Image Databases: Search and Retrieval of Digital Imagery Edited by Vittorio Castelli, Lawrence D. Bergman Copyright  2002 John Wiley & Sons, Inc. ISBNs: 0-471-32116-8 (Hardback); 0-471-22463-4 (Electronic) 6 Storage Architectures for Digital Imagery HARRICK M. VIN University of Texas at Austin, Austin, Texas PRASHANT SHENOY University of Massachusetts, Amherst, Massachusetts Rapid advances in computing and communication technologies coupled with the dramatic growth of the Internet have led to the emergence of a wide variety of multimedia applications, such as distance education, on-line virtual worlds, immersive telepresence, and scientific visualization. These applications differ from conventional distributed applications in at least two ways. First, they involve storage, transmission, and processing of heterogeneous data types — such as text, imagery, audio, and video — that differ significantly in their characteris- tics (e.g., size, data rate, real-time requirements, and so on). Second, unlike conventional best-effort applications, these applications impose diverse perfor- mance requirements — for instance, with respect to timeliness — on the networks and operating systems. Unfortunately, existing networks and operating systems do not differentiate between data types, offering a single class of best-effort service to all applications. Hence, to support emerging multimedia applications, existing networks and operating systems need to be extended along several dimensions. In this chapter, issues involved in designing storage servers that can support such a diversity of applications and data types are discussed. First, the specific issues that arise in designing a storage server for digital imagery are described, and then the architectural choices for designing storage servers that efficiently manage the storage and retrieval of multiple data types are discussed. Note that as it is difficult, if not impossible, to foresee requirements imposed by future applications and data types, a storage server that supports multiple data types and applications will need to facilitate easy integration of new application classes and data types. This dictates that the storage server architecture be extensible, allowing it to be easily tailored to meet new requirements. 139 140 STORAGE ARCHITECTURES FOR DIGITAL IMAGERY The rest of the chapter is organized as follows. We begin by examining tech- niques for placement of digital imagery on a single disk, a disk array, and a hierar- chical storage architecture. We then examine fault-tolerance techniques employed by servers to guarantee high availability of image data. Next, we discuss retrieval techniques employed by storage servers to efficiently access images and image sequences. We then discuss caching and batching issues employed by such servers to maximize the number of clients supported. Finally, we examine architectural issues in incorporating all of these techniques into a general-purpose file system. 6.1 STORAGE MANAGEMENT 6.1.1 Single Disk Placement A storage server divides images into blocks while storing them on disks. In order to explore the viability of various placement models for storing these blocks on magnetic disks, some of the fundamental characteristics of these disks are briefly reviewed. Generally, magnetic disks consist of a collection of platters, each of which is composed of a number of circular recording tracks (Fig. 6.1). Platters spin at a constant rate. Moreover, the amount of data recorded on tracks may increase from the innermost track to the outermost track (e.g., in the case of zoned disks). The storage space of each track is divided into several disk blocks, each consisting of a sequence of physically contiguous sectors. Each platter is associated with a read or write head that is attached to a common actuator. A cylinder is a stack of tracks at one actuator position. In such an environment the access time of a disk block consists of three components: seek time, rotational latency,anddata-transfer time. Seek time is the time required to position the disk head on the track, containing the desired data and is a function of the initial start-up cost to accelerate the disk head and the number of tracks that are traversed. Rotational latency, on the other hand, is the time for the desired data to rotate under the head before it can be read or Platter Track Actuator Head Direction of rotation Figure 6.1. Architectural model of a conventional magnetic disk. STORAGE MANAGEMENT 141 written and is a function of the angular distance between the current position of the disk head and the location of the desired data, as well as the rate at which platters spin. Once the disk head is positioned at the desired disk block, the time to retrieve its contents is referred to as the data-transfer time; it is a function of the disk block size and data-transfer rate of the disk. The placement of data blocks on disks is generally governed by contiguous, random, or constrained placement policy. Contiguous placement policy requires that all blocks belonging to an image be placed together on the disk. This ensures that once the disk head is positioned at the beginning of an image, all its