.Net Foundation On .NET, 26 Sep 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Walking Down the Memory Lane with Bing – Part 3

Description

In this episode, Maoni (CLR) and Mukul (Bing) talk about the new GC profiling APIs in .NET Core 3.0 and how to take advantage of them.

Follow Maoni and Mukul on Twitter:
https://twitter.com/maoni0
https://twitter.com/mjsabby

A

Hi, we're back with part 3 of walking down memory, lane,.

A

Talking about diagnostics. check it out.

A

Hi. welcome to the next episode of the on.net show., I'm [inaudible] ,, the owner of.net core and garbage collector.

A

hi., I'm michael from the bing performance team.

B

So michael, last time we ended at talking about frozen objects,.

A

And I mentioned that you and I also did a bunch of diagnostics, work.

B

Right. in 30 time, frame.

A

So I think that would be really interesting to talk about,, especially for people who are really interested in how to figure things out..

A

So I think this all started actually quite a way.

A

Back when you either commented or started an issue on github,.

A

Where you requested for a lightweight gc profiling, api.

B

Yeah,, maybe two or three years: ago. yeah,, two years ago.

A

So tell me a bit about why you wanted it and all that.

A

So, as you can imagine,.

B

Having large memory, workload, you'll have managed memory, problems,, leaks, and all things., so one of the things we wanted to figure out is: how can we use the gc profiling apis without having to turn off concurrent gc,.

B

Which the heavyweight gc profiling apis do?.

A

Right., so that's really a disadvantage,, because we definitely don't want you to turn off background gc..

A

So when you said that, I was like, "I wholeheartedly agree,.

A

I want it too", and then years have gone by and nothing was happening.

A

I e-mailed you again, one day.

B

Yeah. earlier this year,.

A

You sent me e-mail., I think you had three things on it.. This was one of them.

A

yeah.. I just felt bad.. I was like, "gosh".

A

two years have gone by and nothing's happening for this,, and this is a simple thing to do., so I was like, "I'm just going to do it today." yeah., so I planted that day, and I checked it. In.

A

That's where we started our trend of adding new gc profiling, apis.

B

A

Profiling, apis is really the only way that you can get a real-time diagnostics in.

A

it is that be in the event is not real time.

A

You cannot do something like when gc happens at this moment,. I want to do something.

B

Do something else.

A

Yeah., so the lightweight api gives you real-time juicy start and stop, and generational ranges.

A

So how are you using that, in your environment?.

B

Yeah. so the lightweight profiling apis, which we've called,.

B

I mean, we're not the best naming experts, basic gc.

A

It's not our forte.

B

Naming is not our forte., it's not our forte.. It's called basic gc., it's flag, you turn on..

B

What we use it for is getting ranges which are.

B

Really helpful for finding out things like, "okay.

B

Where is our heap?" from beginning to end for all the different segments,.

B

For all the different gc segments,, where is the data?.

B

Because you can find out things to grow, and it also tells you where the object's ended.

B

If you wanted to walk it for example,, you could do things like that., so we use it for just getting the idea of ranges,.

B

Which we couldn't do with the previous api, without disabling gcs.

A

Exactly. yeah., so that was really helpful..

A

Then from there, you start adding a bunch of new profiling api support..

A

I think the next one that you added was medium weight, gc, profiling, api.

B

That's right., as we said, we're not naming experts.

A

We added the medium weigh api.

B

The medium weight api is mutually exclusive from the basic.

B

So if you wanted both.

A

You have to enable both.

B

Right., you have to enable basic and this medium weight.

B

The medium weight api basically gives you actual information about object, movements.

B

When an object moves in gcs, where it does move,.

B

You get that information.

A

So you only get this during compacting gcs.

B

You only get this during compacting gcs and then you actually will be able to do,. If you wanted to write a profiler tool that did memory leak detection,. This is now possible. Again, without disabling concurrent gcs.

A

Exactly,, which is very nice., which is very nice.

B

Yeah. I think,.

A

We talked about frozen objects last time and one of the things that you had to do obviously was to,.

A

Given the object you wanted to work as a references.

A

So you also added an api for doing just that.

B

Right., so when I started on the frozen objects stuff that we talked about last time,, I was wanting to enumerate, object references and I said,.

B

"Hey, profiling, api makes exact sense.

B

it'll be useful because you'll be able to pause, gc and just enumerate the objects from there.", but it does exist, an api which was cumbersome for me to use,.

B

Which, basically gives you all objects that survived,.

B

Removed, and all their references., so it was awkward to use that.

A

I think we're just packed up everything in a buffer., so if you want to know a particular object's references,.

A

You have to go through that buffer and pick out the ones that you care, about.

B

Buffering was an impediment, because you were buffering for every object,, all its references.

