Chain performance improvement

[pythonberg1997]

Summary

The performance of cosmos based chain is poor.

Problem Definition

The cosmos chain has low TPS. I think there are two obvious existential problems:

• No pre-cache/pre-fetch mechanism
  For example, a txn may be executed(baseapp.runTx()) in three different stages(PrepareProposal/ProcessProposal, CheckTx, DeliverTx). So I think we could find a way to do caching.
• The global mutex in tendermint
  There's a global mutex in tendermint localClient. But some operations are read operations and can be executed concurrently like abci query.

Work Breakdown

I was wondering if you guys had any thoughts on the poor performance so far.

[alexanderbez]

Hey @pythonberg1997 thanks for opening an issue, but this is sort of a blanket critique that isn't really capturing what's going on, especially without context and raw numbers. A few things:

• Cosmos SDK never sought ultra high TPS as a feature -- this was never really an aim, although chains like Sei are pushing the bounds of PBFT-based protocols, i.e. where you have quadratic communication.
• "Caching" of txs is tricky because different contexts are based/branched off of different states and managing that is non-trivial.
• We have a few major themes in the works that will drastically improve things, such as:
  • Optimistic execution RFC and implementation
  • Store v2
  • CometBFT storage and execution (such as mutex management) improvements

Perhaps you can share some profiles or numbers with us to see where exactly you're experiencing bottlenecks?