Queuing architecture including a plurality of queues and assocated method for controlling admission for disk access requests for video content

Abstract

A queuing architecture and method for scheduling disk drive access requests in a video server. The queuing architecture employs a controlled admission policy that determines how a new user is assigned to a specific disk drive in a disk drive array.

Description

[0002] The invention relates to methods of scheduling disk access requests in a video server, and, more particularly, to statistical scheduling methods that improve the effective disk bandwidth provided by video servers.

BACKGROUND OF THE DISCLOSURE

[0003] Video-on-demand systems allow subscribers (also referred to as users or clients) to request video programs from a video library at any time for immediate viewing in their homes. Subscribers submit requests to a video service provider via a communication channel (e.g., telephone lines or a back channel through the distribution network that carries the video to the subscriber's home), and the requested video program is routed to the subscriber's home via telephone or coaxial television lines. In order to provide such movie-on-demand services, video service providers use a video server to process subscriber requests, retrieve the requested programs from storage, and distribute the programs to the appropriate subscriber(s). One exemplary system for providing video-on-demand services is described in commonly assigned U.S. patent application Ser. No. 08/984,710, filed Dec. 3, 1997, which is incorporated herein by reference.

[0004] In order for video servers to provide good performance, it is crucial to schedule video storage (disk) access requests such that disk bandwidth is maximized. Also, once a subscriber is watching a program, it is imperative to continuously deliver program content to the subscriber without interruption. In addition to distributing content to subscribers, disk bandwidth in a video server is typically also utilized for operations such as loading content, disk maintenance and file system meta-data syncing. Disk bandwidth may also be reserved for reducing latency in data transfer to subscribers. The number of subscribers that can be properly served concurrently by a video server therefore depends on effective disk bandwidth, which in turn depends on how disk access requests are scheduled.

[0005] One of the problems facing current disk scheduling methods is the potential variation in time required to service disk accesses. For example, the internal transfer rate of a Seagate Cheetah disk varies from 152 Mbps on inner tracks to 231 Mbps on outer tracks, and the seek time can vary from 0 ms to 13 ms depending on how far apart the segments of data are from one another. Given these variations in seek and transfer times and the fact that the server may contain sixteen or more disk drives, it is difficult to determine the effective disk bandwidth of a video server. As a result, current disk scheduling methods allocate a fixed amount of time for every disk access request, regardless of whether the access finishes early. This results in a deterministic system in which the available disk bandwidth is known, but since the fixed amount of time must be large enough to accommodate a worst-case disk access, disk bandwidth is wasted.

[0006] Therefore, there is a need in the art for a method and apparatus for scheduling disk access requests in a video server without allocating worst-case access times, thus improving disk bandwidth utilization.

SUMMARY OF THE INVENTION

[0007] The disadvantages associated with the prior art are overcome by a method of the present invention, called Statistical Disk Scheduling (SDS), which exploits the fact that disk access times are on average significantly less than the worst case access time. The SDS finds use in improving video server functionality by increasing the bandwidth utilization of the storage medium in the following manner: worst case performance is used for priority operations (e.g., user read operations) but the bandwith created by better than worst case performance is used for non-priority operations such as loading content onto the disk drives and disk maintenance. As a result, bandwidth for loading content and disk maintenance, or file system meta-data syncing does not have to be specifically reserved, thus increasing the number of users that can be served simultaneously by the video server.

[0008] SDS maintains at least two queues and a queue selector. The first queue is an access request queue for access requests from a current user that is presently viewing a program and the second queue is for all other forms of access requests. The second queue may comprise multiple queues to provide a queuing hierarchy. The requests are ordered in each of the queues to optimize the bandwidth and ensure that the data to the current users is not interrupted such that a display anomaly occurs. The queue selector identifies the queue that will supply the next access request to a disk queue. The selected requests are sent to the disk queues for execution. The disk queues are generally located on the disk drives and are generally not accessible except to place a request in the queue for each disk drive. The requests are then executed on a first-in, first-out manner. In effect, the invention defers disk use to the latest possible moment because once the request is in the disk queue it is more difficult to change. The inventive queue structure provides opportunities to alter the disk access requests and their execution order prior to sending the requests to the disk queue. If a disk queue is not used, i.e., the disk drive does not have an internal queue, then the access requests are sent one at a time from the SDS to the disk drive for execution.

[0009] More specifically, one embodiment of the SDS maintains three queues for each disk based on the type and priority of disk access requests, and a queue selector for managing queue selection. Selected requests are forwarded from the three queues to the disk such that bandwidth utilization is maximized, while giving highest priority to subscribers currently viewing a program so that their program streams are generally not interrupted. (Subscribers currently viewing a program are referred to as “steady-state” subscribers.) SDS dynamically monitors bandwidth utilization to determine when lower-priority requests can be scheduled without affecting on-time completion of the higher priority steady-state subscriber requests. In order to keep the disks busy and maximize disk bandwidth utilization, disk command queuing may be employed to ensure that the disk can begin seeking for the next access immediately after it finishes the data transfer for the current disk access.

[0010] In another embodiment of the invention, the queue selector uses a controlled admission algorithm that identifies a particular queue in which a new user (subscriber) will be assigned. The queue assignment process computes whether the user request can be serviced within a particular service period without causing other user requests to miss their data access deadlines. If the request can be performed in a timely manner, the user is assigned to the particular new user queue. However, if the request cannot be performed without detrimentally impacting other user requests within that service period, the request is reallocated to another entry point (i.e., a new user queue associated with another disk drive) into the disk array.

BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The teachings of the present invention can be readily understood by considering the following detailed description in conjunction with the accompanying drawings, in which:

[0012] FIG. 1 depicts a high-level block diagram of a video-on-demand system that includes a generic video server incorporating the present invention;

[0013] FIG. 2 depicts the queuing architecture of the Statistical Disk Scheduler used to perform the method of the present invention;

[0014] FIG. 3 depicts a flowchart specification of the SDS Selection Procedure;

[0015] FIG. 4 depicts a flowchart specification of the Scheduling Interval Procedure;

[0016] FIG. 5 depicts a round-robin version of the Scheduling Interval Procedure;

[0017] FIG. 6 depicts a flowchart specification of the Command Completion Procedure;

[0018] FIG. 7 depicts a flowchart specification of the method of the present invention;

[0019] FIG. 8 shows the software process architecture for a preferred multi-threaded implementation of the method of the present invention;

[0020] FIG. 9 depicts a block diagram of an alternative embodiment of the present invention;

[0021] FIG. 10 depicts a timing diagram for enabling access to the system of FIG. 9;

[0022] FIG. 11 depicts a probability distribution diagram for access time versus various extent sizes;

[0023] FIG. 12 depicts an expanded view of the 768 k extent size graph of FIG. 12; and

[0024] FIGS. 13A and 13B together depicts a flow diagram representing the operation of the alternative embodiment of the present invention of FIG. 9;

[0025] To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures.