B

So the new api that I added essentially allows you to enumerate object references just the way the gc does and have you get called back for each reference without having to buffer up anything,, and you can do it for the object that you care about., so an object of interest,. You can just enumerate it's references.

A

Just that one. yeah. also,, as we talked about last time,,.

A

That all your frozen objects code was a managed code.

A

So it turn out that you actually didn't use this api..

A

Can you talk a bit about that,? Why you didn't end up using it?.

B

Yeah., so I didn't end up using it because, with the idea of you wanted to make frozen objects already experimental feature that we want to get traction, having to require the profiling apis.

B

Be a thing in your tool, was an impediment., so we basically said, it'd, be easier to just have a full one match. Code.

A

Add-On, and it's an explicit addition.

A

now,, you have another dl that you have to load.

B

Plus, you may have had a profiler running already, so it was just cumbersome.

A

So that's interesting because, even though you didn't end up using that,, I thought it was so useful to keep it in the product., because, if you're already a profiling company,.

A

Your company's product is the profiler,.

A

We have quite a few of those, and you're already using the gc profiling api anyway.

A

So now you have a much faster away and much more convenient way.

A

To you enumerate object, references.

B

The customers of said tools would.

B

Also benefit because now they wouldn't be told you have to run..

B

If you want to use a profiler with these features,, you have to disable convergency.. Now you can use it anywhere. You want.

B

you can enable and disable it at will,, as your application is running.

A

Exactly. yeah., so what else did you add??

A

I think you also added a large objects. Allocation, api.

B

Yeah. actually, talking about frozen objects,.

B

I also added an api that tells you if an object is frozen or not.

A

Because, now that we're expecting people to start using.

B

Frozen objects and there will be lot heaps that may have it,.

B

You want to be able to diagnose, and profile. Authors would want to know.. If this object is frozen,, I would never want to maybe go through its object references because we know this is going to be.

A

You want it treated differently from the read-write objects.

B

So you can actually now find out if a object is frozen and then you can essentially find the frozen segments as well.

B

Then the loh stuff,, because now in 30, the loh the threshold is configurable.

A

It is configurable,, the threshold is configurable.

B

You can increase the size., so I was thinking. Wouldn't it be cool if we could just get object, allocation events for large objects.

A

Right,, because right now,, we do have profiling apis. That allows you to just get it for every object. Allocation.

A

You can imagine the amount overhead that would ask, because we have.

A

These really fast allocation help for us to handle the non-large and the small object, heap objects.

A

Now, by enabling profiling, the allocation apis, they go through slow, helpers.

A

But large object allocations already needed to go through a call,.

A

So you can just have a hook up right, there.

B

So that makes much more sense to have a specific option for those.

B

If you wanted to just do care about,, if your team trying to improve your loh allocations,.

B

You could just enable this., you only get the loh allocations.. If you configured it to a higher number,.

B

Let's say you decide. Even 85 kilobytes, which is the current size by default, is too high,. You can set it to 200 and use that.

A

Too low. yeah. right. yeah.. If the number is too low,.

B

You can enable it to too high., so you can imagine a way of doing.

A

This prof diagnostics is, you could set it very high.

B

If you allocate it a few like 250 megabyte objects,,.

A

You can just set the threshold there.

B

So you only get the profiling api to tell you about them.

A

So you can make a determination, like, "do. I want to keep them or not,". Then you can keep lowering these thresholds..

A

So that's a way to diagnose large objects, more conveniently.

B

Controlling the amount of impact that you'd have to your program.

B

another api in concert with this is you can actually find out what.

B

The loh configuration threshold.

A

Is from the profiling apis,, which is useful.

A

yeah., so these are all things that we really wanted. Other customers to.

A

Also give us feedback on because a lot of people had told me that well,, you have gc profiling apis, but we couldn't use them on production environment for example.

B

Which the lightweight you could totally use that on production.

A

I don't think you can really detect any prof overhead.

B

Correct., I think the basic gc is something that, if you could turn on, get that information essentially for free.

A

Right., so I think for the existing profiler authors,.

A

You have an additional tool available for you.

A

For people who use to avoid using.

A

Profiling apis for gc on their production,.

A

Now they do have a way to do that if they want to.

B

In bing,, we run with the basic profiling turned on essentially everywhere.

A

Right., you also have plans to use,.

A

For example, the medium weight and all that.

B

The medium weight is something that you can, for example, turn on as part of the leak detection, analysis.

A

You don't have to restart a program., you don't have to do it all. The time.

B

Correct. yeah., you can use this for that.

A

Yeah., so I think this would be really useful..

A

It's a very useful addition to our products.

A

yeah., as usual, you're welcome to give us feedback,.

A

Opinions, and requests and what-not on twitter.

A

Our twitter links will be down below the video and.

A

I think that wraps up our walking down the memory lane,, the three parts.

A

it does. bye. bye.