, Hellerstein et al. 9] have shown that the reliability of an array of 1000 disks which protects against one error, even with periodic daily or weekly repairs, has a lower reliability than an individual disk. Most systems that are available currently handle only one or two disk failures ....

[Article contains additional citation context not shown here]

Mark C. Holland. On-Line Data Reconstruction In Redundant Disk Arrays. PhD thesis, Carnegie Mellon University, 1994.

....layouts presented in [11, 16] satisfy the first three properties; and depending on how the client data units are mapped to the stripes, they can be made to satisfy either large write optimization or maximal parallelism, although it was not known how to satisfy both simultaneously. Holland [12] studied the problem of rearranging the placement of client data units to achieve better parallelism while maintaining large write optimization, although he did not achieve maximal parallelism. The time requirement to compute a BIBD based mapping is reasonably small, although the space requirement ....

Mark Calvin Holland. On-Line Data Reconstruction In Redundant Disk Arrays. PhD thesis, Department of Electrical and Computer Engineering, Carnegie Mellon University, 1994.

.... 116(2) 6 102(6) 108(7) 108(7) 114(8) 114(2) 114(2) 9 108(7) 108(5) 108(5) 108(4) 117(3) 117(1) 12 108(5) 108(4) 108(4) 108(3) 108(1) 108(1) 18 108(4) 108(3) 108(3) 108(3) 108(1) 108(1) 36 108(3) 108(2) 108(2) 108(2) 108(1) 108(1) Table 3: Case Study Results Using standard techniques (see [6]) we can also compute values for the mean time to service loss (MTTSL) and the probability of service loss within a given amount of time (we use the term service loss instead of data loss see Section 4.1. If the mean time to failure for a disk is 500,000 hours, we get that the MTTSL of a disk ....

Holland, M.C. On-Line Data Reconstruction In Redundant Disk Arrays. PhD Dissertation, Dept. of Electrical and Computer Engineering, Carnegie Mellon University, 1994.

....and layout of a disk array is that of modeling reads and writes. In a single error correcting disk array the basic unit is a parity stripe or group. There can be no overlap of disk accesses. One check disk is assigned to each data disk in each group. Two basic types of writes have been described [8]. If more than half of the parity stripe is being written then all of the remaining data disks in the stripe are pre read, the parity for the stripe is computed and the new data and parity are written [8] If however, the new data writes to less than half of the parity stripe, a read modify write ....

....is assigned to each data disk in each group. Two basic types of writes have been described [8] If more than half of the parity stripe is being written then all of the remaining data disks in the stripe are pre read, the parity for the stripe is computed and the new data and parity are written [8]. If however, the new data writes to less than half of the parity stripe, a read modify write is preferred [8] In this case the requested data disks and their associated parity disks are pre read, the new parity is computed and then the new data and parity is written back out. 2 In a multiple ....

[Article contains additional citation context not shown here]

Mark C. Holland. On-Line Data Reconstruction In Redundant Disk Arrays. PhD thesis, Carnegie Mellon University, 1994.

....to an existing code base, thereby restricting a designer s ability to explore the design space, confining experimentation to limited departures from the current code structure. Finally, researchers are investigating more aggressive redundant disk array architectures to boost performance [Bhide92, Blaum94, Cao93, Menon93, Stodolsky93, Holland94]. The acceptance of these proposals is put at risk due to their further increases in the complexity of error handling and the difficulty of modifying existing code structure. Forward error recovery has been used with arguable success in the design of single disk systems and filesystems. Single ....

Mark Holland, "On-line data reconstruction in redundant disk arrays." Ph.D. dissertation, Carnegie Mellon University School of Computer Science technical report CMU-CS-94-164, May 1994.

....to an existing code base, thereby restricting a designer s ability to explore the design space, confining experimentation to limited departures from the current code structure. Finally, researchers are investigating more aggressive redundant disk array architectures to boost performance [Bhide92, Blaum94, Cao93, Menon93, Stodolsky93, Holland94]. The acceptance of these proposals is put at risk due to their further increases in the complexity of error handling and the difficulty of modifying existing code structure. Forward error recovery has been used with arguable success in the design of single disk systems and filesystems. Single ....

Mark Holland, "On-line data reconstruction in redundant disk arrays." Ph.D. dissertation, Carnegie Mellon University School of Computer Science technical report CMU-CS-94-164, May 1994.

....provides libraries of mapping and encoding routines, nodes, graphs, and disk queueing policies which architects may either draw upon or enhance when creating new arrays. At the time of this writing, we have completed implementation of RAID levels 0, 1, 4, 5, 6, as well as parity declustering [Holland94] distributed sparing [Holland94] interleaved declus H R d R d T fault free read fault free small write H T R p R d R d XOR W d W d W p Figure 1 RAID Level 5 Operations Represented as DAGs RAIDframe: rapid prototyping for disk arrays 1 William V. Courtright II, 2 Garth Gibson, 1 ....

....and encoding routines, nodes, graphs, and disk queueing policies which architects may either draw upon or enhance when creating new arrays. At the time of this writing, we have completed implementation of RAID levels 0, 1, 4, 5, 6, as well as parity declustering [Holland94] distributed sparing [Holland94] interleaved declus H R d R d T fault free read fault free small write H T R p R d R d XOR W d W d W p Figure 1 RAID Level 5 Operations Represented as DAGs RAIDframe: rapid prototyping for disk arrays 1 William V. Courtright II, 2 Garth Gibson, 1 Mark Holland, 2 Jim Zelenka 1 ....

[Article contains additional citation context not shown here]

M. Holland, "On-Line Data Reconstruction in Redundant Disk Arrays," Ph.D thesis, Computer Science Technical Report CMU-CS-94-164, Carnegie Mellon University, (1994).

....After Commit Point 2.5 Reconstructing Data On line When a Disk Fails In Chapter One we introduced the need for a process in which the array restores itself to the fault free state following a disk failure. In this section, we provide a brief description of a disk oriented algorithm (taken from [Holland94]) for reconstructing lost data into spare disk space. For a more complete discussion of reconstruction algorithms, including performance evaluations and optimizations of the disk oriented algorithm, please refer to Chapter Four in [Holland94] 2.5.1 Disk Oriented Reconstruction Not only must a ....

....description of a disk oriented algorithm (taken from [Holland94] for reconstructing lost data into spare disk space. For a more complete discussion of reconstruction algorithms, including performance evaluations and optimizations of the disk oriented algorithm, please refer to Chapter Four in [Holland94]. 2.5.1 Disk Oriented Reconstruction Not only must a single fault tolerant disk array recover from the loss of a disk, it should be able to effect this recovery without taking the system off line. This is implemented by maintaining one or more on line spare disks in the array. When a disk fails, ....

Holland, M. On-line Data Reconstruction in Redundant Disk Arrays, Carnegie Mellon University, 1994.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC