Maoni Stephens

A Thread - a guideline doc for memory performance analysis for #dotnet #dotnetcore For the past few weekends I've been working on a (long) document that aims to help you with finding efficient approaches to mem perf problems. github.com/Maoni0/mem-doc…

hi folks! it’s been quite a while since I had time to write a blog post. with the upcoming .NET 10 release, I’m finally sharing more details on DATAS and tuning tips you might find useful. check it out here: maoni0.medium.com/preparing-for-…

@djcp86 @davidfowl @dank_army @dariogriffo @konradkokosa if you hit problems, please feel free to open an issue in the runtime repo on GH (I don't look at X much these days - I was informed of your post by my kind coworkers :-)). I'm still improving it in 10.

@davidfowl @maoni0 @dank_army @dariogriffo This stuff is fascinating, feel like a child, also waiting for @konradkokosa to release his book about .net memory.

Images of the same app running in #dotnet 8 and #dotnet 9 same load in my local machine 3 rps The only thing that changes is the target framework. Oh my f4king god. @davidfowl and the rest of the team are smashing like crazy. We will upgrade this week.

@djcp86 @davidfowl @dank_army @dariogriffo I was going to talk about this at .NET DevDays in Warsaw but I broke my ankle so that didn't happen😅

@maoni0 @davidfowl @dank_army @dariogriffo Btw, is there any recent talk (NETconf?) talking about this?

@djcp86 @davidfowl @dank_army @dariogriffo I mentioned an example in one of my blog posts maoni0.medium.com/is-the-regress….

@djcp86 @davidfowl @dank_army @dariogriffo sure, here's our official doc page about it: #name" target="_blank" rel="nofollow noopener">learn.microsoft.com/en-us/dotnet/c…

@davidfowl @dank_army @dariogriffo I only ported functional changes back to 8, not tuning changes (too much to port). for anyone who wants to try it out, please either try out 9 or get the clrgcexp.dll from 9 and use it with your 8 SDK by replacing the GC with the GCName.

@dank_army @dariogriffo It’s not just on by default in .NET 9, LOTS of work was done in 9 to improve the implementation. I don’t know if that was backported to 8 cc @maoni0

@dariogriffo @davidfowl is the app using Server GC (if this is an asp.net app it uses Server GC by default)?

@StudentDanLuis sure thing!

@maoni0 Thank you so much! That’s what I have been doing I just wanted to make sure I’m not missing any filters, I really appreciate you taking the time to answer and sending samples!

recently we worked with a team that specified only the provider name in dotnet-trace without keywords/level. dotnet-trace will include ALL keywords on VERBOSE level. this can easily degrade perf. DotNETRuntime provider will also induce GCs (due to the GCHeapCollect keyword).

@StudentDanLuis the only thing you can specify is that EventCounterIntervalSec filter value

@StudentDanLuis the name of the runtime counter provider is System.Runtime. this is mentioned here #systemruntime-counters" target="_blank" rel="nofollow noopener">learn.microsoft.com/en-us/dotnet/c… albeit not obvious that's the name of the provider.

@StudentDanLuis logman, eg - logman query providers Microsoft-Windows-DotNETRuntime

@maoni0 Is there a way to know which keywords are available per provider?

so really you should specify the keyword/level🙂 or use one of the built-in profiles

I've requested the diagnostics folks to the change the default: github.com/dotnet/diagnos…. note that this means no keywords enabled on informational level.

@fiigii will definitely keep that in mind. I'll try to find time to do it🙂

@maoni0 We can start with a simple case, I believe it would still be helpful for many novices like me

if you use PerfView, have you noticed the GCTriggeredStacks cmd arg? to see callstacks for induced GCs, use this lightweight commandline - PerfView /GCCollectOnly /GCTriggeredStacks /nogui collect (this is less verbose than what mem-doc currently shows)

profiling tip - if you upgrade to a newer version of a product and observe perf regression, make sure to grab some perf traces before you revert (if you revert all usage and can't easily get traces later). without comparative traces the investigation will be much harder

@fiigii a major problem is interesting data to look at usually isn't shareable publicly. to make it shareable it would require a lot of time (for someone inept at video editing like me) to post process😅

@maoni0 It would be very useful if you could make a YouTube video to show your memory profiling workflow 😆

@chnasarre @konradkokosa @KooKiz also probably better to ask questions like this in email instead of on twitter. in general I look at twitter very infrequently these days (I just happened to look because I made a tweet myself recently)🙂

@chnasarre @konradkokosa @KooKiz (by expose I just meant expose it directly, because you can calculate it yourself)

Dear @maoni0, we wonder with @KooKiz and @chnasarre why new_allocation is not exposed via events, fe. in PerHeapHistory?