Skip to main content

CDC performance on PowerPC

1 reply [Last post]
Joined: 2004-03-10

Having managed to get the CDC to build on a PowerPC platform, I decided to run some tests to gauge the resulting floating point performance. The results stunned me - on x86 I was seeing a 10x slowdown compared with the desktop client VM, which I put down to the way hardware floating point is implemented on that platform. However, on the PowerPC I am seeing a 30x slowdown, even with all of the relevant optimisations turned on! Is there anything in the PowerPC CDC that could account for this? I really was hoping it would be somewhere in the ballpark of the desktop vm...

Reply viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.
Joined: 2006-10-16

CDC performance is in the ballpark of Desktop VMs. Please send us details on how you build and run CVM, and also details on your platform and the benchmarks you are running.

On x86, yes FP support is missing from the JIT. This won't show up in most benchmarks, and those that don't do much FP should not end up 10x slower than a desktop VM. Once again, we need build, execution, and platform information.

There are also other things to consider. The PowerPC port won't do as well compared to a desktop VM on something like a G5, which has many deep pipelines (more and deeper pipelines than a G4). I noticed a big difference in the relative performanc of CDC and J2SE on a PowerMac G4 vs. a PowerMac G5. CDC on a G4 was something like >80% of JaveSE, but on a G5 it was more like 50%. I believe this is mainly due to the lack of instruction scheduling in the CDC JIT.