Improve serializer performance #57327

steveharter · 2021-08-13T00:54:47Z

Primarily addresses serializer overhead of a given call, but there are general gains throughout.

Changes include:

Manual inlinining \ lifting of serializer calls. This avoids some if statements and the smaller size appears to help the jitter.
Avoids a cast against JsonTypeInfo<T> in unnecessary cases during serialization; this is expected to help perf for source-gen when falling back to the serializer.
Use of [MethodImplOptions.AggressiveOptimization] in hot, looping methods.
Optimized check for null when serializing either Nullable<T> and non-Nullable<T>.
Misc others, including some consistency changes for variable and method names.

It primarily resolves this issue with overhead:
resolves #56993

Overhead

A class with no members:

5.0

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	85.53 ns	0.375 ns	0.351 ns	-	-
Deserialize	268.87 ns	1.627 ns	1.522 ns	0.0049	32 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	106.5 ns	0.55 ns	0.49 ns	-	-
Deserialize	267.9 ns	0.55 ns	0.49 ns	0.0049	32 B

6.0 with this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	85.88 ns	0.527 ns	0.493 ns	-	-
Deserialize	263.78 ns	2.824 ns	2.503 ns	0.0049	32 B

Serialize overhead: 1.24x faster
Deserialize overhead: 1.015x faster

Also, if this test is changed to add 2 String properties:

5.0

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	196.3 ns	1.31 ns	1.16 ns	-	-
Deserialize	341.1 ns	0.74 ns	0.70 ns	0.0161	104 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	218.2 ns	1.06 ns	0.94 ns	-	-
Deserialize	334.5 ns	0.71 ns	0.63 ns	0.0161	104 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	198.4 ns	2.19 ns	2.05 ns	-	-
Deserialize	324.2 ns	2.35 ns	2.08 ns	0.0161	104 B

Serialize: 1.1x faster compared to before this PR
Deserialize: 1.03x faster compared to before this PR

but also should resolve these (verification needed for some):

resolves #57093
resolves #52640
resolves #52311
resolves #43413

Improvements in benchmarks with this PR

note: threshold is 2%
summary:
better: 31, geomean: 1.057
worse: 2, geomean: 1.071
total diff: 33

Slower	diff/base	Base Median (ns)	Diff Median (ns)	Modality
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.11	1386.49	1540.57	bimodal
System.Text.Json.Serialization.Tests.WriteJson.Serializ	1.03	377918.23	390056.88

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.19	119.98	100.69
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.16	137.06	117.90
System.Text.Json.Serialization.Tests.WriteJson.SerializeObjectProperty	1.09	246.90	225.56
System.Text.Json.Serialization.Tests.WriteJson.SerializeObjectPr	1.09	505.15	462.27
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.08	7678.84	7140.75
System.Text.Json.Serialization.Tests.WriteJson.Seria	1.07	251.04	234.99
System.Text.Json.Serialization.Tests.WriteJson.Serial	1.06	622.98	586.76
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8By	1.06	303.54	286.05
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf	1.06	393.55	371.04
System.Text.Json.Serialization.Tests.WriteJson.Seria	1.06	412.13	388.94
System.Text.Json.Serialization.Tests.WriteJson.Serial	1.06	862.66	815.38
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.06	124.69	118.10
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.05	559.86	530.84
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.05	331.31	314.80
System.Text.Json.Serialization.Tests.WriteJson.Seria	1.05	225.08	214.83
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.05	810.09	774.23
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.05	7249.32	6936.89
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.04	334.56	320.96
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.04	390.61	374.94
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.04	729.07	700.01
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.04	1076.71	1035.66
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.04	779.50	750.04
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.04	543.79	523.71
System.Text.Json.Serialization.Tests.WriteJson.Serial	1.04	658.87	636.45
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.03	74.08	71.59
System.Text.Json.Serialization.Tests.WriteJson.Serial	1.03	774.83	748.87
System.Text.Json.Serialization.Tests.WriteJson.Serializ	1.03	379279.76	367081.10
System.Text.Json.Serialization.Tests.ReadJson.Deseria	1.03	1145.78	1114.74
System.Text.Json.Serialization.Tests.ReadJson.Deseria	1.03	1556.95	1515.01
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.03	298.83	290.81
System.Text.Json.Serialization.Tests.WriteJson<HashSet>.SerializeToStrin	1.03	6444.12	6282.43

6.0 status (with this PR) of benchmarks compared to 5.0

note: threshold is 2%
summary:
better: 22, geomean: 1.268
worse: 2, geomean: 1.065
total diff: 24

