SUNFLOWER Simulator: A few Questions?

[Yasir]

I am working on a research paper targeted to compare a mathematical
model of a cache energy and throughput with a real-world system. I
have come across SUNFLOWER simulator and I must appreciate you for
making it open source. I am planning to carry out my work on this but
have a few questions if you can answer please
1. When I disable cache in the simulator the Power consumption remain
the same as of a 100% miss in cached system. I assume that disabling
the cache should make it power gated-off so that it does not
contribute towards processor power consumption. And in that case the
power consumption shall be lesser than the case when its ON.
2. There are a number of precompiled benchmark binaries of SPEC 2000
benchmarks. I don’t have any license of SPEC 2000 benchmarks. Is it ok
for me to publish results based on your compiled binaries. (I will
sure be acknowledging your work)
Regards

[Phillip]

Hello,

Sunflower implements three forms of energy estimation--(a) one based
on
an instruction-level hardware measurement, (b) a circuit activity
estimation,
and (c) energy prediction based on coarse-grained active/idle
measurements
(e.g., taken from a processor data sheet). These estimation modes are
useful primarily for the embedded platforms.

To get specifically to your questions:
1. When using the instruction-level power estimation, disabling the
cache simply disables the functionality of the cache, but effectively
leaves it electrically "on".

The rationale is, if you want to model cache power dissipation
realistically,
you would need to know what process technology the specific cache
is implemented in. For realistic leakage power estimates you would
also
need to know, e.g., what temperature the die will be operating at.
Any
realistic estimation would therefore be better off modeled using,
e.g., the
CACTI or eCACTI tool. If you wanted, you could do the analysis in
CACTI and plug the values back into Sunflower.

Perhaps, an interesting project would be to implement an interface
between CACTI and Sunflower, along the lines of the ideas expressed
in http://drops.dagstuhl.de/opus/volltexte/2007/1107/pdf/07041.SWM.ExtAbstract.1107.pdf



2. As far as I know, you should be able to run the SPEC benchmark
binaries without a license (other simulators also distribute compiled
SPEC binaries, and I believe it is OK).



cheers,
phillip

[Phillip Stanley-Marbell]

At the completion of simulation, a brief summary will be printed
at the console output (and details written to sunflower.out). At
this point, the console version of the simulator will exit.

At any point, you can use the "d" or "dumpall" command to force
the generation of a statistics output file.

Best regards,
phillip

> Will the simulator notify me that the benchmark execution is over or
> do I have to quit it manually using OFF command (and if that is the
> case how can I be sure when the benchmark has completed execution).

[Yasir]

Hi
In connection to this mail. May I ask whether the power overhead in
terms of cache accesses is incorporated in the simulator or do we have
to find out that using CACTI and add to the total power consumption.
Regards

[Phillip Stanley-Marbell]

Hello,

The instruction-level power estimates are based on
measurements on an actual processor, and therefore
account for all components for _that_ processor.

For any realistic analysis of another system in which
cache power dissipation plays an important role, you
should use dedicated cache structure power estimating
tools like CACTI as a start.

Best regards,
phillip