DETAILED DESCRIPTION

[0026] FIG. 1 depicts a video-on-demand system that utilizes a generic video server incorporating the teachings of the present invention. Specifically, video-on-demand system 100 contains a video server 110 that communicates with a plurality of disks 120 via a Statistical Disk Scheduler (SDS) 170. In addition to the SDS 170, video server 110 contains a CPU 114 and memory element 117. SDS 170 is coupled to disks 120 by paths 130 (e.g., fiber channel), and memory 117 by data path 177. The video server sends access requests along paths 130 to disks 120, and each disk 120 has its own internal queue 125 for buffering access requests. Data read from the disks are transmitted back to the video server along paths 130n (where n is an integer greater than zero). The paths 130n are “daisy chained” to form a data transfer loop 131, e.g., a fiber channel loop. Although one loop is depicted, multiple loops may be employed to interconnect subsets of the disk drives such that the data transfer rate amongst the disk drives and the video server is increased over that of a single loop system. The video server contains a Distribution Manager 180 that receives the data transmitted along paths 130n and loop 131 and distributes this data to subscribers 160 via a transport network 140. Additionally, disks 120 send messages called command completion messages (to be discussed later) to the SDS 170 along paths 130.

[0027] The transport network 140 is typically, but not exclusively, a conventional bi-directional hybrid fiber-coaxial cable network. Subscribers 160 are coupled to the transport network 140 by paths 150 (e.g., coaxial cable). Additionally, transport network 140 forwards subscriber access requests along path 175 to the SDS 170, and receives video data from Distribution Manager 180 via path 185.

[0028] Commonly assigned U.S. patent application Ser. No. 08/984,710, filed Dec. 3, 1997, which is incorporated herein by reference, describes an information distribution system, known as the OnSet™ system, that uses a video server that may benefit from the present invention. Additionally, the video server of the OnSet system is described in U.S. Pat. Nos. 5,671,377 and 5,581,778 which are both herein incorporated by reference.

[0029] The SDS 170 performs the method of the present invention. A logical representation of the SDS data architecture is shown in FIG. 2. In a physical representation, the outputs of each queue are connected to the data loop (131 of FIG. 1). In the depicted embodiment, the SDS queuing architecture contains three queues for each disk 120 and a queue selector 205 for managing queue selection, i.e., the queue selector determines which queue is to receive the next access request for a disk drive. For simplicity, the logical representation is more easily understandable. Although FIG. 2 depicts three queues for each disk drive, a greater or lesser number of queues may be used to fulfill the invention, i.e., at least two queues should be used; one for the “steady-state” requests and one for all other requests.

[0030] In the three queue embodiment of the SDS 170, a steady-state subscriber queue (SSQ) 221 is used for “steady-state” subscriber disk reads for active streams (i.e., continuous content retrieval for distribution to subscribers currently watching a program.) Disk access requests in SSQ 221 are assigned the highest priority. A new subscriber queue (NSQ) 222 is for subscriber requests to begin viewing a program or perform other program related commands, i.e., non-steady state commands such as fast forward or rewind that in essence are a request for a new data stream. Disk access requests in NSQ 222 are assigned medium priority. The other request queue (ORQ) 223 is for all non-subscriber operations, such as loading content, disk maintenance, and file system meta-data syncing. Disk access requests in ORQ 223 are assigned the lowest priority.

[0031] Queues 221n, 222n, and 223n are collectively called the SDS queues 200n, where n is an integer greater than zero that represents a disk drive 120n in an array of disk drives 120. For each disk 120n, the queue selector 205 selects requests from the three SDS queues 221n, 222n, and 223n and forwards the requests to the corresponding disk queue 125n. Each request has an associated worst-case access time based on the type of request and data transfer size. The worst-case access time can be fixed, or dynamically computed based on prior access time statistics. Additionally, each steady-state subscriber request has a time deadline for when the request must complete in order to guarantee continuous video for that subscriber. Disk requests in the NSQ and ORQ generally do not have time deadlines.

[0032] Requests in the SSQ 221n, are ordered by time deadline so that the request at the front of the queue has the earliest deadline. Consecutive SSQ requests with the same time deadline are ordered by logical disk block address according to an elevator algorithm. The elevator algorithm is a disk scheduling algorithm well-known in the art in which the disk head travels in one direction over the disk cylinders until there are no more requests that can be serviced by continuing in that direction. At this point, the disk head changes direction and repeats the process, thus traveling back and forth over the disk cylinders as it services requests. Since requests in the NSQ and ORQ do not generally have deadlines, they may be ordered on a first come first serve basis, or according to some other desired priority scheme.

[0033] In order to keep the disks 120 busy and maximize disk bandwidth utilization, disk command queuing may be employed to ensure that the disk can begin the seek for the next access immediately after it finishes the data transfer for the current disk access. When a steady-state request needs to access a sequence of multiple disks, the request is initially added to the SSQ 2211, of the first disk 1201. After this request is selected for servicing by the first disk 1201, the request is added to the second disk's SSQ 2212 as soon the video server begins sending the data that was recalled from the first disk 120n to the subscriber. Steady-state requests are similarly added to the SSQ 221n, of each successive disk 120n.

[0034] The queue selector 205 employs an SDS Selection Procedure to select requests from the three SDS queues 200n and forward the requests to an associated disk queue 125n located within each of the disk drives 120n. The SDS Selection Procedure uses worst-case access times, request priorities, and time deadlines in determining which request to forward to the disk queue. The general strategy of the SDS Selection Procedure is to select a non-SSQ request only when such a selection will not cause any of the SSQ 221n requests to miss their time deadlines, even if the non-SSQ request and all requests in the SSQ 221n were to take their worst-case access times. If such a guarantee cannot be made, then the first request in the SSQ is always selected. As an optional step, once a request is selected, the SDS Selection Procedure checks whether the data for the selected read request is already in cache (if caching is used). If this is the case, the disk access can be discarded and the Selection Procedure is repeated. Otherwise, the selected request is removed from the SDS queue 221n, and forwarded to an associated disk queue 125n.