blocks can be retrieved without incurring any seek or rotational latency. Unfortunately, the contiguous placement policy results in disk fragmentation in environments with frequent image creations and deletions. Hence, contiguous placement is well suited for read-only systems (such as compact discs, CLVs, and so on.), but is less desirable for a dynamic, read-write storage systems. Storage servers for read-write systems have traditionally employed random placement of blocks belonging to an image on disk [1,2]. This placement scheme does not impose any restrictions on the relative placement on the disks of blocks belonging to a single image. This approach eliminates disk fragmentation, albeit at the expense of incurring high seek time and rotational latency overhead while accessing an image. Clearly, the contiguous and random placement models represent two ends of a spectrum; whereas the former does not permit any separation between successive blocks of an image on disk, the latter does not impose any constraints at all. The constrained or the clustered placement policy is a generalization of these extremes; it requires the blocks to be clustered together such that the maximum seek time and rotational latency incurred while accessing the image does not exceed a predefined threshold. For the random and the constrained placement policies, the overall disk throughput depends on the total seek time and rotational latency incurred per byte accessed. Hence, to maximize the disk throughput, image servers use as large a block size as possible. 6.1.2 Multidisk Placement Because of the large sizes of images and image sequences (i.e., video streams), most image and video storage servers utilize disk arrays. Disk arrays achieve high performance by servicing multiple input-output requests concurrently and by using several disks to service a single request in parallel. The performance of a disk array, however, is critically dependent on the distribution of the workload (i.e., the number of blocks to be retrieved from the array) among the disks. The higher the imbalance in the workload distribution, the lower is the throughput of the disk array. To effectively utilize a disk array, a storage server interleaves the storage of each image or image sequence among the disks in the array. The unit of data interleaving, referred to as a stripe unit, denotes the maximum amount of 142 STORAGE ARCHITECTURES FOR DIGITAL IMAGERY logically contiguous data that is stored on a single disk [82,3]. Successive stripe units of an object are placed on disks, using a round-robin or random allocation algorithm. Conventional file systems select stripe unit sizes that minimize the average response time while maximizing throughput. In contrast, to decrease the frequency of playback discontinuities, image and video servers select a stripe unit size that minimizes the variance in response time while maximizing throughput. Although small stripe units result in a uniform load distribution among disks in the array (and thereby decrease the variance in response times), they also increase the overhead of disk seeks and rotational latencies (and thereby decrease throughput). Large stripe units, on the other hand, increase the array throughput at the expense of increased load imbalance and variance in response times. To maximize the number of clients that can be serviced simultaneously, the server should select a stripe unit size that balances these trade-offs. Table 6.1 illustrates typical block or stripe unit sizes employed to store different types of data. The degree of striping — the number of disks over which an image or an image sequence is striped — is dependent on the number of disks in the array. In relatively small disk arrays, striping image sequences across all disks in the array (i.e., wide-striping) yields a balanced load and maximizes throughput. For large disk arrays, however, to maximize the throughput, the server may need to stripe image sequences across subsets of disks in the array and replicate their storage to achieve load balancing. The amount of replication for each image sequence depends on the popularity of the image sequence and the total storage-space constraints. 6.1.2.1 From Images to Multiresolution Imagery. The placement technique becomes more challenging if the imagery is encoded using a multiresolution encoding algorithm. In general, multiresolution imagery consists of multiple layers. Although all layers need be retrieved to display the imagery at the highest resolution, only a subset of the layers need to be retrieved for lower resolution displays. To efficiently support the retrieval of such images at different resolu- tions, the placement algorithm needs to ensure that the server access only as much data as needed and no more. To ensure this property, the placement algorithm should store multiresolution images such that: (1 ) each layer is independently Table 6.1. Typical Block or Stripe Unit Size for Different Data Types Data Type Storage Block or Stripe Requirement Unit Size Text 2–4 KB 0.5–4 KB Gif Image 64 KB 4–8 KB Satellite Image 60 MB 16 KB MPEG Video 1 GB 64–256 KB STORAGE MANAGEMENT 143 Layer 1 block Layer 2 block Layer n block . . . . Data