godmodegalactus / mango_bencher Goto Github PK

From the profiler, create_ask_bid_transaction taxes 21% of time. Half of this is due to pk_from_str_like. I think it is a low-hanging fruit to get publickeys once for vector of PerpMarketCache.
Other observations:

• serde_json::from_str takes 6% of total time
• 16% of time is spent in futex_wait not related to send_transactions

It might be useful to pass node_id to the client.
The PR on bench-tps side is solana-labs/solana#29929 but lets wait until it lands.

This change will allow implementing the following:

• create a validator node on the same machine with client
• use this validator's identity as node_id also for the client

This way we could overcome unstaked connection limitations

My understanding is that the current validator + keeper + bencher setup generates non-determenistic load on the validator due to the unpredictable time of arrival of ask/bid orders (due to network uncertainty). The solution might be added to add an optional time feed that would specify for each ask/bid time they should have arrived in the orderbook.

The transactions sent by keeper are not really tracked. They should be tracked and store in transaction log.

it would be nice to have an option for random fees like in bench-tps: https://github.com/solana-labs/solana/blob/master/bench-tps/src/cli.rs#LL69C25-L69C25

0xd errors are there because market maker is trading on more than 5 markets. This can be avoided if user only trades on 5 random markets at a time.

Initially we close all the orders on all the markets for the market maker during startup phase.
Then we set random 5 markets where user can trade on.

We can then increase number of default market makers to 100 instead of 50.

In order to have a more active testnet, we would like to run mango_bencher not only for short tests like we did so far but also for long ones (12h).

It implies some changes in the code:

1. reporting side. Currently it reports once: https://github.com/godmodegalactus/mango_bencher/blob/master/src/main.rs#L217
   It would be nice to have a loop that it reports every minute or couple of minutes. Reporting side might be more time demanding (requires rpc calls) than doing the actual work, so might be impossible to analyze all the slots that has been created during time interval. The best effort approach is fine
2. checks for accounts balances. Currently, we update balances before the start of the program using cli tools. If we run for a long time, it would be simpler to check in the code that accounts have enough lammports and if not add some from authority like it is currently implemented in bench-tps.

Currently, the project uses both git submodules and cargo crates. This approach makes crates dependencies resolution to be more complicated. The proposal is to move submodules to Cargo.toml (one can specify github link and commit, if he/she needs a specific version)

Adding identity to pass identity to the client.

Add sequence enforcing program to bench it.