[0035] FIG. 3 depicts a flow diagram of the SDS Selection Procedure 300. First, the Selection Procedure checks whether the first entry in the NSQ can be selected while guaranteeing that all SSQ requests will meet their time deadlines in the worst case (step 320), where worst case is defined by the system. Generally, the worst case value is the access value having a per user error rate that is acceptable.

[0036] Each queue maintains “a sum of the worst case values” selector that performs a worst case analysis and selects the queue that will be used (i.e., steps 320 and 330) to send the next command to the disk drive. The following pseudocode represents the operation of such a selector.

[0037] 1) perform worst case analysis returns remaining time (the amount of time left on the SSQ if all commands take worst case time to execute, if the SSQ is empty, the remaining time is infinity)

[0038] 2) if NSQ is !empty && NSQ.head.worstcase<remaining time take request off NRQ else if NSQ is empty && ORQ is !empty && ORQ.head.worstcase<remaining time take request off ORQ else if SSQ is !empty take request off SSQ if request.deadline−request.worstcase>current time request missed deadline, terminate request, try selector again else no requests pending

[0039] Preference is given to the NRQ over the ORQ, only take things off the ORQ if the NSQ is empty.

[0040] The ORQ.head.worstcase and NSQ.head.worstcase are the respective worstcase access times to fulfill the next request in the ORQ and NSQ. The “remaining time” value is computed as follows:

remaining time disk Q Remaining Time (SSQn) - disk Q
worst case (PQn)
disk Q Remaining Time (Q, now) {
sum = 0
min = MAX
for each entry in Q {
sum + = entry → worstcase
left = entry → deadline + sum − now;
if (left < = 0 || entry → deadline > now) { /*
out of time */
min = 0;
break;
}
if (min > left)
mm = left; /* there is now less time remaining
*/
}
return min;
}

[0041] The worst case time value may be dynamically computed or empirically measured to be a cut off time that defines a period in which accesses have an acceptable error rate. If the first entry fulfills the requirement, then this first entry is selected (step 340); otherwise, the Selection Procedure checks whether the first entry in the ORQ can be selected while guaranteeing that all SSQ requests will meet their time deadlines in the worst case (step 330). If so, then this first entry is selected (step 350); otherwise, the procedure proceeds to step 315, wherein the procedure queries whether the first entry in the SSQ can be executed within its time deadline assuming the worst case access. If the request cannot be executed in time, the request is discarded at step 325 and the procedure returns to step 320.

[0042] If, however, the request can be executed in the allotted time, the first entry of the SSQ is selected at step 360. The selected request is then removed from its queue (step 370). Alternatively, if caching is used, the Selection Procedure checks whether data for the selected request is already in cache (step 380) (the caching step 380 is shown in phantom to represent that it is an optional step). If the request is cached, the selected request is discarded and the Selection Procedure is repeated. Otherwise, the selected request is forwarded to the associated disk queue (step 390).

[0043] The SDS executes the Selection Procedure during two scheduling events, called the scheduling interval and the command completion event. The scheduling interval is a fixed, periodic interval, while a command completion event occurs every time one of the disks completes a command. (Note that it is possible, although highly unlikely, that multiple disks complete a command simultaneously at a command completion event.) At each scheduling interval, a procedure called the Scheduling Interval Procedure is executed, and at each command completion event, a procedure called the Command Completion Procedure is executed. In the case that a scheduling interval and a command completion coincide, the Command Completion Procedure is executed first (i.e., the Command Completion Procedure is given priority over the scheduling Interval Procedure). Alternatively, if the disk queue has a depth that is greater than one, then the execution priority of these routines is reversed. Such reversal leaves more time available to do other operations.

[0044] In the Scheduling Interval Procedure, steady-state requests are added to the next SSQ, if possible. (Recall that a steady-state request can be added to the next SSQ as soon as the data is output from the video server to the subscriber), and all SSQs are reordered to maintain correct time deadline order. The first entries in each of the SSQs are then sorted based on time deadlines, which determines the order with which the disks are serviced. For each disk, the Selection Procedure 300 is repeatedly executed as long as the associated disk queue is not full, at least one of the three SDS queues (SSQ, NSQ, ORQ) is not empty, and there is a request in one of the three SDS queues that satisfies the Selection Procedure criteria. For example, if in a three-Disk system when the disk queues are not full the first entry in Disk 1's SSQ has a time deadline of 35, the first entry in Disk 2's SSQ has a time deadline of 28, and the first entry in Disk 3's SSQ has a time deadline of 39, then the disks would be serviced in the following order: Disk 2, Disk 1, Disk 3. Once the disk order has been established, then the SDS Selection Procedure is performed for each disk in that order.

[0045] Generally, in a video server application, the extents for the data are very long (e.g., hundreds of kilobytes) such that the disk queues have a depth of one. In other applications using shorter data extents, the disk queues may have various depths, e.g., five requests could be stored and executed in a first-in, first-out (FIFO) manner. The extent size is inversely proportioned to disk queue depth where data delivery latency is the driving force that dictates the use of a large extent size for video server applications. For other applications where the extent size is relatively small, the disk queue depth is dictated by the desire to reduce disk drive idle time.

[0046] FIG. 4 shows a formal specification of the Scheduling Interval Procedure 400 in flowchart form. First, the Scheduling Interval Procedure adds steady-state requests to the appropriate SSQs, if possible (step 420), and reorders all the SSQs by time deadlines (step 430). The disk that has the earliest deadline for the first entry in its SSQ is then selected (step 450). The Selection Procedure is performed for the selected disk (step 300), and then the Scheduling Interval Procedure checks whether a request satisfying the Selection Procedure criteria was selected (step 460). If not, the disk with the next earliest deadline for the first entry in its SSQ is selected (steps 475, 480, 450) and the Selection Procedure is repeated for this disk (step 300). Otherwise, the Scheduling Interval Procedure checks whether the selected disk's queue is full, or if all three SDS queues for the selected disk are empty. If either of these conditions are true, then the disk with the next earliest deadline for the first entry in its SSQ is selected (steps 475, 480, 450) and the Selection Procedure is repeated for this disk (step 300). If, however, both conditions are false, the Selection Procedure is repeated for the same selected disk. Thus, the disks are processed sequentially, ordered by the corresponding SSQ's first deadline, where “processing” means that the Selection Procedure is invoked repeatedly until the disk queue is full or there are no more requests for that disk.

