JsonRpcMessage.Converter.Read optimization

[Scooletz]

This PR replaces the copying behavior of JsonRpcMessage.Converter.Read with a single pass deserialization of it. It uses a Union struct that deserializes the raw content of JsonRpcMessage to a bag of properties. Then it uses it to check the conditions and properly construct underlying objects. This makes deserialization ~50% faster and also reduced the allocations.

Motivation and Context

I noticed the double deserialization in JsonRpcMessage.Converter.Read and wanted to make it faster.

How Has This Been Tested?

Benchmarks. Allocations are ~halfed.

Before

Method	Mean	Error	StdDev	Gen0	Allocated
DeserializeRequest	1.113 us	0.0167 us	0.0148 us	0.0572	720 B
DeserializeNotification	1.044 us	0.0080 us	0.0062 us	0.0515	664 B
DeserializeResponse	1.053 us	0.0057 us	0.0048 us	0.0515	648 B
DeserializeError	1.333 us	0.0138 us	0.0129 us	0.0954	1200 B

After

Method	Mean	Error	StdDev	Gen0	Allocated
DeserializeRequest	528.4 ns	9.85 ns	9.67 ns	0.0324	408 B
DeserializeNotification	443.9 ns	6.79 ns	6.35 ns	0.0305	384 B
DeserializeResponse	472.4 ns	6.77 ns	6.00 ns	0.0291	368 B
DeserializeError	732.5 ns	7.05 ns	6.59 ns	0.0687	864 B

Breaking Changes

Nope

Types of changes

• Optimization
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have added or updated documentation as needed

[eiriktsarpalis]

@stephentoub has pointed out that this call is wholly unnecessary. I'm curious what percentage of the performance improvement can be attributed to simply removing this and passing the root directly into JsonSerializer further down.

[Scooletz]

The question that I'd ask, and I might be missing something here, is whether you want to create the root at all? Clearly we'd benefit from not var rawText = root.GetRawText(); and this is one part of the coping. The other is creating JsonDocument that has a few fields and does some parsing. With Union we don't allocate at all beside for the properties that will be used in any case.

What do you want me to do then? Just remove this and rerun benchmarks and provide them in a comment?

[eiriktsarpalis]

The Union approach is clearly the fastest, but it is also the most elaborate. Before we commit to that, I'd like to see how much we can gain by fixing the obvious inefficiencies.

What do you want me to do then? Just remove this and rerun benchmarks and provide them in a comment?

No need to undo any changes yet, but if you could share benchmarks comparing all three versions would be great.

[Scooletz]

👍 Will do tomorrow (CEST)