Slower	diff/base	Base Median (ns)	Diff Median (ns)	Modality
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.10	482.68	530.03
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.03	668.86	690.92

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	2.13	47511.77	22281.48
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	2.10	47456.53	22649.40
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Byte	2.07	46860.63	22584.86
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.88	67083.49	35639.60
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Byte	1.76	63251.21	35890.00
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.74	63594.48	36651.60
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.14	45314.11	39891.80
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.12	45394.56	40385.56
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromUt	1.11	11901.75	10722.24
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.11	21191.19	19168.64
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromSt	1.10	12139.81	11039.48
System.Text.Json.Serialization.Tests.ReadJson<ImmutableSortedDictionary<String,	1.10	75880.24	69099.21
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.09	47138.25	43118.00
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.09	29213.97	26797.27
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.08	20351.18	18857.99
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf	1.07	27666.04	25944.15
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.07	22345.68	20956.26
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.06	1105.41	1039.88
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.06	1178.27	1113.83
System.Text.Json.Serialization.Tests.ReadJson<ImmutableSortedDictionary<String,	1.05	75538.87	72250.75
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.04	587.58	564.27
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromSt	1.03	12883.77	12530.95

ghost · 2021-08-13T00:54:52Z

Tagging subscribers to this area: @eiriktsarpalis, @layomia
See info in area-owners.md if you want to be subscribed.

Issue Details

Primarily addresses serializer overhead of a given call, but there are general gains throughout.

Changes include:

Manual inlinining \ lifting of serializer calls. This avoids some if statements and the smaller size appears to help the jitter.
Avoids a cast against JsonTypeInfo<T> in unnecessary cases during serialization; this is expected to help perf for source-gen when falling back to the serializer.
Use of [MethodImplOptions.AggressiveOptimization] in hot, looping methods.
Optimized check for null when serializing either Nullable<T> and non-Nullable<T>.
Misc others, including some consistency changes for variable and method names.

It primarily resolves this issue with overhead:
resolves #56993

Overhead

A class with no members:

5.0

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	85.53 ns	0.375 ns	0.351 ns	-	-
Deserialize	268.87 ns	1.627 ns	1.522 ns	0.0049	32 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	106.5 ns	0.55 ns	0.49 ns	-	-
Deserialize	267.9 ns	0.55 ns	0.49 ns	0.0049	32 B

6.0 with this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	85.88 ns	0.527 ns	0.493 ns	-	-
Deserialize	263.78 ns	2.824 ns	2.503 ns	0.0049	32 B

Serialize overhead: 1.24x faster
Deserialize overhead: 1.015x faster

Also, if this test is changed to add 2 String properties:

5.0

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	196.3 ns	1.31 ns	1.16 ns	-	-
Deserialize	341.1 ns	0.74 ns	0.70 ns	0.0161	104 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	218.2 ns	1.06 ns	0.94 ns	-	-
Deserialize	334.5 ns	0.71 ns	0.63 ns	0.0161	104 B

6.0 before this PR

Method	Mean	Error	StdDev	Gen 0	Allocated
Serialize	198.4 ns	2.19 ns	2.05 ns	-	-
Deserialize	324.2 ns	2.35 ns	2.08 ns	0.0161	104 B

Serialize: 1.1x faster compared to before this PR
Deserialize: 1.03x faster compared to before this PR

but also should resolve these (verification needed for some):

resolves #57093
resolves #52640
resolves #52311
resolves #43413

Improvements in benchmarks with this PR

note: threshold is 2%
summary:
better: 22, geomean: 1.055
worse: 4, geomean: 1.034
total diff: 26

Slower	diff/base	Base Median (ns)	Diff Median (ns)
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.04	1386.49	1442.69
System.Text.Json.Serialization.Tests.WriteJson<Dictionary<String, String>>.Seria	1.04	10160.87	10520.05
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.03	28357.84	29236.77
System.Text.Json.Serialization.Tests.WriteJson<Dictionary<String, String>>.Seria	1.03	10253.93	10566.39