[0047] As disclosed in FIG. 4, the Scheduling Interval Procedure fills each of the disk queues one at a time, which is most efficient for small disk queues. In the preferred embodiment, a small disk queue is used, as it facilitates the latency reduction. In particular, as soon as the servicing of a request extends past its worst-case access time, the request is aborted by the SDS, i.e., the SDS “times-out” waiting for the request to be serviced and then moves on the next procedural step. To assist in error handling when using a disk queue with a depth that is greater than one such that the server may determine which request was not fulfilled within a predefined time period, the server maintains a disk mimic queue that mimics the content of the disk queue of each of the disk drives. As such, the server can poll the mimic queue to determine the nature of the errant request and send an “abort” command to the disk drive for that request. The disk drive will then process the next request in the disk queue and the server updates the mimic queue.

[0048] In the case of large disk queues, however, filling the disk queues in a round-robin fashion may be more efficient. A round-robin version of the Scheduling Interval Procedure for large disk queues is shown in FIG. 5. As in the previous embodiment of the Scheduling Interval Procedure, steady-state requests are first added to the appropriate SSQs (step 520), and disks are ordered by the deadlines of the first entry in each disk's SSQ (step 530). In this round-robin version, however, the Selection Procedure is executed only once for a disk, and then the next disk is selected. Once all disks have been selected, the round-robin Scheduling Interval Procedure goes through each of the disks once again in the same order, executing the Selection Procedure once per disk. This process is continued until no more requests can be added to any of the disk queues.

[0049] Specifically, a vector D is defined as an ordered list of all the disks, where the order is based on the time deadlines of the first entry in each disk's SSQ (step 530). A Boolean variable SELECT is initialized to false, and an integer variable i is initialized to 1 (step 540). The following condition is then tested: if i=n+1 and SELECT=false (step 550). As will be seen shortly, this condition will only be true when all of the disks have been selected and no requests could be added to any of the disk's queues. Next (555), if i=n+1 (i.e., the last disk had been selected in the previous iteration), then i is set to 1 (start again with the first disk). If disk Di's disk queue is full (step 560), or all three of Di's SDS queues are empty (step 570), then the next disk is selected (step 585). The Selection Procedure is performed for Di (step 300), and if a request satisfying the Selection Procedure criteria was found, SELECT is set to true (step 580), and the next disk is selected (step 585). Thus the SELECT variable indicates whether a request was added to one of the disk queues during a pass over the vector of disks.

[0050] The Command Completion Procedure is executed, on a first-in, first-out basis, every time a disk completes a command. Thus, for each completed command, the Command Completion Procedure executes in the order in which the commands are completed, i.e., using the FIFO command handling step 605. As such, the Command Handling Procedure begins at step 610, proceeds to step 605 and ends at step 690.

[0051] Alternatively, the procedure can be adapted to handle simultaneous command events. In this procedure, it is first determined if multiple disks have completed a command simultaneously at the command completion event. (Most likely only one disk will have completed a command at the command completion event, but the multiple-disk situation is possible.) If more than one disk has completed a command, then the first entries in the SSQs of these disks are sorted based on time deadlines, determining the order in which the disks are serviced. Once the disk order has been established, the SDS Selection Procedure is performed for each disk in order in the same manner as the Scheduling Interval Procedure. That is, for each disk, the Selection Procedure is repeatedly executed as long as the associated disk queue is not full, at least one of the three SDS queues (SSQ, NSQ, ORQ) is not empty, and there is a request in one of the three SDS queues that satisfies the Selection Procedure criteria.

[0052] A formal specification of both forms of the Command Completion Procedure is shown in flowchart form in FIG. 6. Step 605 represents the standard FIFO command handling procedure, while the dashed box 615 represents an alternative procedure capable of handling simultaneous command occurrences. In this alternative version, the Command Completion Procedure determines which disks have just completed a command, and the disk that has the earliest deadline for the first entry in its SSQ is then selected (step 650). Just as in the Scheduling Interval Procedure, the Selection Procedure is performed for the selected disk (step 300), and then the Command Completion Procedure checks whether a request satisfying the Selection Procedure criteria was selected (step 660). If not, the disk with the next earliest deadline for the first entry in its SSQ is selected (steps 675, 680, 650) and the Selection Procedure is repeated for this disk (step 300). Otherwise, the Command Completion Procedure checks whether the selected disk's queue is full, or if all three SDS queues for the selected disk are empty. If either of these conditions are true, then the disk with the next earliest deadline for the first entry in its SSQ is selected (steps 675, 680, 650) and the Selection Procedure is repeated for this disk (step 300). If, however, both conditions are false, the Selection Procedure is repeated for the same selected disk.

[0053] As disclosed in FIG. 6, the Command Completion Procedure fills each of the disk queues one at a time, i.e., the disk with a complete event is refilled. Note that since it is highly unlikely that more than one disk is serviced on a command completion event, the choice of whether to employ round-robin or sequential filling of the disk queues in the Command Completion Procedure has essentially no impact on performance.

[0054] In both the Scheduling Interval and Command Completion Procedures, the ordering of requests within the disk queues are managed by the video server CPU, and not the disks themselves. (Any reordering operations normally performed by the disk must be disabled.) While reordering by the disks would improve the average seek time, managing the disk queues by the CPU is required to preserve the time deadlines of the user requests.

[0055] A formal specification of the method of the present invention is shown in flowchart form in FIG. 7. Whenever a command completion event occurs (720), the Command Completion Procedure is invoked (600), and whenever a scheduling interval occurs (730), the Scheduling Interval Procedure is invoked (400). As shown in the figure, if both a scheduling interval and a command completion event occur simultaneously, the command completion is given priority and the Command Completion Procedure is executed first. Alternatively, as discussed above, when a disk queue having a depth that is greater than one is used, the execution priority for these procedures is reversed.

