Could you please recommend a C++ performance profiler for Windows 64-bit that can attach to a running process and collect stack traces? The profiled program is massively-multithreaded: thousands of threads that are mostly waiting for network activity. Some tools do not work with my program because my program throws exceptions.
-
Which profiler has problems with exceptions? A profiler should not be interested in exceptions at all. – Thomas Weller Jul 2 '20 at 19:12
-
Why are 1000 threads waiting? Isn't that quite useless? And - if they are waiting, what should the profiler tell you other than they are waiting? Do 1000 threads make sense on a 16 core machine? – Thomas Weller Jul 2 '20 at 19:16
-
@ThomasWeller, it's softwareverify.com/performance-profiling.php that can't profile programs that throw exceptions. As to multithreading, I tried VTune, MSVS 2019 built-in profiler, VerySleepy and many more - they all can't handle 11K threads. Yes, these threads are waiting, but they are waiting in a library out of my control. The library however has an interface that doesn't require 1 thread per connection, but it is very complicated. The profiler should give me insights about the threads that are not waiting. – Serge Rogatch Jul 2 '20 at 20:39
Add a comment
|
Since you do not have many requirements, I suggest starting with Windows Performance Toolkit (WPT). WPT is not available as a standalone tool. It is part of the Windows Assessment and Deployment Kit (WADK).
WPT is not a single tool. It's made of two parts: Windows Performance Recorder (WPR) and Windows Performance Analyzer (WPA).
By default, the recorder (WPR) will capture the activity of the whole system (that's ~2800 threads here currently) without any issues.
And here's an example of what a stack analysis looks like for the Windows Compatibility Telemetry process, done with WPA after recording:
The blue highlighting in the CPU graph corresponds to the selection of MSXML6 calling SchemaCache::validate(). You can see how many times it was called and how many milliseconds were spent there.
Windows Performance Toolkit
- is made by Microsoft and works on Windows
- can analyze C++ or any language that compiles to x86 code. It's not so good for stuff that compiles to intermediate languages (IL) such as C# or Java.
- is gratis (free of cost, but not open source)
- is capable of analyzing 2000+ threads
- does not care about exceptions in your program (when your program crashes, you'll still have the data up to that point)
answered Jul 2 '20 at 19:44
-
Unfortunately, Windows Performance Recorder produces a 320GB file, and then Windows Performance Analyzer fails to load it, complaining about running out of storage (1.5TB free space before running WPA). – Serge Rogatch Jul 3 '20 at 6:07
-
-
-
@SergeRogatch: I think I will delete this answer. Would it be ok for you to integrate all the comments into your question? I think you'll get better responses when we know more requirements / have a better understanding what you already tried. – Thomas Weller Jul 3 '20 at 8:48
-
I think your answer may be fine for someone else, and I've already upvoted it :) . – Serge Rogatch Jul 3 '20 at 9:10
