Playing the Trace at Accelerated Rates

Figure 11: Comparison of TPC-C throughput^M on alternative disk array configurations as we vary the I/O rate. The^M total number of disks is a constant (36). The size of the delayed^M write buffer is 10,000 blocks.

We play the TPC-C trace at accelerated rates in the experiments reflected in Figure 11. It compares I/O response times of various array configurations as we maintain a constant total of 36 disks for each configuration. (Note that the $D_s \times D_r = 36 \times 1$ SR-Array configuration is simply a conventional 36-way stripe.) In these experiments, we perform replica propagation in the background. (A maximum of 10,000 blocks can be buffered on an array for background propagation.) The EW-Array has the best response time under all arrival rates and it generally delivers much higher sustainable throughput than conventional configurations. For example, the maximum sustainable throughput rates (expressed in terms of the speedup rate over the original trace speed) on a 36-way stripe, a RAID-10, a $2 \times 9 \times 2$ EW-Array, and a $1 \times 22 \times 1.6$ EW-Array are approximately $5 \times$, $8 \times$, $10 \times$, and $14 \times$, respectively. As we raise the request arrival rate, idle time becomes more scarce and the replica propagation cost becomes felt by all configurations. We must successively reduce the degree of replication for both the SR-Array and the EW-Array. Thanks to the very low write latency of the EW-Array, however, the replica propagation burden on the EW-Array represents a much lighter load. The $2 \times 9 \times 2$ EW-Array remains a configuration of choice that offers sub-5 m$s$ response times for an arrival rate that is as high as $9 \times$ the original, a rate that has rendered replica propagation a costly luxury that the other approaches cannot afford. The payoff of replication is better read response time than that on the other configurations. In addition to eager-writing, two other factors contribute to the EW-Array's superior throughput. One is the greater flexibility afforded by its SATF-EW local disk scheduler. The other is the better load-balancing opportunities afforded by the array-wide scheduling heuristics as writes are scheduled to disks with shorter queues and reads are serviced by choosing among multiple candidate copies on different disks.