McPAT version 1.0

[Alexandru Iordan]

Hello

On the 3rd of September, the latest version of McPAT was published (https://code.google.com/p/mcpat/downloads/list). I know you did some modification to version 0.8 in order to make it compatible with the Intel CPUs you model in Sniper. Can you provide me with the necessary information so I can modify version 1.0 of McPAT and use it for power estimation with Sniper?

Alexandru Iordan

[Wim Heirman]

Hi Alexandru,

You can find all patches we made to the 0.8 version of McPAT at [1]. They don't apply directly to 1.0 though, we're still working on porting them. However these patches aren't required, you can use an unmodified version of McPAT as well (it will just be slower since the CACTI results won't be cached). Also the voltage patch is unnecessary as they have exposed Vdd in the XML file as well -- although using this feature will require you to modify tools/mcpat.py to support this. Once you compiled your own version of McPAT, drop it in the sniper/mcpat directory, making sure to remove all other versions (see the mcpat_bin() function of mcpat.py, which tries a few options depending on 32/64 bit etc).

Regards,

Wim

[1] http://snipersim.org/w/McPAT

--
--
--
You received this message because you are subscribed to the Google
Groups "Sniper simulator" group.
To post to this group, send email to snip...@googlegroups.com
To unsubscribe from this group, send email to
snipersim+...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/snipersim?hl=en
 
---
You received this message because you are subscribed to the Google Groups "Sniper simulator" group.
To unsubscribe from this group and stop receiving emails from it, send an email to snipersim+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Ermao Cai]

Hi Wim,

I have the same question. I hope to use the McPAT 1.0 version. When I put the new compiled binary file mcpat into the directory sniper/mcpat directory and modified the function mcpat_bin() in the mcpat.py to make sure it would execute my mcpat, I found there was no valid output from mcpat when running Sniper. 

Could you tell me where the problem may be in?

Thanks a lot!

Ermao

[Wim Heirman]

Ermao,

What is the output that mcpat does give? Is there some other output to the terminal, and what is the contents of power.txt?

Regards,

Wim

[Ermao Cai]

Hi Wim,

Thanks for your reminding. I checked with McPAT v0.8. It ran perfectly with replace of orginal mcpat binary files in directory sniper/mcpat. But McPAT v1.0 still failed in such a try. When I checked with the output file. In the power.txt, there was only such a sentence "some value(s) of number_of_cores/number_of_L2s/number_of_L3s/number_of_NoCs     is/are not correct!", nothing else. I never changed the value of any parameter. And I used 4 the number of cores. And I checked with the *.xml files for cases both in v0.8 and v1.0. And I found both of them are same 

number_of_cores    4

number_of_L2s       4

number_of_L3s       1

number_of_NoCs    1.

However, v0.8 worked and v1.0 didn't.

How should I solve this issue?

Thanks,

Ermao

[Wim Heirman]

Ermao,

It looks like McPAT 1.0 no longer accepts the system.mem component in the XML input file. You can remove this by removing this part in tools/mcpat.py (should be lines 1075 to 1098). As far as I know this component wasn't even used as McPAT doen't provide memory power values, and anyway the mcpat.py script already contains its own DRAM power model. I don't think anything else is needed for McPAT 1.0 support, at least a few quick runs I did here return sensible values. I haven't yet made a full comparison of results between 0.8 and 1.0, if you will be doing one I'd be interested in hearing about the results.

Regards,

Wim

[Ermao Cai]

Hi Wim,

I used my version of McPAT 0.8, which I modified from the source code provided by HP for specific use. And it is quite slow compared to the one you provided in sniper/mcpat. It is because you cache the CACTI results. I just wonder that what I should do to cache the CACTI results after the first time configurations. How should I make use of the patch you provided?

Thanks,

Ermao

On Thursday, September 19, 2013 8:04:02 AM UTC-4, Wim Heirman wrote:

[Wim Heirman]

Ermao,

You should be able to just apply the patch (available at [1]) to an unmodified version of McPAT 0.8 which will enable the CACTI caching. After that you can re-apply your own modifications. Doing it the other way round might be possible as well but you may have to fix some merge conflicts in case you updated the same parts of the code.

Regards,

Wim

[1] http://snipersim.org/w/McPAT

[Huan Zhang]

Hi Wim,

I have created a patch for McPAT 1.0 supporting results caching. Some minor changes to mcpat.py are also necessary, so there is a separated patch for mcpat.py. They work well on my machine.

In my tests, McPAT 1.0 gives a little bit higher power numbers than McPAT 0.8.

I have attached my patches for SniperSim 5.2, and hopefully you can include McPAT 1.0 support in next release.

Best,

Huan

[Wim Heirman]

Huan,

Thanks! We'll make sure to include this in the next version, which should hopefully be out in a week or so.

Regards,

Wim

[Umar Dastgir]

Hello! So I am working with McPat to extract the power and area stats. I actually got the error:
"Some value(s) of number_of_cores/number_of_L2s/number_of_L3s/number_of_NoCs is/are not correct!."
I tried to play around with the file and realized that for a single core, I can't have NoC so I deleted the NoC part from the file and got the following error:
"ERROR: no valid data array organizations found."
I have tried playing around with the cache configurations, but the error still persists. I am clueless as how to solve it. I have attached the XML file. Any help in this matter will be highly appreciated.