- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Community,
I am using Intel Profiler to profile an application on Intel core i7 to perform microarhcitecture exploration. I found that form hardware mode sampling, the minimum sampling interval can go till 0.01ms.
While profiling with 0.01ms, the vtune generates a warning that the sampling interval is too low and can cause inaccurate results. So I tried with 0.1ms and 1ms. The warning doesn't comes up this time but the profiling results of micro-architecture exploration i.e., the front-end bound, bad speculation, back-end bounds (as a percentages) are different in both the cases.
Why such difference? Which one is more accurate? I need a justification as to which value to be used which I am unable to find on the VTune documentation. The Intel just mentions that 1ms is recommended value.
Regards.
Shlok
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
Thanks for posting in Intel Communities.
We have contacted the development team regarding such behaviors. We will get back to you soon with an update once we get confirmation from them.
Could you please share the following details which will help us to debug further-
1. Sample reproducer code.
2. VTune version.
3. OS, Kernel and Processor details.
Regards,
Diya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
We have not heard back from you.
Could you please share the above details so that we can try to investigate the issue from our end?
Regards,
Diya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I am sorry Diya, I thought no one from the team is looking at your query. Thanks for considering the question.
OS details: I am using Ubuntu 20.04, Kernel 5.15.0-86-generic
The processor details are as follows
___________________________________
Reproduction: In VTune GUI for microarchitecture analysis, the executable application and its parameter can be found at the drive link https://drive.google.com/drive/folders/13LJmSoURoY6Gr090HdWjEWDfH5sP_dTu?usp=sharing
Here, the file named 'cadical' is an application that receives the second "cnf" file as an application parameter or input argument. The 'cadical' application takes about 5 minutes to execute completely on the given parameter and then VTune shows the profiling results.
_____________
VTUNE Version:
Intel(R) VTune(TM) Profiler 2023.0.0 (build 624757) Command Line Tool
Copyright (C) 2009 Intel Corporation. All rights reserved.
__________________________________
I have tried for 0.1ms and 1ms and uarch exploration results (the various bounds like front-end, backend, core bound) are different in both the cases.
For 0.01ms there is an warning shown that sampling interval is too low and can cause instability.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
Thank you for providing the details.
We are working on it internally and will get back to you with an update soon.
Regards,
Diya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
Please find the below observations:
1. The amount of performance monitoring overhead from the sampling driver (perf and/or sep and/or vtsspp) increases as the sampling interval decreases.
To know more about it you can follow this document: https://www.intel.com/content/www/us/en/docs/vtune-profiler/user-guide/2024-0/minimizing-collection-overhead.html
2. When the sampling interval is less than 1 ms, you should see about 5-9% overhead.
When the sampling interval is >= 1 ms, you should see less than 5% overhead.
It comes down to a tradeoff of accuracy (smaller sampling interval) vs. overhead (i.e., reduced relevance of data when you are measuring VTune's sampling driver instead of target workload).
3. Please try to disable call stack collection when specifying a sampling interval less than 1 ms. Specifying 0.5ms sampling interval should be OK for hotspot analysis which collects fewer performance data metrics than micro-architecture exploration but for micro-architecture analysis try not to specify less than 1 ms.
If this resolves your issue, kindly mark it as a solution as this would help others with a similar issue.
Regards,
Diya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
We have not heard back from you .
Did it clarify your doubts?
If you have any queries please get back to us.
Regards,
Diya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi ,
Good day to you.
We have not heard back from you.
We assume that your issue is resolved.
This thread will no longer be monitored by Intel. If you need further assistance, please post a new question.
Regards,
Diya
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page