We had another good question about performance comparisons with regard to our study. As we have maintained, we have not attempted to produce a head to head comparison. JVMs execute code differently and we wanted to be free of such direct comparisons. We were more interested in the parallel performance characteristics relative to an internal reference.
There are clearly two boundaries of performance - the ideal case where n tasks executed by n parallel CPUs complete in the same time as 1 task on 1 CPU; and the parity case where n tasks executed by n parallel CPUs complete in the same time as n tasks on 1 CPU, serially. These provide a frame for gauging observed performance. With this in mind, we have produced a technical note that shows how this visual aid can help someone determine at a glance the meaning of performance results, using our previous results as examples. Although again, we caution that these are simple tests, we hope the examples provide ample illustration of our boundary criteria and demonstrates an isolation from value for value comparison. You can find the document at http://www.amitysolutions.com.au/documents/Threads_aid-technote.pdf. We hope this answers adequately questions raised on direct comparison of performance results. Regards, JonB.
smime.p7s
Description: S/MIME cryptographic signature
