Optimize ABI API to have less allocation

[LukaszRozmej]

Encoding and Decoding ABI allocates a lot, rewrite it to a version that will allocate a lot less (stream-like or spans?) and update usages.

[ssonthal]

Hi @LukaszRozmej can I work on this?

I read a bit about streams and spans. Streams are likely to work better when we have large chunk of data of variable sizes while spans work better with structured fixed-size data and the latter offers better performance enhancement, maintains concurrency in a better way.

What would you suggest is a better approach? I am currently not aware about the data that Abi encoding deals with.

[LukaszRozmej]

Hi @LukaszRozmej can I work on this?

Sure, take a look at: https://github.com/NethermindEth/nethermind/blob/master/src/Nethermind/Nethermind.Abi/AbiEncoder.cs and everything under it (AbiTypes etc). This now allocates a ton of objects when you Decode and ton of byte arrays when you Encode.

Think how can you allocate as little as possible. For inspiration you can check how our Rlp Encoding/Decoding works with NettyRlpStream that is based on pooled memory.

[ssonthal]

@LukaszRozmej can we try to optimize this? -

In AbiBytes class's Encode method PadRight method is used like this - ,

The PadRight method create new byte array allocations like this and this method is used at multiple places.

[ssonthal]

The Slice method in the PadRight method also create new allocations and used heavily ->

[ssonthal]

Hey @LukaszRozmej , gentle nudge for your advise on this.

[LukaszRozmej]

We definitely should be able to, but this is just tip of the iceberg and I think there are way more ways to optimise ABI

[ssonthal]

Maybe we can break it down to subtasks for incremental progress?

[ssonthal]

Hey @LukaszRozmej so for PadLeft and Slice methods, we can start changing the concerned files by using object pools instead of new allocations.

Can I take his as the first subtask for this issue?

[LukaszRozmej]

Hey @LukaszRozmej so for PadLeft and Slice methods, we can start changing the concerned files by using object pools instead of new allocations.

Can I take his as the first subtask for this issue?

I think using object pools might be hard as a lot of those arrays are then later returned upstream, so the upstream would have to track and eventually return them.

But if we try changing to Span<> or maybe better here to Memory<> in a lot of places we might be able to still reduce allocations substantially.

[ssonthal]

Got it! So I will try to use Span/Memory instead of creating new byte arrays.