- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
All hardware PMU event-based sampling and analysis types are accurate, since they are hardware based:-)
My opinions are:
1. As first step, you can use lightweight-hotspots analysis to know CPU_CLK_UNHALTED.CORE (aka clockticks) & INST_RETIRED.ANY to evaluate performance and workload.
2. Secondary, you can use "General Exploration" analysis to detect PMU events, most of events (branch mispredict, L2 cache miss, resource stall, etc) impact on performance result. There are 10+ events can be monitored ata run.
Regards, Peter
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
Let me clarifyfrom backgroud.
If the user set many events, buthardware can only monitor limited eventsat a session- thus, old VTune will reorganize these events into several groups then run several sessions.
In new product, the tool uses "multiplexing events" and collect data in different time slice - for example,"t1, t5, t9,..." will serve onsome events, "t2, t6, t10,..." will serve on other events, ...
Yes. Some samples will be missed (not captured), butresults will reflects program's overallevents situation.
- "allow multiple runs" means that tool will not use "multiplexing events", BUT all you selelected events (attached file) - they are allowed in one session, and there is no sample missed.
If you wants to run one event only at a session, you can create several projects for several events.
Hope that I have answered your questions.
Thanks, Peter
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page