我可以使用 protobuf-net 将 ConcurrentDictionaries 序列化和反序列化为同一个文件，然后读取它吗？

您可以声明一个容器（结构或类），其中包含您的字典作为成员，然后序列化/反序列化此容器。下面是一个示例：

[ProtoContract]
struct MyDictionaries
{
    [ProtoMember(1)]
    public ConcurrentDictionary<string, int> A;
    [ProtoMember(2)]
    public ConcurrentDictionary<string, char> B;
}

public static void Main()
{
    MyDictionaries dictionaries = new();
    dictionaries.A = new();
    dictionaries.A.TryAdd("A", 1);
    dictionaries.A.TryAdd("B", 2);
    dictionaries.B = new();
    dictionaries.B.TryAdd("C", 'c');
    dictionaries.B.TryAdd("D", 'd');
    Console.WriteLine($"dictionaries.A: {String.Join(", ", dictionaries.A)}");
    Console.WriteLine($"dictionaries.B: {String.Join(", ", dictionaries.B)}");
    MemoryStream stream = new();
    Serializer.Serialize(stream, dictionaries);
    Console.WriteLine();

    stream.Position = 0;
    var deserialized = Serializer.Deserialize<MyDictionaries>(stream);
    Console.WriteLine($"deserialized.A: {String.Join(", ", deserialized.A)}");
    Console.WriteLine($"deserialized.B: {String.Join(", ", deserialized.B)}");
}

输出：

dictionaries.A: [A, 1], [B, 2]
dictionaries.B: [D, d], [C, c]

deserialized.A: [A, 1], [B, 2]
deserialized.B: [D, d], [C, c]

在线演示。

文档明确指出支持类型，源代码中也有类型。IDictionary<TKey,TValue>ConcurrentDictionarySerializer

There is a bit to unpack here.

To serialize multiple independent objects to the same stream you can use / . This should allow you to serialize objects one after each other. See ProtoInclude for how to handle inheritance.Serializer.SerializeWithLengthPrefixSerializer.DeserializeWithLengthPrefix

To serialize each object I would tend to prefer to convert the objects into a separate type that is only used for serialization. Sometimes called a Data Transfer Object or DTO. This lets you separate the concerns of serialization from all kinds of domain logic, at the cost of some duplication of code.

There are a few ways to manage size. One approach is to only change changes to state, not the entire state. Something similar is sometimes used for games, where you only need to record the user input to allow you to replay the entire game. Another approach is compression, since states probably do not change that much. LZ4 claims to be one of the faster algorithms around. You might also want to keep things in memory if possible, since even the fastest SSD is much slower than memory.

I would highly recommend setting up a simple test environment. I.e. start by serializing a simple object, continue with a complex object, a dictionary of complex objects, and so on. This is also a good opportunity to measure performance.