[0056] In one embodiment, the method of the present invention is implemented as a multi-threaded process. FIG. 8 shows the software process architecture 800 for this embodiment of the invention. The media control thread 810 receives new-subscriber request messages from the transport network 140 and path 175, and forwards these requests through message queues 815 to the Ts thread 820. The Ts thread 820 is a top level scheduler responsible for two primary functions: first, it maintains all state information necessary to communicate with the disk interfaces 835 and video server memory 840; second, it performs the Scheduling Interval Procedure using a period of, for example, 100 ms. The Ts Loop thread allocates the commands to the SDS queues 875, where each disk drive is associated with a set of queues (e.g., ssa, NSQ and other queues) generally shown as queues 8250, 8251, . . . 825N. At the startup condition, when the disks are idle, the initial commands (startup commands) from the Ts loop thread 820 are sent from the SDS queues 825 directly to the disk interfaces 835. Under steady-state operation, a response thread 830 communicates the commands from the SDS queues 825 to the disk drive interfaces 835. Each interface 835 communicates to individual disk drives through a fiber channel loop. Response thread 330 also receives command completion messages from the disk interfaces 835. Upon receiving these messages the response thread performs the Command Completion Procedure. Media control thread 810, Ts loop thread 820, and response thread 830 are all executed by video server CPU 114 of FIG. 1.

[0057] FIG. 9 depicts a block diagram of another embodiment of the present invention comprising a video server system 900. The system 900 comprises a plurality of client's 902 (also referred to as subscribers or users), a server 904, and a plurality of disk drives 906. The server 904 comprises a queue selector 908 and a plurality of queues 910. The queues 910 comprise at least one new user queue (NUQ) 910A and a steady-state queue (SSQ) 910B that are coupled to each disk drive 906. The queue selector 908 operates using a disk scheduling algorithm that is executed by the server 904 that can determine priority (new user vs. steady state) through analysis of the actual distribution of steady state user disk requests, giving the new user request priority when doing so would not jeopardize the requirement that the steady state requests miss no access deadlines, but otherwise giving the steady state requests priority. For each new user request, the queue selector 908 operates only once, at the time the new user requests occur (not every time each disk completes a read). In this case, the NUQ 910A is always given priority over SSQ 910B when disk reads are issued at step 911. In the following description of this alternative embodiment of the invention, the scheduling algorithm shall be referred to as facilitating a “controlled admission policy” for controlling disk drive access.

[0058] One basis for the controlled admission policy is the existence of at least two separate queues 910A and 910B for each disk 906n in a server's disk loop 912. This is shown schematically in FIG. 9, which depicts a server with M clients 902 and D disks 906. Each disk 906 has its own new user queue 910A and steady state queue 910B. When a user (client) “enters” the server by requesting access to a disk drive (i.e., the user requests delivery of a video program), the user's first disk request is performed from the NUQ; all subsequent requests from that user are serviced from the SSQ's (until, that is, the user requests a new file or new file position, in which case the first request for the new file or new file position is serviced from a NUQ).

[0059] Recall that the steady state disk scheduling consists of a user sequentially accessing each disk in the disk loop by enqueuing a data request, i.e., the user request is enqueued for an extent read on disk J at time T, and on disk J+1 at time T+SP and on disk J+2 at time T+2*SP, and so on where SP is the service period within which all steady state users must be serviced. The fact that that user must read from disk J between times T and T+SP means that the user's request for data from disk J is enqueued into disk J's SSQ at time T, and that within the next service period the disk loop must service that request. Although the request is entered into the SSQ at time T, if many clients are using the server, it is likely that the request is not actually serviced for some time, during which the request sits pending in the queue while the request waits for its time to actually read from disk. The existence of the NUQ's (distinct from the SSQ's) allows the server to give steady state user requests priority over pending new user requests.

[0060] In order to perform a controlled admission policy, the server must be able to ascertain whether allowing a new user disk read to proceed ahead of any already enqueued steady state requests (on the same disk) will potentially cause missed deadline problems with the already existing steady state users of the server. In order to be able to make such a determination, the server must know each user's disk deadlines and the extent size of the requested data. (The extent size determines the expected time that the disk read will take.) If there are D disks in the disk loop, each user enqueues a read from each disk periodically once every D service periods. The full set of a user's disk deadlines can thus be summarized with just one number, e.g., the user's first deadline on disk 0, since all the other deadlines are simply periodic extensions from that one. Rather than use actual numeric deadlines, the following discussion uses “enqueue times”, which are one SP before each deadline. Each user's steady state scheduling requirements can thus be fully summarized with two numbers: (1) the extent size E of the data requested by the user, and (2) a normalized time value T0 (such that 0≦T0<D*SP), that tells when the user will enqueue a read from disk 0.

[0061] FIG. 10 schematically shows these two numbers for a full set of all users requesting information from the server. The loop 1000 shown is a time axis, running from time 0 to time D*SP, with the time value D*SP being equivalent to time 0 (and thus the time axis wrapping around to form a circular loop). Each client using the server is represented by a single vertical vector 1002 sitting on the time axis disk loop 1000. The height of each user's vector 1002 is proportional to that user's extent size (and thus to that user's bitrate), and the position that the vector 1002 sits on the loop 1000 is the normalized time T0 at which that user will periodically enqueue a read from disk 0.

[0062] The dashed vector 1004 represents a disk usage request from a new client attempting to enter the server. That user's extent size is E_new and that user's normalized disk 0 request time is T0(new). T0(new) is calculated as follows: suppose the time the server receives the new user request is T_req and suppose the request is for a read from disk 1. Then, if the server were to allow that request to proceed (preempting any already enqueued, but not started, steady state reads) then that user would be requesting a read from disk 0 (D−1) service periods later (since there are D disks in the loop), at time

T 0 (new)_unnormalized=T_req+(D−1)*SP.

[0063] Assuming all times are measured with respect to some time 0 that occurred before any client entered the server, T0 (new) is then simply the remainder of T0 (new)_unnormalized after dividing by D service periods:

T0 (new)=T0 (new)_unnormalized % (D*SP).

[0064] Note that the service period prior to T0 (new) has three steady state users requesting disk 0 reads in it: users 1, 2 and 3 with extent sizes E1, E2 and E3. Similarly, the service period following when the new user would read from disk 0 has three users 4, 5 and 6 requesting disk 0 reads.

