I am working on a framework that allows applications to perform per-rank memory allocation. I would like to use Intel PCM to evaluate the DRAM power consumption of various configurations, but I am struggling to understand some of the measurements reported by the PCM tool.
For each interval, the PCM tool reports the 'CKE Off Residency' and 'CKE Off Average Cycles' for each rank, as well as the DRAM power consumption (in Watts). I was expecting that these measurements would all be directly correlated, but that is not what my experiments are showing. Specifically, I find that the 'CKE Off Average Cycles' may decrease by an order of magnitude (compared to an idle configuration), while the 'CKE Off Residency' shows that the rank is mostly idle. What exactly are each of these numbers reporting and is there an explanation for the apparent discrepancy?
Here are the details of my experimental configuration:
With these experiments, I am trying to determine the potential power savings of per-rank allocation. I created a set of experiments that allocate and access memory in different ways. My three configurations are shown below:
mcolor-t0t0: four processes continually access memory on only rank 0 mcolor-t0t1: two processes access memory on rank 0, two processes access memory on rank 1 idle: no activity
I find that, although the 'CKE Off Residency' seems to show that rank 1 is mostly idle for the mcolor-t0t0 experiment, the 'CKE Off Average Cycles' value decreases by an order of magnitude compared to an idle system. The power consumption of mcolor-t0t0 is not significantly different than mcolor-t0t1. If you would like me to post the actual output from PCM for each configuration, I can do that as well.
I am using a Dell T5600 with an Intel E5-2603 processor, with 4 4GB DIMMs of DDR3 SDRAM, each connected to its own channel.