Bug Idle (!) performance much worse in 2020.2+ versus 2020.1 and earlier

I've been seeing a very significant performance regression in a DOTS project for a while (several months now, but I thought it was my fault).

Performance is much worse in 2020.2 and later than in earlier versions.

This affects both Edit Mode and Playmode (in the editor), Windows builds of the project seem less affected but I haven't had the time to benchmark.

This issue occurs with both URP 9 and 10 (URP 10 seems worse but may be a nocebo effect).

There seems to be something running, even if the scene is literally empty, that eats between 5 and 25 ms (sic) of CPU time. This time doesn't show up in the profiler other than diffuse time spent in EditorLoop. In fact, the profiler shows about the same CPU time in either unity version, but then it is actually 5x or more as much in the Stats gizmo (first screenshot), and this can be felt (under minimal load; with more cameras or scene views open, this drops my frame rate to below 40 and sometimes into the 20s).

On my system spec (i7-8086k, RTX 2080 Super), I would normally see 500-1000 fps.

This has been going on since 2020.2 beta, and also persists in 2020.2.1f1 and I just confirmed it for 2021.1.0b1 (to a lesser extent - cpu time in 2021.1.0b1 is only ~4x as much as in 2020.1, while 2020.2 eats ~15x as much CPU time per frame - it was worse in 2021.1.0a9 or earlier).

Unity 2020.2.1f1 - 40 fps in edit mode (empty scene "New Scene"), sometimes this hovers around 80 fps; the actual content of the scene doesn't seem to matter much:


Play mode is slightly better at 11+ ms cpu time... but about 1/10th the performance I would get in 2019.4:


And now the shocker ... both in play mode and edit mode, CPU usage is under 1 ms in the profiler? But Editorloop seems to account for many of the missing milliseconds.

The performance I see in the profiler is the performance I could expect and actually see with earlier versions of Unity.


In Unity 2020.1.17f1 - 700+ fps (fps in empty scene "New Scene", mind you this is with various systems but still only about 50% the performance I would expect from my System)



I've tried many permutations of packages. I've stripped out things like Odin Inspector and Rainbow Folders to ensure it's not an Editor extension causing this.

This performance issue persists whether I have G-Sync on or off, and whether I use Unity on one screen or two. It does seem to get worse with multiple scene views, so the "4 Split" layout gets the worst performance.

Some profiling at some stages showed me that the following things take "long" (despite the profiler insisting there's no time spent):
- Editor Loop
-- Rendering Scene Views
-- Rendering the Entity Debugger
-- Rendering Inspectors - deselecting an object in Scene View will help performance a good bit, but I took these screenshots with nothing selected.

The performance stays terrible even if I close the entity debugger.

Here are my packages. I've tried cleaning out the entire project and installing each at their "current baseline", to no avail. The performance issue persists, especially also with older versions of Burst (1.3) and Input System, etc.

Code (JavaScript):

