6.  Conclusions

      Peak available throughput is only one criterion in most storage system purchasing decisions. Most of the VAX UNIX systems we are familiar with are not I/O bandwidth constrained. Nevertheless, an adequate disk bandwidth is necessary for good performance and especially to preserve snappy response time. All of the disk systems we tested provide more than adequate bandwidth for typical VAX UNIX system application. Perhaps in some I/O-intensive applications such as image processing, more consideration should be given to the peak throughput available. In most situations, we feel that other factors are more important in making a storage choice between the systems we tested. Cost, reliability, availability, and support are some of these factors. The maturity of the technology purchased must also be weighed against the future value and expandability of newer technologies.

      Two important conclusions about storage systems in general can be drawn from these tests. The first is that buffering can be effective in smoothing the effects of lower bus speeds and bus contention. Even though the UDA50 is located on the relatively slow UNIBUS, its performance is similar to controllers located on the faster processor busses. However, the SC780 with only one sector of buffering shows that little buffering is needed if the underlying bus is fast enough.

      Placing more intelligence in the controller seems to hinder UNIX system performance more than it helps. Our profiling tests have indicated that UNIX spends about the same percentage of time in the SC780 driver and the UDA50 driver (about 10-14%). Normally UNIX uses a disk sort algorithm that separates reads and writes into two seek order queues. The read queue has priority over the write queue, since reads cause processes to block, while writes can be done asynchronously. This is particularly useful when generating large files, as it allows the disk allocator to read new disk maps and begin doing new allocations while the blocks allocated out of the previous map are written to disk. Because the UDA50 handles all block ordering, and because it keeps all requests in a single queue, there is no way to force the longer seek needed to get the next disk map. This disfunction causes all the writes to be done before the disk map read, which idles the disk until a new set of blocks can be allocated.

      The additional functionality of the UDA50 controller that allows it to transfer simultaneously from two drives at once tends to make the two drive transfer tests run much more effectively. Tuning for the single drive case works more effectively in the two drive case than when controllers that cannot handle simultaneous transfers are used.