On Thu, Feb 21, 2008 at 2:27 AM, Alan Altmark <[EMAIL PROTECTED]> wrote:
> Hard evidence of the performance of emulated FBA can be found at: > http://www.vm.ibm.com/perf/reports/zvm/html/530scsi.html > > Don't just look at the tables; read the text that goes with them. Since I have not been able yet to measure it myself, I've hesitated before responding to this thread. So you're excused if you want to skip my long post with no real numbers... ;-) As with some of the measurements from others, the significance for Linux on z/VM installations isn't as much as we would hope. The most significant part in these is a potential reduction in CPU usage. That's good news. The "native SCSI" was meant as a realistic option to hold the z/VM data while Linux has its data on FCP directly. That is also the motivation for the measurements in the performance report. The context that Sir Santa introduced is to have Linux run on emulated devices. A simple iozone test does not address the complexity of the performance issues involved. Most of the measurements folks seem to do are "top speed" measurements (like maximum single-user throughput). With Linux on z/VM we want to tune the system to deliver efficient *multi-user* throughput. We rarely care to tune the system such that a single user can get all resources for itself, because in most real-life cases there will be sufficient others who also need resources and need to make progress. Normally that means lower single-user throughput because you introduce latency. If you care for analogies: it's about measuring mileage versus top speed. Many of the algorithms involved do not behave in a linear fashion. Driving a Linux virtual machine at saturation level provides little information about what it will do at lower utilization levels that you may have in real life. We've seen both extremes: code that gets more efficient at high levels as well as code that performs worse under pressure. If you're careful, you can measure the effect of configuration changes on your own workload. But it does not mean you could predict the effect for other workloads. I once counted Linux on z/VM disk I/O to have 17 layers where we block, cache, queue, skip or reorder the I/O. Some of these layers have instrumentation, but many don't. Trying to understand that with benchmarks is like doing weather forecast by looking out of the window. Rob -- Rob van der Heij Velocity Software, Inc http://velocitysoftware.com/