MITRA Introduction -- One Disk Configuration

Go backward to Experimental Design
Go up to Performance Evaluation
Go forward to Multi-disk Configuration

One Disk Configuration

We analyzed the performance of Mitra with a single disk to observe the impact of: 1) alternative mode of operation with the PM-driven scheduling paradigm, 2) block size, 3) different number of groups with GSS, and 4) multiple zones. In the first experiment, we configured the system with 384 KByte block size, g=1, a single zone, with the low and high water marks set to 1 and 2 respectively. In theory, the number of guaranteed simultaneous displays supported by our target disk is 12. This is computed based on the transfer rate of the slowest zone, i.e., 18 Mbps, to capture the worst case scenario where all blocks retrieved during a time period reside in this zone. Mitra realized the theoretical expectations successfully. However, during a time period, the referenced blocks might be scattered across the disk surface, causing the system to observe the average disk transfer rate (26 Mbps). This results in a number of idle slots per time period. As discussed in the PM-driven Section, the PM-driven scheduling paradigm may treat these idle slots in two ways: either (1) busy-wait to exhaust the current time period, or (2) proceed with the retrieval of blocks that should be retrieved during the next time period. The second paradigm results in a lower average latency as compared to the first (0.3 seconds as compared with 2.4 seconds). This paradigm enhances the probability of a new request locating an idle slot during the current time period. Note that this requires more memory at the PM as compared with the first. The amount of required memory can be controlled by the choice of a value for the high water mark.

In the second experiment, we changed the block size from 32 KByte to 64, 128, and 256 KByte. (The remaining parameters are unchanged as compared with the first experiment.) As the block size increases, Mitra supports a higher number of simultaneous displays (6, 8, 10, and 12 displays, respectively). The maximum number of simultaneous displays supported by the available disk bandwidth is 13 and can be realized with a block size of 625 KByte(8). The explanation for this is as follows. With magnetic disks, the block size impacts the percentage of wasted disk bandwidth attributed to seek and rotational delays. As the block size increases, the impact of these delays becomes less significant, allowing the disk to support a higher number of simultaneous displays [GVK+95].

The number of groups (g) with GSS impacts the seek times incurred by the disk when retrieving blocks during a time period. In general, small values of g minimize the seek time. The number of groups (g) has an impact with small block sizes where the seek time is significant. This impact becomes negligable with large block sizes. For example, with a 64 KByte block size, Mitra supports 6 displays with six groups, 7 displays with three groups, and 8 displays with one group. However, with a 384 KByte block, Mitra supports 11 displays with eleven groups, and 12 displays with one group. With this block size, the impact of a choice of value for g is only one display because the disk seek time has been rendered insignificant with the retrieval time of a large block.

In a final experiment, EVEREST was configured to recognize all the 23 zones of the disk. The block size was 539 KByte to guarantee a continuous display with FIXB. In this case, Mitra can store only twelve clips (instead of 22) on the disk because once the storage capacity of the smallest zone is exhausted, no additional clips can be stored (due to a round-robin assignment of blocks to zones). With this configuration, Mitra supports 17 displays with an average startup latency of 35.9 seconds. The higher number of simultaneous displays (as compared to 12 in the previous experiments) is due to the design of FIXB that enables Mitra to harness the average disk transfer rate. The higher startup latency is because a display must wait until the zone containing its first block is activated. The number of logical zones recognized by Mitra is a tradeoff between the number of displays supported by the system, the average startup latency and the percentage of wasted disk space. We now report on several experiments that demonstrate this tradeoff. In the first experiment, we configured EVEREST to recognize two logical zones (the first logical zone consists of zones Z₀ to Z₁₁ while the second consists of the remaining physical zones). In this case, Mitra can store 15 clips on the disk. With this configuration, while the number of simultaneous displays is reduced to 14, the average startup latency is reduced to 0.22 seconds. In a second experiment, we configured EVEREST to recognize one logical zones consisting of only the nine outermost zones (eliminated the remaining 14 innermost zones). With this configuration, Mitra can store twelve clips on the disk. This increases the transfer rate of the disk drive from a logical perspective, allowing Mitra to support 19 displays with an average startup latency of 2 seconds. The higher startup latency is due to a longer duration of a time period. In [GKSZ96], we detail a planner that determines system parameters to satisfy the performance objectives of an application (it desired throughput and maximum startup latency tollerated by its clients).