1. {
2.   "dependencies": {
3.     "com.unity.burst": "1.4.3",
4.     "com.unity.collab-proxy": "1.3.9",
5.     "com.unity.dots.editor": "0.11.0-preview.3",
6.     "com.unity.entities": "0.16.0-preview.21",
7.     "com.unity.ide.rider": "2.0.7",
8.     "com.unity.ide.visualstudio": "2.0.5",
9.     "com.unity.ide.vscode": "1.2.3",
10.     "com.unity.inputsystem": "1.1.0-preview.2",
11.     "com.unity.physics": "0.5.1-preview.2",
12.     "com.unity.render-pipelines.universal": "10.2.2",
13.     "com.unity.rendering.hybrid": "0.10.0-preview.21",
14.     "com.unity.searcher": "4.3.1",
15.     "com.unity.test-framework": "1.1.20",
16.     "com.unity.textmeshpro": "3.0.3",
17.     "com.unity.timeline": "1.3.6",
18.     "com.unity.ugui": "1.0.0",
19.     "com.unity.visualeffectgraph": "10.2.2",
20.     "com.unity.modules.ai": "1.0.0",
21.     "com.unity.modules.androidjni": "1.0.0",
22.     "com.unity.modules.animation": "1.0.0",
23.     "com.unity.modules.assetbundle": "1.0.0",
24.     "com.unity.modules.audio": "1.0.0",
25.     "com.unity.modules.cloth": "1.0.0",
26.     "com.unity.modules.director": "1.0.0",
27.     "com.unity.modules.imageconversion": "1.0.0",
28.     "com.unity.modules.imgui": "1.0.0",
29.     "com.unity.modules.jsonserialize": "1.0.0",
30.     "com.unity.modules.particlesystem": "1.0.0",
31.     "com.unity.modules.physics": "1.0.0",
32.     "com.unity.modules.physics2d": "1.0.0",
33.     "com.unity.modules.screencapture": "1.0.0",
34.     "com.unity.modules.terrain": "1.0.0",
35.     "com.unity.modules.terrainphysics": "1.0.0",
36.     "com.unity.modules.tilemap": "1.0.0",
37.     "com.unity.modules.ui": "1.0.0",
38.     "com.unity.modules.uielements": "1.0.0",
39.     "com.unity.modules.umbra": "1.0.0",
40.     "com.unity.modules.unityanalytics": "1.0.0",
41.     "com.unity.modules.unitywebrequest": "1.0.0",
42.     "com.unity.modules.unitywebrequestassetbundle": "1.0.0",
43.     "com.unity.modules.unitywebrequestaudio": "1.0.0",
44.     "com.unity.modules.unitywebrequesttexture": "1.0.0",
45.     "com.unity.modules.unitywebrequestwww": "1.0.0",
46.     "com.unity.modules.vehicles": "1.0.0",
47.     "com.unity.modules.video": "1.0.0",
48.     "com.unity.modules.vr": "1.0.0",
49.     "com.unity.modules.wind": "1.0.0",
50.     "com.unity.modules.xr": "1.0.0"
51.   }
52. }
53.  

 

This issue may be a duplicate of this:

https://forum.unity.com/threads/2020-2-0b1-performance-reduction.965363/#post-6656287

 

Because the other thread is now in the archived beta section, I'll repost follow-up here.

 

Okay, understood.

The actualy performance difference is very significant though. (talking 10x to 20x, which results in 40-ish fps vs. 400-800-ish; of course this is hyperbolic but it's still almost an order of magnitude if you count time spent instead of fps).

I do think this is something in DOTS causing this, for a while I believed it was my systems dependencies deadlocking each other across frame boundaries, but that seems no longer to be (or never was) the case and it's only the editor taking this much time.

The more Scene Views and the more Inspectors or Focused Inspectors I have, the worse it seems.

I'll spend some time over christmas profiling as you requested and will post the captures here.

 

The profiler can also be very misleading because things can show up several times removed from the source. Which in one sense is actually more correct but also very misleading at a glance. Like in 2020.2 main thread time from multiple things showing up as wait times in the render thread in timeline (and the stats window).

So you really need to look at both the hierarchy and timeline and correlate them to figure out what is really going on.

And then you have to correlate that with the impact of specific settings.

 

I filed a issue for this (Case 1301741). It's definitely happening, also attaching two projects.

The drastic performance issue I see may stem from two separate issues.

Things I tried:

• entitydebugger on and off - doesn't change actualy frame rate, but the time consumed is ascribed to the entitydebugger instead of whatever else it is
• playmode and editmode - playmode has poor performance but it doesn't show in the profiler (however, presentationsystemgroup eats 2-5ms, seems to do that on both versions, and then editorloop accounts for the other 5 or more ms)
• interactionmodes default and no throttling (in edit mode, that makes a HUGE difference, in playmode, not so much)

There is a dependence of the profiling data on the editor layout, especially whether the entity debugger is open or not. I attached two layout files as well.