retrieved for lowest resolution display Data retrieved for highest resolution display Figure 6.2. Contiguous placement of different layers of a multiresolution image. Storing data from different resolutions in separate disk blocks enables the server to retrieve each resolution independently of others; storing these blocks contiguously enables the server to reduce disk seek overheads while accessing multiple layers simultaneously. accessible, and (2 ) the seek and rotational latency while accessing any subset of the layers is minimized. Although the former requirement can be met by storing layers in separate disk blocks, the latter requirement can be met by storing these disk blocks adjacent on disk. Observe that this placement policy is general and can be used to interleave any multiresolution image or video stream on the array. Figure 6.2 illustrates this placement policy. 6.1.2.2 From Images to Video Streams. Consider a disk array–based video server. If the video streams are compressed using a variable bit rate (VBR) compression algorithm, then the sizes of frames (or images) will vary. Hence, if the server stores these video streams on disks using fixed-size stripe units, each stripe unit will contain a variable number of frames. On the other hand, if each stripe unit contains a fixed number of frames (and hence data for a fixed playback duration), then the stripe units will have variable sizes. Thus, depending on the striping policy, retrieving a fixed number of frames will require the server to access a fixed number of variable-size blocks or a variable number of fixed-size blocks [4,5,6]. Because of the periodic nature of video playback, most video servers service clients by proceeding in terms of periodic rounds. During each round, the server retrieves a fixed number of video frames (or images) for each client. To ensure continuous playback, the number of frames accessed for each client during a round must be sufficient to meet its playback requirements. In such an architecture, a server that employs variable-size stripe units (or fixed-time stripe units) accesses a fixed number of stripe units during each round. This uniformity of access, when coupled with the sequential and periodic nature of video retrieval, enables the server to balance load across the disks in the array. This efficiency, however, comes at the expense of increased complexity of storage-space management. The placement policy that utilizes fixed-size stripe units, on the other hand, simplifies storage-space management but results in higher load imbalance across the disks. In such servers, load across disks within a server may become unbal- anced, at least transiently, because of the arrival pattern of requests. To smoothen 144 STORAGE ARCHITECTURES FOR DIGITAL IMAGERY out this load imbalance, servers employ dynamic load-balancing techniques. If multiple replicas of the requested video stream are stored on the array, then the server can attempt to balance load across disks by servicing the request from the least-loaded disk containing a replica. Further, the server can exploit the sequen- tiality of video retrieval to prefetch data for the streams to smoothen out variation in the load imposed by individual video stream. 6.1.3 Utilizing Storage Hierarchies The preceding discussion has focused on fixed disks as the storage medium for image and video servers. This is primarily because disks provide high throughput and low latency relative to other storage media such as tape libraries, optical juke boxes, and so on. The start-up latency for devices such as tape libraries is substantial as it requires mechanical loading of the appropriate tape into a reader station. The advantage, however, is that they offer very high storage capacities (Table 6.2). In order to construct a cost-effective image and video storage system that provides adequate throughput, it is logical to use a hierarchy of storage devices [7,8,9,10]. There are several possible strategies for managing this storage hierarchy, with different techniques for placement, replacement, and so on. In one scenario, a relatively small set of frequently requested images and videos are placed on disks and the large set of less frequently requested data objects are stored in optical juke boxes or tape libraries. In this storage hierarchy there are several alternatives for managing the disk system. The most common architecture is the one in which disks are used as a staging area (cache) for the secondary storage devices and the entire image and video files are moved from the tertiary storage to the disk. It is then possible to apply traditional cache-management techniques to manage the content of the disk array. For very large-scale servers, it is also possible to use an array of juke boxes or tape readers [10]. In such a system, images and video objects may need to be striped across these tertiary storage devices [11]. Although striping can improve I/O throughput by reading from multiple tape drives in parallel, it can also increase contention for drives (because each request accesses all drives). Studies have shown that such systems must carefully balance these trade-offs by choosing an appropriate degree