How to Obtain Raw Data for Simulation Purposes

How to Obtain Raw Data for Simulation Purposes

Imagen de Joshua Ferguson
Joshua Ferguson
Lun, 28/03/2011 - 19:26

Hello all, I'm a newbie to VTune, attempting to use it in an academic project. In brief, I'd like to look at a trace of different workloads on an Apache webserver, and see if I can suggest some machine learning technique to predict memory access, clock halts, (or any other useful metric of how the workload affects the CPU). My question to you guys is: how can I obtain access to a trace for this purpose? I know a lot of effort was put into the analyzer, but it isn't sufficient for my purpose (matlab programming). Can I obtain a raw trace of memory accesses over time or clock halts over time using VTune? Also, does anyone have a suggestion of which CPU metric would be useful to predict? Thanks ahead of time.

RSS Superior
publicaciones de 5 / 0 nuevos
Último envío
Para obtener más información sobre las optimizaciones del compilador, consulte el aviso sobre la optimización.
Imagen de Peter Wang (Intel)
Peter Wang (Intel)
Mar, 29/03/2011 - 22:47

Hi,

If you try to use VTune Amplifier XEwith server application, I suggest to create a new project -> set for "Profile System"as Target. Then you can select "Memory Access"as Analysis Type, which includes CPU_CLK_UNHALTED, INST_RETIREDand many MEM_LOAD_RETIRED events (L1 miss, L2 miss, DTLB miss).

After you start above data collection, you can stop this collector manually on GUI, then get report.

The report can be displayed (grouped by process/thread/module, or function) - so you can observe where (code)and when (timeline) data cache missed. However the tool does NOTprovide any trace for memory access (address).

Yes, you can observe CPU state (Running / Wait) on Function and/or Thread, also event count of CPU_CLK_UNHALTED.CORE indicated code/threadis busy or not.

Hopefully I have answered your questions.

Thank youof having interest of using Intel VTune Amplifier XE.

Regards, Peter

Superior
Imagen de Rob Wolters (Intel)
Rob Wolters (Intel)
Mié, 30/03/2011 - 09:36

Joshua,

You may want to take a look
at http://whatif.intel.com particularly
the Performance Tuning area of this site.
Of interest may be the Intel Performance Tuning Utility and / or the Intel
Platform Modeling with Machine Learning.

If you link to the Intel
Performance Tuning Utility, the page will provide an overview with links to additional
resources such as a Discussion Forum, download, release notes, etc.

Thanks
Rob
Intel Support

Superior
Imagen de Joshua Ferguson
Joshua Ferguson
Mié, 30/03/2011 - 14:57

Peter, Thanks for the response. I'll definitely make use of the metrics you suggest (CPU_CLK_UNHALTED.CORE, cache misses, etc.). I'm rusty on CPU architecture, so you've given me a good jumping point to investigate the utility of these metrics. With regards to the trace, what I'm looking for is something like a file dump that has event information organized chronologically, not necessarily the memory addresses. I'm thinking something like: t (in nanoseconds) : # of L1 misses t=1 : 1 t=2 : 3 . . . For a few experiments, I'm going to need to act on this data in matlab, thus this need. I noticed that VTune has a file that it saves as the report, but when I attempt to open it in notepad/notepad++ it seems compressed/proprietary.

Superior
Imagen de Joshua Ferguson
Joshua Ferguson
Mié, 30/03/2011 - 15:01

Rob, Wow, I'm glad you pointed me to this, it seems like they've already tackled exactly what I was planning to. I'll have to see exactly what kind of equations/techniques they use for prediction. Luckily, machine learning is a fast moving area right now, so I might still have an experiment or two to do. Thanks :D

Superior

Inicie sesión para dejar un comentario.