- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
I am trying to understand the "cache-config" option in the Advisor roofline analysis. I have a two socket system with Platinum 8280 CPU 28 cores/socket. Roofline analysis of the "roofline_demo_samples" case without option "cache-config" (default mode) delivers following values for cache rooflines:
- L1 bandwidth: 511.39 GB/s
- L2 bandwidth: 240.97 GB/s
- L3 bandwidth: 27.15 GB/s
Having the results with default cache configuration I repeat the same analysis explicitly specifying cache hierarchy as follows:
--cache-config=28:8w:32k:64l/28:16w:1024k:64l/1:11w:39424k:64l
As a result I obtain following values:
- L1 bandwidth: 441.4 GB/s
- L2 bandwidth: 186.39 GB/s
- L3 bandwidth: 27.74 GB/s
Since differences in L1 and L2 bandwidth are substantial, I would like to understand the cause of these discrepancies.
Kind regards,
Igor
- Tags:
- Roofline
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Sorry for confusion. I meant that it happens step by step during Trip Counts with Flops enabled collection and shouldn't be overlapped.
You can see it in Collection log of Trip Counts collection. E.g.
Peak bandwidth measurement started...
Peak bandwidth measurement finished. - roofline values are measured
Analysis started... - step for cache simulation
Collection has been started.
Anyway if you still see the dependency at several run, I will raise a ticket so someone can check it.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hello Igor!
Cache config option doesn't affect roof values. Roof values are measured on the system each time Trip Count collection with flops is running. Values are result of running small programs called benchmarks. They measure peak values for your current hardware. As a result values could be different from run to run. It is possible that your system was busy, which could be a reason of these discrepancies.
Btw, you can put custom Roof values in Roofline setting menu (at right top corner at Roofline view). Cache-config option is used for cache-simulation to get data transfers values between different memory levels for your loops/functions.
Hope this helps,
Ekaterina
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Ekaterina,
thank you for clarification of "cache-config" option.
Regression in L1 bandwidth when using "cache-config" option is rather stable and remains after repeating calculation several times. I expect rather, that the discrepancy can be simply an advisor overhead when enabling cache simulation.
Regards,
Igor
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Technically It should not happen, because benchmarks run before Trip count collection with cache simulation on/off. You can see it in collection log.
But if you confirm this problem exists after several runs with cache simulation on/off, I can raise a ticket to investigate why it happens.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Ekaterina,
do you mean, that rooflines are calculated in the "survey" pass? Since rooflines depend on the properties of CPU, what is the purpose of the workload in the first collect "survey" step?
Moreover there is an option "target-pid" which allows to attach the advisor instance to the running process. Does it change the measurement of rooflines?
Regards,
Igor
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Sorry for confusion. I meant that it happens step by step during Trip Counts with Flops enabled collection and shouldn't be overlapped.
You can see it in Collection log of Trip Counts collection. E.g.
Peak bandwidth measurement started...
Peak bandwidth measurement finished. - roofline values are measured
Analysis started... - step for cache simulation
Collection has been started.
Anyway if you still see the dependency at several run, I will raise a ticket so someone can check it.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Igor,
Did Ekaterina's recommendation help you to solve the issue?
If yes, I would mark this forum topic as solved and close the related Online Support Center ticket.
Thanks, Mariya
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Mariya,
yes, Katya answered my question and you can close the topic.
Best,
Igor
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page