of striping for a given workload [11,12]. Table 6.2. Tertiary Storage Devices Disks Tapes Magnetic Optical Low-End High-End Capacity 40 GB 200 GB 500 GB 10 TB Mount Time 0 sec 20 sec 60 sec 90 sec Transfer Rate 10 MB/s 300 KB/s 100 KB/s 1,000 KB/s FAULT TOLERANCE 145 6.2 FAULT TOLERANCE Most image and video servers are based on large disk arrays, and hence the ability to tolerate disk failures is central to the design of such servers. The design of fault-tolerant storage systems has been a topic of much research and develop- ment over the past decade [13,14]. In most of these systems, fault-tolerance is achieved either by disk mirroring [15] or parity encoding [16,17]. Disk arrays that employ these techniques are referred to as redundant array of independent disks (RAID). RAID arrays that employ disk mirroring achieve fault-tolerance by duplicating data on separate disks (and thereby incur a 100 percent storage-space overhead). Parity encoding, on the other hand, reduces the overhead consider- ably by employing error-correcting codes. For instance, in a RAID level five disk array, consisting of D disks, parity computed over data stored across (D − 1) disks is stored on another disk (e.g., the left-symmetric parity assignment shown in Figure 6.3a) [18,19,17]. In such architectures, if one of the disks fails, the data on the failed disk is recovered using the data and parity blocks stored on the surviving disks That is, each user access to a block on the failed disk causes one request to be sent to each of the surviving disks. Thus, if the system is load- balanced prior to disk failure, the surviving disks would observe at least twice as many requests in the presence of a failure [20]. The declustered parity disk array organization [21,22,23] addresses this problem by trading some of the array’s capacity for improved performance in the presence of disk failures. Specifically, it requires that each parity block protect some smaller number of data blocks [for e.g., (G − 1)]. By appropriately distributing the parity information across all the D disks in the array, such a policy ensures that each surviving disk would see an on-the-fly reconstruction load increase of (G − 1)/(D − 1) instead of (D − 1)/(D − 1) = 100% [Fig. 6.3b]. M0.0 M1.1 M2.2 P3 M5.0 Disk1 M0.1 M1.2 P2 M4.0 M5.1 Disk2 M0.2 P1 M3.0 M4.1 M5.2 Disk3 P0 M2.0 M3.1 M4.2 P5 Disk4 M1.0 M2.1 M3.2 P4 M6.1 Disk5 M0.0 M1.0 M2.0 M3.0 P4 Disk1 M0.1 M1.1 M2.1 P3 M4.0 Disk2 M0.2 M1.2 P2 M3.1 M4.1 Disk3 M0.3 P1 M2.2 M3.2 M4.2 Disk4 P0 M1.3 M2.3 M3.3 M4.3 Disk5 (a) Left-symmetric data organization in RAID level 5 disk array with G = D = 5 (b) Declustered parity organization with G = 4 and D = 5 Figure 6.3. Different techniques for storing parity blocks in a RAID-5 architecture. (a) depicts the left-symmetric parity organization, in which the parity group size is same as the number of disks in the array; (b) depicts the declustered parity organization in which the parity group size is smaller than the number of disks. M i.j and P i denote data and parity blocks, respectively, and P i = M i.0 ⊕ M i.1 ···⊕M i.(G−2) . 146 STORAGE ARCHITECTURES FOR DIGITAL IMAGERY In general, with such parity-based recovery techniques, increase in the load on the surviving disks in the event of a disk failure results in deadline violations in the playback of video streams. To prevent such a scenario with conventional fault-tolerance techniques, servers must operate at low levels of disk utilization during the fault-free state. Image and video servers can reduce this overhead by exploiting the characteristics of imagery. There are two general techniques that such servers may use. • A video server can exploit the sequentiality of video access to reduce the overhead of on-line recovery in a disk array. Specifically, by computing parity information over a sequence of blocks belonging to the same video stream, the server can ensure that video data retrieved for recovering a block stored on the failed disk would be requested by the client in the near future. By buffering such blocks and then servicing the requests for their access from the buffer, this method minimizes the overhead of the on-line failure recovery process. • Because human perception is tolerant to minor distortions in images, a server can exploit the inherent redundancies in images to approximately reconstruct lost image data using error-correcting codes instead of perfectly recovering image data stored on the failed disk. In such a server, each image is parti- tioned into subimages and if the subimages are stored on different disks, then a single disk failure will result in the loss of fractions of several images. In the simplest case, if the subimages are created in the pixel domain (i.e., prior to compression) such that none of the immediate neighbors of a pixel in the image belong to the same subimage, then even in the presence of a single disk failure, all the neighbors of the lost pixels will be available. In this case, the high degree of correlation between neighboring pixels will make it possible to reconstruct a reasonable approximation of the original image. Moreover, no additional information will have to be retrieved from any of the surviving disks for recovery. Although conceptually elegant, such precompression image partitioning tech- niques significantly reduce the correlation between the pixels assigned to the same subimage and hence adversely affect image-compression efficiency [24,25]. The resultant increase in the bit rate requirement may impose higher load on each disk in the array even during the fault-free state, thereby reducing the number of video streams that can be simultaneously retrieved from the server. A number of postcompression partitioning algorithms that address this limitation have been proposed [26,27]. The concepts in postcompression partitioning is illustrated by describing one such algorithm, namely, the loss-resilient joint photographic expert group (JPEG) (LRJ). 