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.11	119.98	108.35
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.11	137.06	123.94
System.Text.Json.Serialization.Tests.WriteJson.SerializeObjectProperty	1.10	246.90	224.10
System.Text.Json.Serialization.Tests.WriteJson.SerializeToStream	1.08	213.16	198.04
System.Text.Json.Serialization.Tests.ReadJson.Deseria	1.07	1556.95	1449.40
System.Text.Json.Serialization.Tests.WriteJson.Seria	1.07	225.08	210.27
System.Text.Json.Serialization.Tests.WriteJson.Seria	1.06	251.04	236.62
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.06	559.86	530.03
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.06	729.07	690.92
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.05	331.31	314.87
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.05	37465.99	35639.60
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8By	1.05	303.54	290.19
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.05	124.69	119.23
System.Text.Json.Serialization.Tests.WriteJson.SerializeObjectPr	1.05	505.15	483.23
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.04	779.50	747.46
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.04	298.83	287.05
System.Text.Json.Serialization.Tests.WriteJson.SerializeToUtf8Bytes	1.04	390.61	375.73
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.04	74.08	71.26
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.04	1076.71	1039.88
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.03	334.56	324.45
System.Text.Json.Serialization.Tests.WriteJson.SerializeToString	1.03	810.09	789.99
System.Text.Json.Serialization.Tests.WriteJson<HashSet>.SerializeToStrea	1.02	6130.58	5996.71

6.0 status (with this PR) of benchmarks compared to 5.0

note: threshold is 2%
summary:
better: 22, geomean: 1.268
worse: 2, geomean: 1.065
total diff: 24

Slower	diff/base	Base Median (ns)	Diff Median (ns)	Modality
System.Text.Json.Serialization.Tests.ReadJson.Deseri	1.10	482.68	530.03
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.03	668.86	690.92

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	2.13	47511.77	22281.48
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	2.10	47456.53	22649.40
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Byte	2.07	46860.63	22584.86
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.88	67083.49	35639.60
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Byte	1.76	63251.21	35890.00
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStream	1.74	63594.48	36651.60
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.14	45314.11	39891.80
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.12	45394.56	40385.56
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromUt	1.11	11901.75	10722.24
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.11	21191.19	19168.64
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromSt	1.10	12139.81	11039.48
System.Text.Json.Serialization.Tests.ReadJson<ImmutableSortedDictionary<String,	1.10	75880.24	69099.21
System.Text.Json.Serialization.Tests.ReadJson<ImmutableDictionary<String, String	1.09	47138.25	43118.00
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromStr	1.09	29213.97	26797.27
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.08	20351.18	18857.99
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf	1.07	27666.04	25944.15
System.Text.Json.Serialization.Tests.ReadJson<Dictionary<String, String>>.Deseri	1.07	22345.68	20956.26
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromUtf8Bytes	1.06	1105.41	1039.88
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.06	1178.27	1113.83
System.Text.Json.Serialization.Tests.ReadJson<ImmutableSortedDictionary<String,	1.05	75538.87	72250.75
System.Text.Json.Serialization.Tests.ReadJson.DeserializeFromString	1.04	587.58	564.27
System.Text.Json.Serialization.Tests.ReadJson<HashSet>.DeserializeFromSt	1.03	12883.77	12530.95

Author:	steveharter
Assignees:	steveharter
Labels:	`area-System.Text.Json`, `tenet-performance`
Milestone:	6.0.0

eiriktsarpalis · 2021-08-13T08:31:08Z

...tem.Text.Json/src/System/Text/Json/Serialization/Converters/Object/ObjectDefaultConverter.cs

@@ -16,6 +16,9 @@ internal class ObjectDefaultConverter<T> : JsonObjectConverter<T> where T : notn
    {
        internal override bool CanHaveIdMetadata => true;

+#if NET6_0_OR_GREATER
+        [MethodImpl(MethodImplOptions.AggressiveOptimization)]


Out of curiosity, what are some of the optimizations that were being missed without this attribute? I had tried using it a few months back on the JsonConverter<T>.TryWrite method (which is also fairly large) and did not receive any perf improvements.

I am not sure how to find this out, with tiered jitting and all (e.g. hit a breakpoint after running a 1,000 times to inspect the native code?). Also, I did not run with crossgen2 yet, but typically that has the same results once the warm-up phase is done.

UPDATE: the assembly can be viewed by a checked build and using DOTNET_JitDisasm=methodName.

The results are small when there are a couple properties but should be greater when there are more properties. I had to run benchmarks a few times to make sure noise isn't a factor. I welcome investigation from others on this.

Here's some results with and without AggressiveOptimization. Again, I assume the more properties, the more savings. Also I tend to focus on the Min column more than Mean.

With AO and without

Using a class with 2 string properties; I ran 5 times and took the fastest.

With AO:

Method Mean Error StdDev Gen 0 Allocated

Serialize 197.1 ns 0.34 ns 0.27 ns - -

Deserialize 330.1 ns 0.42 ns 0.38 ns 0.0161 104 B

Without AO:

Method Mean Error StdDev Gen 0 Allocated

Serialize 198.8 ns 0.10 ns 0.09 ns - -

Deserialize 333.1 ns 1.14 ns 1.01 ns 0.0161 104 B

Running ReadJson<LargeStructWithProperties> and WriteJson<LargeStructWithProperties> since it has 10 properties. I ran the benchmarks 3 times and took the fastest. The "Stream" cases didn't get any faster -- not sure why, although that does have a different code path in the object converter.