[0065] Suppose user 3's request would come 5 msec before the new user's request, and user 4's would come 5 msec after the new user's request, so that

T 0 (new)−T0 (3)=T0 (4)−T0(new)=5 msec.

[0066] Further suppose at the time the server receives the new user's request (T_req), user 3's request is still pending on disk 1's SSQ (the request was enqueued on disk 1's SSQ 5 msec earlier). The goal of the server's controlled admission algorithm is to ascertain whether preempting user 3's pending request on disk 1 with the new user's request, and delaying user 4's, will cause any missed deadlines by any existing server clients (user 3 or 4, or any others).

[0067] The amount of time it takes to read a fixed amount of data from a hard disk is not completely deterministic. In fact, the data access time depends on the distance the hard disk arm, needs to seek, on whether multiple “on-track” attempts need to be made to successfully read serve data, and on the rotational latency experienced once the read head is determined to be on-track. These factors cause the statistical distribution of access times required to read a certain extent size to range over 100 msec.

[0068] FIG. 11 shows such access time distributions for nine different extent sizes E ranging from 4 kB to 2 MB. Of course, as the amount of data read increases, the average access time increases too. But even for a given extent size, the distribution of access times has most of its probability density over a range of 30 msec. In fact, FIG. 12 shows a detailed view of one of the curves (E=768 kB) in FIG. 11, now plotted logarithmically, so that the small probability tail can be seen. It is seen that, on the disks this data was extracted from, the reads of 768 kB extents ranged from taking 33 msec to taking 139 msec. Most of the probability density ranges from 50-80 msec, but there is some small chance that very long access times (over 100 msec) will occur.

[0069] The controlled admission algorithm must thus allow for the stochastic nature of the disk access process. To provide further understanding of this issue, a probabilistic analysis is performed below for the following problem: given a sequence of Q disk requests for extent sizes E1, E2, . . . , EQ, how long must the system wait before it is known that the probability that all Q accesses have completed is greater than 1−z, for some small number z. The total time the system must wait is called T_Q(E1, E2, . . . , EQ). For z=10−6, the probability that all Q reads are not finished after waiting T_Q is smaller than one in a million.

[0070] The analysis is performed for a single fixed extent size E. Let

[0071] p(n,E)=probability that a disk read of extent size E takes between n and n+1 msec. p(n,E) is shown in FIG. 11 as a function of n, for 9 different extent sizes E.

[0072] J_Q(n,E)=probability Q successive accesses of extent size E takes longer than n msec total.

[0073] The following shows a method of calculating J_Q(n,E) given the p(n,E) as input.

[0074] Pi=probability a single access takes between i and i+1 msec.

[0075] Pi=0 for i≧N=>N msec=max access time. 1$\sum _{i=0}^{N-1}{P}_i=1.$2$\begin{array}{c}{I}_1\left(n\right)=\mathrm{probability}\mathrm{of}\mathrm{one}\mathrm{access}\mathrm{taking}\le n\mathrm{msec}.\\ =\sum _{i=0}^{n-1}{P}_i,\left(n\ge 1\right).\end{array}$

[0076] I1(N)=1. I1(0)=0. 3$\begin{array}{c}{I}_2\left(n\right)=\mathrm{probability}\mathrm{of}\mathrm{two}\mathrm{successive}\mathrm{accesses}\mathrm{taking}\le n\mathrm{msec}.\\ =\sum _{i=0}^{n-1}{P}_i{I}_1\left(n-i-1\right)\end{array}$$\begin{array}{c}{I}_Q\left(n\right)=\sum _{i-0}^{n-1}{P}_i{I}_{Q-1}\left(n-i-1\right),\left(\mathrm{for}Q\ge 2\right)\\ =\mathrm{probability}\mathrm{of}Q\mathrm{successive}\mathrm{accesses}\mathrm{taking}\le n\mathrm{msec}.\end{array}$

[0077] Define JQ(n)=1−IQ(n)=probability of Q accesses taking>n msec.

[0078] Then the above equations become 4$\{\begin{array}{cc}{J}_1\left(n\right)=\theta \left(N-1-n\right)\sum _{i=n}^{N-1}{P}_i& \\ J_Q\left(n\right)=J_1\left(n\right)+\sum _{i=0}^{n-1}{P}_i{J}_{Q-1}\left(n-i-1\right)& \left[Q\ge 2,n\ge 1\right]\\ J_Q\left(0\right)=J_1\left(0\right)=1.& \end{array}$

[0079] Here θ(n≧0)=1 and θ(n<0)=0.

[0080] The minimum time to wait for all Q accesses to finish with probability greater than 1−z is T_Q(E, E, . . . , E). T_Q(E, E, . . . E can then be directly obtained from J_Q(n,E) as the minimum “n” such that J_Q(n,E)<z.

[0081] The foregoing analysis can be extended to multiple constant bit rate (mCBR), i.e., to different extent sizes. The first problem to be dealt with is that the extent size can vary continually. In the single extent size (CBR) case, the system only needs the p(n,E) distribution of the one extent size E as input. But in the mCBR case, E can vary, so the system needs to be able to estimate (or interpolate) p(n,E) for any E. As discussed below, this is performed by noting that the system can pick a baseline E (called E_0 there), and estimate the other p(n,E) as rigid shifts of p(n,E_0). With that approximation, the following shows how to calculate T_Q(E1, . . . EQ) given the p(n,E_0) and the shift parameters as input.

[0082] From FIG. 11 it is seen that the p(n,E) curves are roughly shifted versions of the same curve. So assume the system is given a set of pSH(nE) curves that are shifted versions of a baseline psh(nE°) curve:

p sh (n,E)=psh(n−Δ(E),E°)

[0083] A preliminary result needed is that JQ(E1 . . . EQ) is independent of the order of the E1 . . . EQ: 5$I_1\left(n,E\right)=\sum _{i=0}^{n-1}p\left(i,E\right)$

[0084] (n≧1) 6$\begin{array}{c}{I}_2\left(n,E_1{E}_2\right)=\mathrm{probability}\mathrm{that}\mathrm{read}{E}_1\mathrm{then}\mathrm{read}{E}_2\mathrm{takes}\le n\mathrm{msec}\\ =\sum _{i-0}^{n-1}p\left({\mathrm{iE}}_1\right)I_1\left(n-i-1,E_2\right)\\ =\sum _{i=0}^{n-2}[p\left({\mathrm{iE}}_1\right)I_1(n-i-1,E_2]\mathrm{since}{I}_1\left(n=0,E\right)=0.\\ =\sum _{i=0}^{n-2}\sum _{j=0}^{n-i-2}p\left({\mathrm{iE}}_1\right)p\left({\mathrm{jE}}_2\right)\\ =\sum _{i=0}^{n-2}\sum _{j=0}^{n-2}p\left({\mathrm{iE}}_1\right)p\left({\mathrm{jE}}_2\right)\theta \left(n-2-i-j\right)\\ =I_2\left({\mathrm{nE}}_2{E}_1\right),\mathrm{since}\mathrm{the}\mathrm{foregoing}\mathrm{is}\mathrm{symmetric}\mathrm{in}i\to j\end{array}$

[0085] The general result then follows recursively from this.