Some highlights:

Entitydebugger "taking the blame" on both platforms, but note how one takes much more time in the 2nd thumbnail, that's 2020.2:


... where it's actually something else in the editor loop.


Amazing result with nothrottling, but ONLY in edit mode:


Playmode: PresentationSystemGroup being slow as a snail (there's nothing in the scene that converts to an entity):


The poor PSG performance also seems to be mostly an (unnecessary) waitstate, though?

 

Projects for anyone interested in trying to reproduce. Under 60kb each, but will download a bunch of packages from the Unity registry.

Observe, in editmode and playmode alike, 10x worse performance in Unity 2020.2 compared to 2020.1. I get 130 fps in 2020.2 while I get over 1000 fps in 2020.1 with the same* project.

*2020.2 force upgrades URP to 10 from 9, and 10 is not availaible in 2020.1

Note: I accidentally deleted the render pipeline asset, it doesn't make a big difference in this empty scene test case but you may want to create a new one and set it up in Project Settings. Notably, doing this with the out of the box URP asset, this takes me down to 800 fps versus the builtin renderer.

SRP is the future though, don't use builtin, kids.

 

??? I think you misunderstood. I was pinged by the OP in the other thread so I thought I'd provide that relevant piece of info here. Did you assume I was trying to use this as a new continuation thread or something?

@Thygrrr As I'm not experiencing the editor perf reduction anymore in the newest 2020.1 and 2020.2 versions, can you try deleting your Library folders and then run the tests again? Last week, we encountered a weird issue with the DOTS suite of packages that impacted URP render performance (in the range of 10x - 20x like you described). We tried re-importing entire project, downgrading URP versions... But ultimately only deleting the Library folder solved the issue.

 

I did that about a dozen times, and about twice today for each Unity version while I was creating the bug report.

 

You don't say?
I mention it because it appears to be a load-dependent overhead, rather than a constant overhead. I could live and go on developing with 4 wasted milliseconds, but I can't live or work with 25 wasted millis.

I'm literally getting juddery gameplay even at "130 fps" here, and it often drops into the 20s on Unity 2021.1 and 2020.2 if I don't mitigate by reducing the camera and sceneview count. And it really FEELS like 20 fps.

Also ran into this loveable Christmas scoundrel today for the first time in 2 Months.
Maybe it's telling me to also go into vacation mode. Well. This IS my vacation.

(no worries, it's unrelated to my tests, it happened after importing the 2D Sprite package)

 

I ran both of your projects. But I'm not experiencing the performance issue in 2020.2. In fact 2020.2 seems to be more performant in playmode FPS. Here's a gif (right side is 2020.2):



Also just to be sure, I ran them individually (1 Unity instance at a time). Both were about 3-4ms (and both at ~2.2ms when scene view is closed).

My specs: Intel Core i7-7700K @ 4.20GHz; NVIDIA GeForce GTX 1080 (Windows 10 Home 19041.685)

 

All good! I was getting a little confused about the threads (beta, SRP, DOTS...) as well.

 

It's not that but good guess. I've been investigating exactly that for quite some time.

I wonder what it is... it does indeed feel like absent physics body smoothing. It is only one scene though.

(well, I figured out something, of course my code that runs in dynamic update has issues if Input System updates in FixedUpdate - this only affects some of the judder I have seen in my UI; as the judder I was mentioning before was actually on a camera following an interpolated physicsbody... will check if it actually runs after interpolation or before - but all judder remains as is, regardless of how I treat input, in Editor)

 

I have a i7-8086k, 32 GB ram, RTX 2080 Super.

This scene runs at 1000 fps on 2020.1 and at 130 in 2020.2 for me.

1000 fps is the minimum I expect for an empty scene. If I reduce the throttling to zero, I get 2000-3000 fps (only in 2020.1)

The fact that you have different figures makes me wonder if this may be an editor preferences issue. And to keep you worrying, BOTH your performances are pretty darn bad for an empty scene.

