[Question] For the Java PB object generated by Protobuf, the serialization and deserialization using Fury are slower than using Protobuf.

[a1342772]

GrpcService.ModelInferRequest` resp = GrpcService.ModelInferRequest
                    .newBuilder()
                    .addAllInputs(inputs)
                    .build();

             byte[] data = fury.serializeJavaObject(resp);
             fury.deserializeJavaObject(data, GrpcService.ModelInferRequest.class);

             byte[] data = resp.toByteArray();
             GrpcService.ModelInferRequest.parseFrom(data); 

[a1342772]

@chaokunyang

[chaokunyang]

@a1342772 Java PB object generated by Protobuf is an internal object representation in Protobuf, it hold many internal state for protobuf usage only. You should not serialize such objects using other seiralization framework. You could define your data using POJO and pass such obejcts to Fury instead.

[chaokunyang]

@a1342772 If you do want serialize such Protobuf generated objects using Fury, you could do following optimization:

• Implement serializers by extends org.fury.serializer.Serializer for com.google.protobuf.Internal$IntList/DoubleList/FloatList/LongList/BooleanList/ShortList
• Implement serializer for com.google.protobuf.UnknownFieldSet
• Register all those serializers to Fury.

At last, Protobuf generated objects cached many fields data such as booleanArrayMemoizedSerializedSize/longArrayMemoizedSerializedSize/memoizedHashCode/memoizedIsInitialized/memoizedSize/bitField0_/, they are all redundant and only meaningful to Protobuf. You shoud find a way to skip serialization of those fields.

[a1342772]

1.The Java POJO is too large, and directly serializing it with Fury takes a long time. Therefore, it's necessary to convert it into a Protobuf (pb) object before serialization. Do you have any solutions to this problem?
2.Do you have examples of Java serializing Protobuf objects? @chaokunyang chao

[chaokunyang]

@a1342772 could you share the full proto definition here for the data you want to serialize? We can't do anything with the code you post in this issue

[chaokunyang]

1.The Java POJO is too large, and directly serializing it with Fury takes a long time. Therefore, it's necessary to convert it into a Protobuf (pb) object before serialization. Do you have any solutions to this problem? 2.Do you have examples of Java serializing Protobuf objects? @chaokunyang chao

I don't think serializing the pojo with Fury directly will take a longer time. There are already many companies switches from pb to fury and they all get several times speedup. If you could a fully reproducible code, and share jmh benchmark results, we could take a look at it. Otherwise, i don't think we can do anything.

[a1342772]

The application scenario is search and recommendation, where proto objects need to be constructed for model inference，The proto file is as follows:
grpc_service.proto.txt

The test code is as follows:

static Fury furyTest;

static {
    furyTest = Fury.builder()
            .withLanguage(Language.JAVA)
            .requireClassRegistration(false)
            .withCodegen(false)
            .withRefTracking(false)
            .withAsyncCompilation(true)
            .build();
}

public static void main(String[] args) {
    Map<String, ArrayList<FeaOutputV3>> itemsFeatDataV3 = new HashMap<>();
    int numFeatures = 320;
    int numColumns = 20;

    // 模拟特征数据
    for (int i = 0; i < numFeatures; i++) {
        String featureName = "feature_" + i;
        ArrayList<FeaOutputV3> featureDataList = new ArrayList<>();

        for (int j = 0; j < numColumns; j++) {
            IntList featureValues = new IntArrayList();
            for (int k = 0; k < numColumns; k++) {
                featureValues.add(i * numColumns + k);
            }
            FeaOutputV3 feaOutputV3 = new FeaOutputV3();
            feaOutputV3.setFixSizeList(featureValues);
            featureDataList.add(feaOutputV3);
        }

        itemsFeatDataV3.put(featureName, featureDataList);
    }

    // 构建输入张量
    GrpcService.ModelInferRequest.InferInputTensor.Builder tensorBuilder =
            GrpcService.ModelInferRequest.InferInputTensor.newBuilder();
    GrpcService.InferTensorContents.Builder tensorContentBuilder =
            GrpcService.InferTensorContents.newBuilder();