[0086] Now suppose p(jE2)=p(j−Δ,E1). Then 7$\begin{array}{cc}{I}_2\left({\mathrm{nE}}_1{E}_2\right)=\sum _{i=0}^{n-2}\sum _{j=0}^{n-2}p\left({\mathrm{iE}}_1\right)p\left(j-\Delta ,E_1\right)\theta \left(n-2-i-j\right)& \\ =\sum _{i=0}^{n-2}\sum _{k=-\Delta }^{n-2-\Delta }p\left({\mathrm{iE}}_1\right)p\left({\mathrm{kE}}_1\right)\theta \left(n-2-\Delta -i-k\right)& \{\begin{array}{c}k=j-\Delta \\ k+\Delta =j\end{array}\\ =\sum _{i=0}^{n-2}\sum _{k=0}^{n-2-\Delta }p\left({\mathrm{iE}}_1\right)p\left({\mathrm{kE}}_1\right)\theta \left(n-2-\Delta -i-k\right)& \{\begin{array}{c}\mathrm{since}\\ p\left(n<0\right)=0\end{array}\\ =\sum _{i=0}^{n-2-\Delta }\sum _{k=0}^{n-2-\Delta }p\left({\mathrm{iE}}_1\right)p\left({\mathrm{kE}}_1\right)\theta \left(n-2-\Delta -i-k\right)& \{\begin{array}{c}\mathrm{since}\theta \left(\right)\Rightarrow \\ i<n-2-\Delta -k\end{array}\\ =I_2\left(n-\Delta ,E_1{E}_1\right).& \end{array}$$\mathrm{Now}\mathrm{suppose}p\left({\mathrm{jE}}_2\right)=p\left(j-{\Delta }_2,E_1\right)\mathrm{and}p\left({\mathrm{jE}}_3\right)=p\left(j-{\Delta }_3,E_1\right)\text{:}$$\begin{array}{cc}{I}_3\left({\mathrm{nE}}_1{E}_2{E}_3\right)=\sum _{i=0}^{n-1}p\left({\mathrm{iE}}_3\right)I_2\left(n-i-1,E_1{E}_2\right)& \\ =\sum _{i=0}^{n-1}p\left({\mathrm{iE}}_3\right)I_2\left(n-i-1-{\Delta }_2,E_1{E}_1\right)& \\ =\sum _{i=0}^{n-1}p\left(i-{\Delta }_3,E_1\right)I_2\left(n-i-1-{\Delta }_2,E_1{E}_1\right)& \\ =\sum _{j=-{\Delta }_3}^{n-1-{\Delta }_3}p\left({\mathrm{jE}}_1\right)I_2\left(n-j-1-{\Delta }_2-{\Delta }_3,E_1{E}_1\right)& \{\begin{array}{c}j=i-{\Delta }_3\\ j+{\Delta }_3=i\end{array}\\ =\sum _{j=0}^{n-1-{\Delta }_3}p\left({\mathrm{jE}}_1\right)I_2\left(n-j-1-{\Delta }_2-{\Delta }_3,E_1{E}_2\right)& \{p\left(n<0\right)=0\\ =\sum _{j=0}^{n-1-{\Delta }_2-{\Delta }_3}p\left({\mathrm{jE}}_1\right)I_2\left(n-j-1-{\Delta }_2-{\Delta }_3,E_1{E}_1\right)& \\ \downarrow & \\ [\mathrm{Since}{I}_2\left(n\le 0\right)=0\Rightarrow j<n-1-{\Delta }_2-{\Delta }_3\mathrm{above}.& \\ =I_3\left(n-{\Delta }_2-{\Delta }_3,E_1{E}_1\right).& \end{array}$

[0087] Since JQ(n)=1−IQ(n), the above arguments recursively give the following result:

[0088] If the psh(nE) are probability densities satisfying

[0089] psh(n,E)=0 for n<0, and

[0090] psh(n,Ei)=p(n−Δi,E°),

[0091] then

[0092] JQsh(n;E1 . . . EQ)=JQsh(n−Δ1−Δ2 . . . ΔQ; E°E° . . . E°).

[0093] That implies

[0094] TQsh(E1 . . . EQ)=TQsh(E° . . . E°)+Δ1+Δ2+ . . . +ΔQ

[0095] For example, 8$T_{Q=19}^{\mathrm{sh}}\left[10\left(768\mathrm{KB}\right)+5\left(4\mathrm{KB}\right)+4\left(2\mathrm{MB}\right)\right]=T_{Q=19}^{\mathrm{sh}}\left[19\left(768\mathrm{KB}\right)\right]+5\Delta \left(4\mathrm{KB}\right)+4\Delta \left(2\mathrm{MB}\right)=19\left(75\mathrm{msec}\right)+5\left(-56\right)+4\left(108\right)\mathrm{msec}=1633\mathrm{msec}.$

[0096] This mathematical analysis is combined with simulations to create a critical time T_crit to be used in the controlled admission algorithm, in the following manner. As shown in FIG. 13, when a new user request comes in at time T_req (step 1302), the server at step 1306 first studies the current state of the requested disk's active queue. If a read is active , the estimated time the read will complete is first calculated at step 1308 (using the probability curves shown in FIG. 11) To that expected time that the disk will be free, the method 1300 adds the expected access times for any NUQ requests already in the NUQ for that disk (that are allowed to start by the current calculated time) (step 1310). At this point, the method has the expected time that the disk will be free of higher priority (active and previous new user) requests. Call this time T_new_1. The method at step 1312 now normalizes the time T_new_1 to the disk 0, (0, D*SP) range as explained above, to obtain the normalized time T_new_1_norm. The method 1300 must determine whether insertion of this disk read cause any existing users in the system to miss one of their disk read deadlines

