network: DSENT power model
Review Request #1818 - Created April 9, 2013 and submitted
| Information | |
|---|---|
| Owen Chen | |
| gem5 | |
| default | |
| Reviewers | |
| Default | |
| beckmann, tushar | |
network: DSENT power model This patch removes Orion 2.0 and adds the DSENT to the garnet network as the new NoC power/area model. The power model reads configurations from python scripts and GARNET, and generates power estimation for NoC based on the performance stats provided by GARNET. DSENT assumes tile-based multicore design to calculate link length (see topology scripts). Users can specify the clock frequency and various electrical technology node from python scripts.
Posted (April 11, 2013, 2:21 a.m.)
Wow, that is a rather extensive patch. It looks like a good step in the right direction though, and a step up from Orion. It will take some time to go through this, and before delving into the details, the high-level impression I get is: 1) It is far from the ~2:1 code:comment ratio of the gem5 code base (260k:120k), more extensive documentation would certainly be a big plus. 2) The integration with the gem5 clocking system is not there presently, and all the DSENT internal keeping of frequencies etc should be replaced with the native clocks (ClockedObject etc) 3) Similar to the previous remark, gem5 stats framework should probably be used for all the tracking and also the power accounting/output (in my opinion). 4) As above, the same can be said for debug/trace output. 5) The libutil seems rather redundant from a first look. Perhaps I missed something, but it seems to me STL and gem5 already provides most of this functionality 6) There are a number of style issues that'd be nice to see fixed, e.g. the 80 char lines. Thanks for all the effort in making this happen.