[Feature]: add SignatureCache for checkTx and DeliverTx

[chengzhinei]

Summary

Problem Definition

For a TX, when it is called through abci to deliverTx and checkTx, verification of the signature is used. However, it was found through pprof that the performance overhead of verifying the signature is still significant.
To save performance, the result of verifying the signature for the same TX can be cached during the first verification. This way, when the signature is needed for verification again, no decrypting and encrypting operations need to be performed, thereby saving performance.

Proposal

The code change is relatively simple, and this Pull-Request can be used as a reference.
https://github.com/cosmos/cosmos-sdk/pull/15778/files

https://github.com/okex

[alexanderbez]

Can you post benchmarks please? This is going to be dwarfed by the fact that apps might have to verify 100+ signatures in ProcessProposal, which shouldn't take that much time.

[chengzhinei]

Can you post benchmarks please? This is going to be dwarfed by the fact that apps might have to verify 100+ signatures in ProcessProposal, which shouldn't take that much time.

thanks for your reply.
In our okc(based on cosmos-sdk), we supported the evm, and we found that evm-tx verify the signature for many times, and it cost so much time. While for cosmos-tx, it only verify twice in DeliverTx and CheckTx, so maybe it has little effect.
Maybe it is not obvious when a block has not many txs, but when we have the pressure test, the performance is greatly affected.

These conclusions are based on pressure testing data, and haven't yet been compared to benchmarks. I'll perform some benchmarks next and provide feedback as soon as results are available.

Besides, you said you already use Hashi's LRU cache, and the LRU cache is used for getting the signBytes, but it still has to do the pubkey.VerifySignature, and this operation of verifying signatures has a performance overhead.

[IdaTucker]

Hi @chengzhinei, any updates about the benchmarks? If not, someone from Zondax could look into doing them, so that a final decision can be made whether the code change is worth it?

[educlerici-zondax]

this will be on hold until the implementation of ADR071 is completed.
After that, we can assess the possibility of applying these changes.