With AO:

Method Mean Error StdDev Median Min Max Gen 0 Allocated

SerializeToString 650.1 ns 5.49 ns 4.87 ns 648.0 ns 643.8 ns 656.6 ns 0.0765 488 B

SerializeToUtf8Bytes 618.7 ns 7.81 ns 7.31 ns 615.9 ns 609.7 ns 633.3 ns 0.0577 376 B

SerializeToStream 766.6 ns 10.86 ns 9.63 ns 766.1 ns 754.1 ns 783.0 ns 0.0368 232 B

SerializeObjectProperty 831.6 ns 13.94 ns 11.64 ns 828.7 ns 819.1 ns 862.5 ns 0.1347 848 B

Without AO:

Method Mean Error StdDev Median Min Max Gen 0 Allocated

SerializeToString 667.1 ns 2.42 ns 2.15 ns 667.5 ns 661.5 ns 669.4 ns 0.0754 488 B

SerializeToUtf8Bytes 620.1 ns 3.00 ns 2.66 ns 620.0 ns 615.0 ns 624.4 ns 0.0598 376 B

SerializeToStream 760.7 ns 0.80 ns 0.67 ns 760.6 ns 759.8 ns 762.1 ns 0.0370 232 B

SerializeObjectProperty 839.5 ns 8.38 ns 7.84 ns 838.3 ns 826.0 ns 851.7 ns 0.1340 848 B

With AO:

Method Mean Error StdDev Median Min Max Gen 0 Allocated

DeserializeFromString 1.114 us 0.0083 us 0.0070 us 1.114 us 1.104 us 1.131 us 0.0318 200 B

DeserializeFromUtf8Bytes 1.042 us 0.0118 us 0.0111 us 1.040 us 1.028 us 1.065 us 0.0289 200 B

DeserializeFromStream 1.519 us 0.0189 us 0.0176 us 1.516 us 1.493 us 1.556 us 0.0478 328 B

Without AO:

Method Mean Error StdDev Median Min Max Gen 0 Allocated

DeserializeFromString 1.151 us 0.0046 us 0.0041 us 1.152 us 1.143 us 1.156 us 0.0276 200 B

DeserializeFromUtf8Bytes 1.063 us 0.0141 us 0.0132 us 1.065 us 1.043 us 1.081 us 0.0295 200 B

DeserializeFromStream 1.497 us 0.0053 us 0.0047 us 1.496 us 1.489 us 1.507 us 0.0479 328 B

In the end, after running thousands of times, the jitted (native) code may be more or less the same with or without the [AO]. But [AO] should produce faster code sooner.

Can we please avoid [AO]?
Methods with [AO] aren't instrumented so won't be able to benefit from PGO (devirtualize calls, move cold parts of methods further from hot ones, etc) - e.g. run your benchmark again with DOTNET_ReadyToRun=0, DOTNET_TC_QuickJitForLoops=1, DOTNET_TieredPGO=1 with and without [AO]

Also, when a method is promoted to tier1 naturally (without [AO]`) we can:

Get rid of static initializations in codegen

Convert static readonly fields to constants

Inliner is able to resolve call tokens and produce better results

/cc @AndyAyersMS

[AO] should be avoided. It causes methods to be optimized prematurely. We use it only in a handful of places where we feel like we can't afford to ever run unoptimized code.

Thanks, I'll re-run with those settings.

See #58209 for reverting the 4 usages of [AO]. Note I was going to keep the 2 in JsonConverter<T> since there does seem to benefits at least sometimes, but for V7 I think it makes sense to remove them based upon offline discussion with @EgorBo and "DynamicPGO" scenarios. I don't plan on porting that PR to V6.

eiriktsarpalis · 2021-08-13T08:37:55Z

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs


+#if NET6_0_OR_GREATER
+        [MethodImpl(MethodImplOptions.AggressiveOptimization)]


I was not able to squeeze out any performance gains when applying aggressive optimizations to this method in the past.

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializerOptions.cs

...ries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializerOptions.Converters.cs

...raries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializer.Write.ByteArray.cs

...ibraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializer.Write.Element.cs

eiriktsarpalis · 2021-08-13T09:18:47Z

...ibraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializer.Write.Helpers.cs


-                WriteCore(converter, writer, value, jsonTypeInfo.Options, ref state);
+            // For performance, the code below is a lifted WriteCore() above.
+            if (converter is JsonConverter<TValue> typedConverter)


Assuming generic type tests are expensive, is this also an opportunity for us to improve the hot path?