6.2.1 Loss-Resilient JPEG (LRJ) Algorithm As human perception is less sensitive to high-frequency components of the spec- tral energy in an image, most compression algorithms transform images into the FAULT TOLERANCE 147 frequency domain so as to separate low- and high-frequency components. For instance, the JPEG compression standard fragments image data into a sequence of 8 × 8 pixel blocks and transforms them into the frequency domain using discrete cosine transform (DCT). DCT uncorrelates each pixel block into an 8 × 8 array of coefficients such that most of the spectral energy is packed in the fewest number of low-frequency coefficients. Although the lowest frequency coefficient (referred to as the DC coefficient) captures the average brightness of the spatial block, the remaining set of 63 coefficients (referred to as the AC coefficients) capture the details within the 8 × 8 pixel block. The DC coefficients of successive blocks are difference-encoded independent of the AC coefficients. Within each block, the AC coefficients are quantized to remove high-frequency components, scanned in a zigzag manner to obtain an approximate ordering from lowest to highest frequency and finally run-length and entropy-encoded. Figure 6.4 depicts the main steps involved in the JPEG compression algorithm [28]. The loss-resilient JPEG (LRJ) algorithm is an enhancement of the JPEG compression algorithm and is motivated by the following two observations: • Because the DC coefficients capture the average brightness of each 8 × 8 pixel block and because the average brightness of pixels gradually changes across most images, the DC coefficients of neighboring 8 × 8 pixel blocks are correlated. Consequently, the value of DC coefficient of a block can be reasonably approximated by extrapolating from the DC coefficients of the neighboring blocks. Discrete cosine transform Run-length and Huffman encoding Quantization Compressed image data Image data B(i, j−1) B(i, j) Differential encoding of DC coefficients Zig-zag reordering of AC coefficients JPEG compression algorithm DC(B(i,j−1)) DC(B(i,j)) d = DC(B(i, j)) − DC(B(i, j−1)) Figure 6.4. JPEG compression algorithm. 148 STORAGE ARCHITECTURES FOR DIGITAL IMAGERY • Owing to the very nature of DCT, the set of AC coefficients generated for each 8 × 8 block are uncorrelated. Moreover, because DCT packs the most amount of spectral energy into a few low-frequency coefficients, quantizing the set of AC coefficients (by using a user-defined normalization array) yields many zeroes, especially at higher frequencies. Consequently, recov- ering a block by simply substituting a zero for each of the lost AC coefficient is generally sufficient to obtain a reasonable approximation of the original image (at least as long as the number of lost coefficients are small and are scattered throughout the block). Thus, even when parts of a compressed image have been lost, a reasonable recovery is possible if: (1 ) the image in the frequency domain is partitioned into a set of N subimages such that none of the DC coefficients in the eight- neighborhood of a block belong to the same subimage, and (2 ) the AC coefficients of a block are scattered among multiple subimages. To ensure that none of the blocks contained in a subimage are in the eight-neighborhood of each other, N should be at least 4 [27]. To scatter the AC coefficients of a block among multiple subimages, the LRJ compression algorithm employs a scrambling tech- nique, which when given a set of N blocks of AC coefficients, creates a new set of N blocks such that the AC coefficients from each of the input blocks are equally distributed among all of the output blocks (Fig. 6.5). Once all the blocks within the image have been processed, each of the N subimages can be independently encoded. A2 A3 A4 A1 A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 A15 B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12 B13 B14 B15 A4 A8 A12 A1 A5 A9 A13 A2 A6 A10 A14 B1 B5 B9 B13 B2 B6 B10 B14 B3 B7 B11 B15 B4 B8 B12 C2 C6 C10 C14 C3 C7 C11 C15 C4 C8 C12 D1 D5 D9 D13 D2 D6 D10 D14 D3 D7 D11 D15 D4 D8 D12 A7 A11 A15 A3 Scrambling AC coefficients C1 C5 C9 C13 A0 A0 B0 B0 C2 C3 C4 C1 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D11 D12 D13 D14 D15 C0 C0 D0 D0 Figure 6.5. Scrambling AC coefficients. Here A 0 , B 0 , C 0 ,andD 0 denote DC coefficients, and ∀i ∈ [29, 30] : A i ,B i ,C i ,andD i represent AC coefficients.