[0097] To answer that question, the system calculates at step 1316 the T_Q value for the new user as follows. The new user itself, and each user with normalized disk 0 access time within one SP greater (i.e., users 4, 5 and 6 in FIG. 10) are counted to form that user's set of Q accesses. The T_Q time is also calculated for each of the users with normalized disk 0 access time within one SP less than the new user (i.e., users 1, 2 and 3 in FIG. 10). Note that the value of Q may vary for each of these users: the new user may have Q=4 in FIG. 10 (including itself, and users 4, 5 and 6), but user 1 may have Q=5 (including itself, the new user and users 2, 3 and 4). Similarly, the T_Q(E1, E2, . . . , EQ) calculation depends on the (potentially different) extent sizes of each of the Q users, so each user calculates a different TQ value. The method can be assured that preempting any steady state reads (by allowing the new user request to proceed at time T_new_1) will not cause missed deadlines (at least, not with probability greater than 1 in several million), if, at step 1318, the T_Q values calculated for each of these users is less than T_crit. If the query at step 1318 is affirmatively answered, the method proceeds to step 1320 where the new user is put in the NUQ for disk d. At step 1326, the method 1300 stops.

[0098] If, at step 1318, one of the T_Q<T_crit comparisons fails, then the method does not allow admission of the new user at that user's optimally requested time (i.e., at T_new_1), for fear that such an admission will generate a missed deadline on an already existing client. The admission algorithm 1300 deals with this case with a multi-tiered method 1300 of FIG. 13B: 1330, If optimal entry is denied, the server at step 1330 determines whether the requested entry point into the new user's data can be re-positioned to lie one service period earlier. If the new user is starting a new track at the beginning, or is requesting data that needs to be delivered at exactly the requested entry point, this entry point shift will be denied, and the algorithm proceeds to step 1344. If the entry point can be backed up, the entry disk is decremented by one in step 1334, and, at steps 1356 and 1358, the entry analysis performed in steps 1316 and 1318 of FIG. 3A is repeated. If entry is denied at step 1358, then step 1340 determines whether a maximum shift in the requested entry point has occurred (e.g., a maximum of 3 disks, equivalent to 3 service periods). If that maximum has not occurred, query 1330 is repeated to see if another backup can occur. (The previous backup may have moved the entry point to the start of the movie track, disabling any further backups.) If that backup is allowed, the procedure is repeated. If not, or if the maximum entry point backup has occurred, the algorithm proceeds to a mode where the entry time is delayed (as compared to the immediate entry at a shifted entry position attempted above). In step 1344, the entry disk is reset to its original value and, in step 1350, an additional delay is added to the user's entry time. Steps 1352 and 1354 are then performed to test for a legal entry in the manner described with respect to steps 1316 and 1318 of FIG. 3A. The process is continued until a legal entry position is found for the disk request. The process then proceeds to step 1360 where the new user is assigned to a NUQ for disk d. The process then stops at step 1362. Besides being the basis for the server's user admissions policy, the algorithm 1300 also allows for a method to determine the maximum load of the server. In particular, simulations can be performed, using this algorithm and real disk access time probability distribution curves, to determine at what client load level the distribution of latencies statistically experienced by all the users becomes unacceptable.

[0099] While this invention has been particularly shown and described with references to a preferred embodiment thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A queuing architecture for scheduling disk drive access requests in an information server having a plurality of disk drives arranged in a disk loop, comprising: a plurality of first queues for disk access requests from users currently receiving information provided by the information server, where each of the first queues are coupled to a disk drive; a plurality of second queues for all other disk access requests, where each of the second queues are coupled to a disk drive; and a queue selector for selecting a particular first queue or a particular second queue for enqueueing a request for data, where said selection is defined by a controlled admission policy.

2. The queuing architecture of claim 1, where said controlled admission policy ensures that fulfillment of a request is not excessively delayed.

3. The queuing architecture of claim 1, where said controlled admission policy defines a critical time period within which a new user request must be fulfilled, if said new user request can be fulfilled within the critical period, the particular request is enqueued in a particular second queue; otherwise, the controlled admission policy enqueues the request in a second queue of another disk drive.

3. The queuing architecture of claim 1, where said queue selector gives highest priority to requests in said first queue and lower priority to requests in said second queue.

4. The queuing architecture of claim 1, wherein said information server is a video server and said information is a video program.

5. A method of scheduling access requests for a plurality of disk drives in an information server, said method comprising the steps of: (a) providing at least two queues for each disk drive, where a first queue contains steady state access requests from users being supplied information from the information server and a second queue for all other types of access requests; (b) selecting a particular second queue to enqueue a new user access request for a particular disk drive, if such selection does not cause a steady-state access request to miss a time deadline within which the steady-state access request must be completed to ensure that the information being viewed is not interrupted; (c) otherwise, selecting another second queue to enqueue the new user access request.

6. The method of claim 5 further comprising the step of enqueuing each steady-state access request in the first queues as the steady-state access requests arrive.

7. A method of accessing data in a disk array comprising: receiving a new user access request; assigning the new user access request to disk d; determining when new user access request will be processed, if the new user access request will be processed within a predefined period of time, placing the new user access request into a new user queue for disk d; otherwise, assigning the new user access request to another disk.

8. The method of claim 7 wherein said another disk is up to three disks distant from disk d.

9. The method of claim 7 wherein, if said new user cannot be assigned to another disk, waiting a predefined period of time; and assigning the new user access request to disk d.

10. In a video on demand system comprising a plurality of disk drives, where video information is striped across the plurality of disk drives, a method of accessing the video information comprising: receiving a user access request for a next portion of video information required by said user to present a streaming video sequence to the user; assigning a user to a disk d that contains the next portion of video information; determining when the user access request will be processed, if the user access request will be processed within a predefined period of time, placing the user access request into a new user queue for disk d; otherwise, assigning the user access request to another disk.

11. The method of claim 10 wherein said another disk is up to three disks distant from disk d.

12. The method of claim 10 wherein, if said user cannot be assigned to another disk, waiting a predefined period of time; and assigning the user access request to disk d.