Figure 15 – Single disk unbuffered WCE throughputs for basic and dynamic volumes. Basic and dynamic volumes have similar unbuffered WCE throughput. For clarity, three, four, and eight deep reads have been omitted from these graphs as their results were identical to that of two deep. Similarly, two, three, four, and eight deep writes and WCE writes have also been omitted, as they were identical to the one deep results. Dynamic volumes have up to 25% more overhead than basic volumes at small request sizes. At 2KB requests, basic volumes show a 10% higher throughput when compared to dynamic volumes. Throughput for reads deeper than one-deep show improved throughput due to increased parallelism. Unbuffered writes show no improvement from increased request depth. The write plateau is a property of the disk, as explained later.
Small requests (2KB and 4KB) are expensive: 2KB requests show an overall 40% utilization across both processors in Task Manager. Small reads on dynamic volumes have more CPU consumption than small requests on basic volumes, as the fixed overhead per request is higher. Figure 16 shows as request size is increased, the difference in overhead between basic and dynamic volumes decreases. This implies that dynamic volumes have a higher fixed cost per request than basic volumes, but the incremental costs per byte thereafter are similar. Our measurements support this theory, as the fixed cost for reads on a basic volume was 50s per request. There was a marginal cost of .48s per KB. For dynamic volumes, the fixed cost was 58s per request while the marginal cost was .43s per KB. Dynamic volumes have a higher overhead per request, but as the fixed request overhead gets amortized over more bytes, the overheads between dynamic and basic volumes become similar. Still, buffered IO has twenty times more processor consumption than unbuffered IO.
|