    List<GrpcService.ModelInferRequest.InferInputTensor> inputs = new ArrayList<>();
    itemsFeatDataV3.forEach((name, featList) -> {
        int l1 = featList.size();
        if (CollectionUtils.isNotEmpty(featList.get(0).getFixSizeList())) {
            int l2 = featList.get(0).getFixSizeList().size();
            IntList features = new IntArrayList(l1 * l2);

            featList.forEach(feaOutput -> features.addAll(feaOutput.getFixSizeList()));
            inputs.add(featureBuildWithFast(tensorBuilder, tensorContentBuilder, features, name, l1, l2));
        }
    });

    // 测试序列化性能
    for (int count = 0; count < 50; count++) {
        long startSerialization = System.nanoTime();
        GrpcService.ModelInferRequest resp = GrpcService.ModelInferRequest
                .newBuilder()
                .addAllInputs(inputs)
                .build();

        byte[] data = furyTest.serializeJavaObject(resp);
        furyTest.deserializeJavaObject(data, GrpcService.ModelInferRequest.class);

        //protobuf
        //byte[] data = resp.toByteArray();
        //GrpcService.ModelInferRequest.parseFrom(data);
        long endSerialization = System.nanoTime();
        long serializationTimeMs = (endSerialization - startSerialization) / 1_000_000;

        System.out.println("Serialization time (milliseconds): " + serializationTimeMs + " size: " + data.length);
    }
}



public static GrpcService.ModelInferRequest.InferInputTensor featureBuildWithFast(
        GrpcService.ModelInferRequest.InferInputTensor.Builder tensorBuilder1,
        GrpcService.InferTensorContents.Builder tensorContentBuilder2,
        IntList features, String featureName, int l1, int l2) {

    GrpcService.ModelInferRequest.InferInputTensor.Builder tensorBuilder =
            GrpcService.ModelInferRequest.InferInputTensor.newBuilder();
    GrpcService.InferTensorContents.Builder tensorContentBuilder =
            GrpcService.InferTensorContents.newBuilder();
    if (features == null) {
        features = new IntArrayList();
    }
    tensorContentBuilder.addAllIntContents(features);
    GrpcService.ModelInferRequest.InferInputTensor input = tensorBuilder.setName(featureName)
            .setDatatype("INT32")
            .addShape(l1)
            .addShape(l2)
            .setContents(tensorContentBuilder)
            .buildPartial();
    return input;
}`

[a1342772]

@chaokunyang The bytes obtained by directly serializing a Java object "itemsFeatDataV3" with Fury are several times larger than those obtained by serializing the object after constructing the Protobuf object.

[chaokunyang]

You disabled FURY JIT, the serialization will be slow:

furyTest = Fury.builder()
            .withLanguage(Language.JAVA)
            .requireClassRegistration(false)
            .withCodegen(false)
            .withRefTracking(false)
            .withAsyncCompilation(true)
            .build();

please create fury like this:

fury = Fury.builder()
            .withLanguage(Language.JAVA)
            .withRefTracking(false)
            .build();

And register all the types, please don't disable class registration:

fury.register(GrpcService.ModelInferRequest.InferInputTensor.class);
fury.register(GrpcService.ModelInferRequest.class);
fury.register(GrpcService.InferTensorContents.class);

And please warm serialization several times before benchmark.

[a1342772]

@chaokunyang

The protected object cannot be registered.

[chaokunyang]

@a1342772 You could pass a qualified string instead like fury.register("com.google.protobuf.GeneratedMessageLite.SerializedForm")
instead. Please take a look at the Java API in fury class.

[a1342772]

Version 0.9.0 does not provide the corresponding API.

[chaokunyang]

Use fury.getClassResolver().register() instead. I forgot to add it to BaseFury interface

[chaokunyang]

You can also invoke Class.forname() instead