@eiriktsarpalis what kind of converter is the most popular? (is it DefaultObjectConverter<T> ?)

Possibly, alongside with all the common primitive type converters. Why do you ask?

I was just wondering if we could introduce a fast path for the most popular one if it's sealed. Cast to sealed classes is almost free.

Since this concerns root-level types ObjectDefaultConverter<T> would probably be the most common type. If we do try this it out we should make sure it doesn't regress other common root-level converters types like JsonNodeConverter or custom converters registered by users.

...ibraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializer.Write.Helpers.cs

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializer.Read.String.cs

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

eiriktsarpalis · 2021-08-13T11:57:55Z

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

+        /// Performance optimization. Overridden by the nullable converter to prevent boxing of values by the JIT.
+        /// Although the JIT added support to detect this IL pattern (box+isinst+br), it still manifests in TryWrite().
+        /// </summary>
+        internal virtual bool IsNull(in T value) => value is null;


Ran a few experiments locally and it seems like the in modifier is the culprit for the boxing behavior. Changing to

private bool IsNull(T value) => value is null;

makes the regression go away. It comes at the cost of having to pass the parameter by value but it's negligible compared to boxing. This also avoids needing to make a virtual call. cc @EgorBo

Thanks. Based on this and your other comment around using default(T) is null in that LOC in question, I re-ran some micro benchmarks and found, as expected, that value types such as int are faster, while reference types including string will be a bit slower. However, removing the extra caching and make the code simpler is also a win, so I plan on making these changes in the next commit and re-running the full benchmarks again.

@eiriktsarpalis how can I reproduce the boxing? e.g. I tried sharplab and don't see any boxing.

Could this be a reproduction? Not sure what L000e: call 0x00007ffd016a1b10 is doing but seems consistent with what I'm seeing when I profile the repro:

FWIW here's a commit that reproduces the boxing behavior: eiriktsarpalis@a44c0ee

It can be observed running the benchmark added here dotnet/performance#1920

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

eiriktsarpalis

Thanks!

src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

…information # By dotnet-maestro[bot] (4) and others # Via GitHub * origin/main: (58 commits) Localized file check-in by OneLocBuild Task (dotnet#57384) [debugger][wasm] Support DebuggerProxyAttribute (dotnet#56872) Account for type mismatch of `FIELD_LIST` members in LSRA (dotnet#57450) Qualify `sorted_table` allocation with `nothrow` (dotnet#57467) Rename transport packages to follow convention (dotnet#57504) Generate proper DWARF reg num for ARM32 (dotnet#57443) Enable System.Linq.Queryable and disable dotnet#50712 (dotnet#57464) Mark individual tests for 51211 (dotnet#57463) Fix Length for ReadOnlySequence created out of sliced Memory owned by MemoryManager (dotnet#57479) Add JsonConverter.Write/ReadAsPropertyName APIs (dotnet#57302) Remove workaround for dotnet/sdk#19482 (dotnet#57453) Do not drain HttpContentReadStream if the connection is disposed (dotnet#57287) [mono] Fix a few corner case overflow operations (dotnet#57407) make use of ports in SPN optional (dotnet#57159) Fixed H/3 stress server after the last Kestrel change (dotnet#57356) disable a failing stress test. (dotnet#57473) Eliminate temporary byte array allocations in the static constructor of `IPAddress`. (dotnet#57397) Update dependencies from https://github.com/dotnet/emsdk build 20210815.1 (dotnet#57447) [main] Update dependencies from mono/linker (dotnet#57344) Improve serializer performance (dotnet#57327) ... # Conflicts: # src/mono/wasm/debugger/BrowserDebugProxy/MemberReferenceResolver.cs # src/mono/wasm/debugger/BrowserDebugProxy/MonoProxy.cs # src/mono/wasm/debugger/BrowserDebugProxy/MonoSDBHelper.cs

kunalspathak · 2021-08-17T16:42:37Z

Improvements: dotnet/perf-autofiling-issues#612, dotnet/perf-autofiling-issues#611, dotnet/perf-autofiling-issues#551

Improve serializer performance

bc27b79

steveharter added area-System.Text.Json tenet-performance Performance related issue labels Aug 13, 2021

steveharter added this to the 6.0.0 milestone Aug 13, 2021

steveharter requested review from eiriktsarpalis and layomia August 13, 2021 00:54

steveharter self-assigned this Aug 13, 2021