I am working with Autodesk Revit; running on Windows 7 and Parallels, on an i7 iMac. I have been doing some benchmarking lately to try to figure out what the best Parallels settings are, and I am noticing some very strange results. Attached is a PDF of some data. The benchmark times various tasks in Revit that are executed via a journal file. I did multiple runs on Bootcamp as a baseline, then more runs in Parallels with 8,4,2 & 1 CPU assigned to the VM. Each run's numbers where averaged and totaled. The first thing I noticed is the variation in times. I ran the benchmark in Bootcamp 5 times, and things varied by 1 or 2 seconds, even when the task took nearly a minute to complete. In Parallels I ran the benchmark 3 times for each tested CPU configuration, and sometimes I got times ranging from 30-50 seconds, for a task that in Bootcamp all five runs where 21 or 22 seconds total. So, why so much variation in the times? I would think this kind of performance creep would be murder when playing games. Is this expected results? And is there anything that can be done to stabilize performance a bit? Also, there where some tasks where Parallels was just as fast as Bootcamp, and others where Parallels was much slower. I assume this is just the nature of the work and nothing in the settings will affect it, but maybe not? Lastly, there are some places where the times make perfect sense, Bootcamp is fastest, then Parallels with 8 cores, on down to Parallels with 1 core. But there are other places where Boot camp and Parallels w/ 8 cores come in at 1 & 2, and then Parallels with 1 core will be next fastest, followed by Parallels with 2 cores, and then 4. there are even some parts of the benchmark where Parallels with 1 core is faster than Bootcamp! But it isn't a reverse of the pattern (i.e. Parallels with 2 cores isn't next fastest, 4 cores is!), as if the task is very single threaded and Turbo Boost is impacting things. Any thoughts or explanations? Very perplexing results, to be sure. Thanks! Gordon
I believe it is hard to answer all your questions without very deep workload analysis: 1. from our experience, up to 10-20% of deviation may be due to different services running in background (like disk indexing etc.). When you have VM, you actually have 2 instances of such services - in Mac OS + Windows. When we benchmark we stop all services in both host/guest. 2. if you run VM not on the same Bootcamp partition, then obviously it uses different disk parts with different performance. 3. Some multithreaded tests can really run faster on UP then SMP. This is due to overhead of inter-core communications (like IPI). On real hardware cost is much less noticeable. 4. If you do not limit Bootcamp in memory, then your Bootcamp runs obviously have more RAM vs. VM, and may cache data more efficiently. 5. In general I see that total results do not differ that much for VM and Bootcamp (<20%).
