Fix expensive typo in JsonNode

[stephentoub]

Fixes#78089 (but assuming this is the right fix, we should backport this to release/7.0)

[ghost]

Tagging subscribers to this area: @dotnet/area-system-text-json, @gregsdennis
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Fixes #78089 (but assuming this is the right fix, we should backport this to release/7.0)

Author:	stephentoub
Assignees:	-
Labels:	area-System.Text.Json
Milestone:	-

[stephentoub]

Using the repro benchmark from the linked issue:

usingBenchmarkDotNet.Attributes;usingBenchmarkDotNet.Running;usingSystem.Text.Json.Nodes;[MemoryDiagnoser]publicpartialclassProgram{staticvoidMain(string[]args)=>BenchmarkSwitcher.FromAssembly(typeof(Program).Assembly).Run(args);privateJsonObject[]_jsonObjs=newJsonObject[100_000];[Benchmark]publicvoidSum(){for(inti=0;i<100_000;i++){_jsonObjs[i]=newJsonObject(){["aaa"]=1,["bbbb"]=2,["cccc"]=newJsonArray(){newJsonObject(){["aaaa"]="aaaaa",["ffffff"]="ffffffffffffff"}}};}}}

Method	Runtime	Mean	Error	StdDev	Ratio	Allocated
Sum	.NET 6.0	46.32 ms	0.926 ms	1.067 ms	1.10	127.41 MB
Sum	.NET 7.0	292.87 ms	5.241 ms	7.174 ms	6.89	313.13 MB
Sum	pr	42.42 ms	0.617 ms	0.547 ms	1.00	128.94 MB

[EgorBo]

I assume this scenario is not covered in the dotnet/performance suite or we somehow missed this.

[eiriktsarpalis]

Yup, fix is spot on. We should backport this.

[eiriktsarpalis]

Test failures are unrelated, mergin away.

[eiriktsarpalis]

/backport to release/7.0

[github-actions]

Started backporting to release/7.0: https://github.com/dotnet/runtime/actions/runs/3435456807

[adamsitnik]

@stephentoub we have added the "Allocation ratio" column to BDN based on your request: dotnet/BenchmarkDotNet#722 and I can see that you have removed it here while I would expect this particular issue to be perfect use case for it. May I ask why? Does the implementation do not meet your requirements?

[stephentoub]

I can see that you have removed it here while I would expect this particular issue to be perfect use case for it. May I ask why?

Because in this case I was focusing on the throughput impact rather than on the allocation impact, and removed extra columns to try to focus attention (e.g. RatioSD, Gen0, and Alloc Ratio). The 6x improvement in throughput was the primary thing to show, rather than the 2x reduction in allocation.

[dakersnar]

I can confirm that the newly added unit tests (dotnet/performance#2714) are shown as heavily regressing in the .NET 7 vs 6 GA performance report.

Is there anything more that needs to be done to confirm we have addressed the problem with this backport, or are we confident this is a complete fix?

Details

System.Text.Json.Node.Tests.Perf_ParseThenWrite.ParseThenWrite(IsDataIndented: False, TestCase: LotsOfStrings)

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.31	+12534	Windows 11	Arm64	Microsoft SQ1 3.0 GHz	bimodal
Slower	0.43	+12970	macOS 13.0	Arm64	Apple M1
Slower	0.34	+12660	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.26	+12442	Windows 10	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)
Slower	0.30	+12568	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)	several?
Slower	0.38	+12563	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	0.33	+12660	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.28	+7782	raspbian 11	Arm	ARMv7 Processor rev 3 (v7l)
Slower	0.27	+12513	macOS Monterey 12.6	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

System.Text.Json.Node.Tests.Perf_ParseThenWrite.ParseThenWrite(IsDataIndented: True, TestCase: LotsOfStrings)

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.36	+12469	Windows 11	Arm64	Microsoft SQ1 3.0 GHz	several?
Slower	0.46	+12995	macOS 13.0	Arm64	Apple M1
Slower	0.37	+12691	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	bimodal
Slower	0.29	+12533	Windows 10	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)
Slower	0.33	+12578	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)
Slower	0.42	+12552	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	0.39	+12652	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.29	+7724	raspbian 11	Arm	ARMv7 Processor rev 3 (v7l)
Slower	0.30	+12551	macOS Monterey 12.6	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

System.Text.Json.Node.Tests.Perf_ParseThenWrite.ParseThenWrite(IsDataIndented: False, TestCase: Json400KB)

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.43	+2891666	Windows 11	Arm64	Microsoft SQ1 3.0 GHz	bimodal
Slower	0.56	+2887897	macOS 13.0	Arm64	Apple M1
Slower	0.49	+2913026	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	bimodal
Slower	0.45	+2964348	Windows 10	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)
Slower	0.39	+2765957	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)
Slower	0.58	+2912668	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	0.48	+2937050	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.55	+1675483	raspbian 11	Arm	ARMv7 Processor rev 3 (v7l)	several?
Slower	0.42	+2938707	macOS Monterey 12.6	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

System.Text.Json.Node.Tests.Perf_ParseThenWrite.ParseThenWrite(IsDataIndented: False, TestCase: LotsOfNumbers)

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.51	+10563	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Slower	0.61	+10136	macOS 13.0	Arm64	Apple M1
Slower	0.55	+10666	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.46	+10542	Windows 10	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)
Slower	0.51	+10581	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)
Slower	0.58	+10596	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	0.50	+10711	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	0.44	+6202	raspbian 11	Arm	ARMv7 Processor rev 3 (v7l)
Slower	0.45	+10558	macOS Monterey 12.6	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

(etc for all other benchmarks)

[eiriktsarpalis]

Yes, we're fairly confident that the regression has been addressed. I've verified that the particular benchmark has parity between .NET 6 and the patched .NET 7 branch.