Yi Shen
2018-09-28 04:25:03 UTC
Hi all,
I am using Gem5 to study power delivery network (PDN). For this purpose, I need to process periodic Per CPU Cycle stats. For example, I need to dump stats generated in 1 cpu-cycle every 1000 cpu-cycles.
I am aware of 2 potential ways of doing it, but I dont think they work out of box in my case.
1. In python script, import periodicStatDump from m5.internal.stats
The problem is that this method will dump all cycles, without sampling. Instead of getting 1 cycle every 1000 cycles, I will get all 1000 cycles. It is clearly a huge problem for CPU cost and storage.
1. In rsC script, use m5 utilities.
It seems that this method could do the sampling and periodic dump. However, it only accepts input in nanosecond. Which means the fastest CPU I get is pathetically 1 GHz. I couldnt get per-cycle stats from any CPU operates faster than that. ( I tested that m5 dumpresetstats does not work with delay=0.5).
How can I get this done?
Thank you,
Yi Shen
I am using Gem5 to study power delivery network (PDN). For this purpose, I need to process periodic Per CPU Cycle stats. For example, I need to dump stats generated in 1 cpu-cycle every 1000 cpu-cycles.
I am aware of 2 potential ways of doing it, but I dont think they work out of box in my case.
1. In python script, import periodicStatDump from m5.internal.stats
The problem is that this method will dump all cycles, without sampling. Instead of getting 1 cycle every 1000 cycles, I will get all 1000 cycles. It is clearly a huge problem for CPU cost and storage.
1. In rsC script, use m5 utilities.
It seems that this method could do the sampling and periodic dump. However, it only accepts input in nanosecond. Which means the fastest CPU I get is pathetically 1 GHz. I couldnt get per-cycle stats from any CPU operates faster than that. ( I tested that m5 dumpresetstats does not work with delay=0.5).
How can I get this done?
Thank you,
Yi Shen