 

Is that from the Stats panel or your own FPS counter? Sure I can get above 1k fps in the Stats panel on 2020.1. But that number is pretty much fake.

2020.1:


2020.2:


To get the actual refresh rate, can you use this one below to test (you can drag it onto any GameObject):

Code (CSharp):

1. using UnityEngine;
2. using System.Collections;
3.  
4. public class FPSDisplay : MonoBehaviour {
5.     float deltaTime = 0.0f;
6.  
7.     void Update() {
8.         deltaTime += (Time.unscaledDeltaTime - deltaTime) * 0.1f;
9.     }
10.  
11.     void OnGUI() {
12.         int w = Screen.width, h = Screen.height;
13.  
14.         GUIStyle style = new GUIStyle();
15.  
16.         Rect rect = new Rect(0, 0, w, h * 2 / 100);
17.         style.alignment = TextAnchor.UpperLeft;
18.         style.fontSize = h * 2 / 100;
19.         style.normal.textColor = new Color(0.0f, 1.0f, 0.0f, 1.0f);
20.         float msec = deltaTime * 1000.0f;
21.         float fps = 1.0f / deltaTime;
22.         string text = string.Format("{0:0.0} ms ({1:0.} fps)", msec, fps);
23.         GUI.Label(rect, text, style);
24.     }
25. }

Source: https://wiki.unity3d.com/index.php/FramesPerSecond

 

This whole thread is about Editor performance...

I could live with "real" 120 fps in the editor, but not with 40, or juddery 120.

 

I just converted my game from 2020.1 to 2020.2. I get 47 fps in editor in the main scene in 2020.2 vs 60 fps in 2020.1. I didn't tested build yet, but game performance in the editor is very important, because we spend 99% of our time working in it. Something is so wrong. I see a huge spikes in the profiler every second or so every time I move a camera and the game feels very jittery and unresponsive. At this spikes editor loop eats like 20ms or so (I am using a standalone profiler).
This spikes remand me 19.3 version. It felt so bad too. Also I've noticed that such things happen every autumn Unity release. 20xx.1 version is always better than the other ones.
I am using HDRP and maybe there something wrong with it in 10 version, but so far I can't find a reason of such regression. I will capture profiler samples and leave them here later.
Thank you.

 

The plot thickens, this is in 2020.1 (i.e. not specific to 2020.2!).

I wrote my own fps counter and it's... nothing like what the stats gizmo displays. I didn't realize it was this far off. No wonder there's this much judder. (but bad performance remains, and is even worse in 2020.2)

Game renders at 200+ fps in a build. (verified both with the nvidia HUD and my own counter)
This case is not with an empty scene.

I'll check what is actually slowing it down so much. I suspect it's perhaps connected to me having my views spread across multiple monitors with different refresh rates and orientations.

 

This is not only in editor, its also in Build

I have been working on Unity 2020.1.0f1
then after updating to 2020.2.0f1 i noticed a big performance problem.


from 72 fps on Quest (and 90 on Quest 2) In 2020.1
30 ~ 50 fps on Quest using 2020.2

Also, the Editor Play mode performance has decreased bt a lot, and it's highly noticeable.

I had to Downgrade in order to get the performance fixed for my Xmas update

 

Be interesting to see a Profile Analysis (https://docs.unity3d.com/Packages/com.unity.performance.profile-analyzer@latest/) of the 2 versions

 

i will do it after work

 

volumetric fog in 2020.2 is totaly broken.
switching it ON and you loose all fps you got.

 

2021.2.xf



This bug has not been resolved in recent versions of Unity.
I've tested it with several versions.
Even in an empty scene, if you repeatedly hit keystrokes, frame drops occur.

 

@Thygrrr It may be a long shot but turn off incremental GC if you have it on. It can produce visible overhead.