Performance Profiler Showdown for .NET

I recently had to do some performance profiler evaluation for .NET applications and I figured I'd share my results. Note that it's as scientific as I could make a subjective review (e.g., "friendly UI" might mean something different to you than to me), but maybe it'll help you out. Also, I'm not a "profiler expert" and, while I've used profilers before and understand generally what I'm looking at, this isn't my primary job function.

The five profilers I tried out:

I put an explanation of what each feature "means" in tooltip form, so put your cursor over it if you don't understand what I'm talking about. An "X" in the box means it has the feature.

Testing was done on a dual-2.8GHz processor machine running Windows Server 2008 R2 64-bit and 4GB RAM.

  VSTS 2008 ANTS Perf 5.2 VTune 9.1 dotTrace 3.1 AQtime 6
User Interface
Visual Studio integration X     X X
Standalone application   X X X X
Friendly/easy to use X X   X  
Robust reporting   X ?   X
Measurement Style
Sampling X X X X X
Instrumentation X X X X X
Measurements Recorded
CPU time   X X X X
Wall-clock time X X X X X
Additional perf counters   X     X
Notes
  This requires Visual Studio, which means you have to have VS installed on the machine running the app you're profiling. That said, this was the easiest to get results from and the easiest to interpret. In general this appeared to be the best balance between "robust" and "usable" but I couldn't actually see the report that came out because it locked up the UI thread on the machine and ate 3GB of memory. I've asked about this in the forums. Turns out this is fixed in the next version, 6, currently in EAP. I couldn't actually get a profile to run using VTune since it complained of being "unable to determine the processor architecture." As such, I don't know how well the reporting works. When I ran dotTrace 3.1 on a multi-proc system, I got several timings that came out with negative numbers (-1,000,289 msec?). You can fix this by setting the proc affinity for the thing you're profiling. I tried a nightly build of dotTrace 4.0 and that's fixed. dotTrace 4.0 will also let you profile a remote application - something the others don't support. AQtime has a lot of power behind it but lacks the usability of some of the other profilers. It appears that if you take the time to really tweak around on your profile project settings, you can get very specific data from an analysis run, but doing that tweaking isn't a small feat. I spent a good hour figuring out how to profile an ASP.NET application in the VS dev server and setting it up. Also, while it very well may be due to my lack of experience with the tool, AQtime had the most noticeable impact on the application's runtime performance. It took several minutes for the first page of my app to load in the browser.

For now, it looks like the VSTS profiler is my best bet. If I could figure out the UI problem with ANTS, or if dotTrace 4.0 was out, I'd say those options tie for my second choice. The VTune profiler seems to be the most... technical... but it also desperately needs a UI refresh and seems geared toward profiling unmanaged code, where managed code is a "nice to have" rather than a first-class feature.

UPDATE 1/21/2010: I added AQtime to the list of profilers I tried out. Also, I removed the "VS integration" checkmark from ANTS because, while it adds a menu entry to VS, all it does is start up the external application. I'm not counting that. Finally, I found out my ANTS problem is fixed in the next version, 6, currently in EAP. Since it's not released, I still have to go with the VSTS profiler, but once it's out, I'd vote ANTS.

posted on Tuesday, January 12, 2010 1:32 PM | Filed Under [ GeekSpeak Web Development .NET ]

Comments

Gravatar # re: Performance Profiler Showdown for .NET
by Salva at 1/14/2010 3:53 AM
You should evaluate AQtime profiler.
Gravatar # re: Performance Profiler Showdown for .NET
by Travis Illig at 1/14/2010 8:17 AM
Oh, good call. If I get a chance, I'll try out and add AQtime here. Thanks!
Comments have been closed on this topic.