The TPC-C Trace

The eager-writing arrays are designed to target TPC-C-like transaction applications. We evaluate the effectiveness of EW-Arrays with a trace supplied by HP Labs. It is a disk trace of an unaudited run of the Client/Server TPC-C benchmark running at approximately 1150 tpmC on a 100-warehouse database. It has 4.2 million I/O requests, 54% of which are reads. The I/O rate is about 700 I/Os per second in the steady state. Most of the requests are synchronous I/Os. The total data set is about 9 GB, distributed originally on 54 disks to achieve the desired throughput. The trace was collected on 5/03/94. Two characteristics of the trace may be of concern due to its old age: the data rate and the size of the data set. With comparable number of disks and machines, the current technology can support a much higher data rate. To account for this development, in some of the following experiments, we raise the I/O rate by multiplying it with a ``trace speedup'' factor. For example, when the trace speedup is two, we halve the inter-arrival time of requests. The data set size factor is of less concern. In fact, only a small fraction of the space on the original traced disks was used to achieve the target I/O rate. Although a single modern disk can accommodate the entire traced data set today, it cannot support the data rate of the original trace. We shall vary the number of disks employed in a disk array onto which the traced data set is distributed. We study the effectiveness of various array configurations and the conclusions that we reach are independent of the size of the entire data set.