Clarification on compute time per block for a parachain

[rhuttman]

I am trying to figure out how much compute time, and therefore weight, we can expect per block for a parachain. The substrate documentation says the following:

The maximum block weight should be equivalent to one-third of the target block time, allocating one third for block construction, one third for network propagation, and one third for import and verification.

Since the block time for a parachain is 12 seconds, this implies we should get ~4 seconds of compute time per block. While if we take a look at the parachain template from substrate it says:

/// We allow for 0.5 of a second of compute with a 12 second average block time.

Could you please clarify what the expected computed time is for a 12 second block on a parachain, and if the template is correct then why it is so low?

[rhuttman]

answer:

So, that documentation is referring to a standalone Substrate based chain. Parachains are a bit unoptimized right now with the two phase inclusion - the validators have to import/execute the block in two rounds, but still within the Relay Chain's block time of 6 seconds. And there seems to be some overhead with importing the PoV. So paras are limited to 1/4 of the execution time + only every other block, leading to a limit of 0.5 of 12 instead of 2 of 6.

The good news is that it's a big optimization we're working on called Asyncronous Backing. This will not only let paras use the normal 2 sec of execution time, but will also allow a 6s para block time, so an 8x improvement in execution capacity. There is a pretty detailed writeup in the issue on it: paritytech/polkadot#3779