I am trying to compare architectures accoss a few different platforms running at different frequencies. I am having a hard time trying to figure out how to normalize this in a way that will withstand scrutiny.
- I am running a static workload that takes 1.5hrs on one processor and as long as 64hrs on some
of the others.
- I have a core clock rate as low as 400MHz and as fast as 2.13GHz.
The execution times and clock rates vary significantly so I am not sure how to present data on something like an L2 miss.
The way i have ran my oprofile script is to normalize the sample time to the fastest processor which collects data for 100s. The slowest processor (400MHz - FPGA based) is collects data for 532 seconds. The idea here is collect data during a proportional number of core clock cycles. However since the architecture of the fastest clock is also much faster, i get more L2
misses because the processor is further along in the work load execution. Therefore I am not sure how to plot this data as is probably going to be misleading.
Has anyone dealt with an issue like this yet?
Thanks for your time,