Files
pezkuwi-subxt/substrate/utils/frame/benchmarking-cli/src/overhead
Bastian Köcher ca5f10567a sc-block-builder: Remove BlockBuilderProvider (#2099)
The `BlockBuilderProvider` was a trait that was defined in
`sc-block-builder`. The trait was implemented for `Client`. This
basically meant that you needed to import `sc-block-builder` any way to
have access to the block builder. So, this trait was not providing any
real value. This pull request is removing the said trait. Instead of the
trait it introduces a builder for creating a `BlockBuilder`. The builder
currently has the quite fabulous name `BlockBuilderBuilder` (I'm open to
any better name 😅). The rest of the pull request is about
replacing the old trait with the new builder.

# Downstream code changes

If you used `new_block` or `new_block_at` before you now need to switch
it over to the new `BlockBuilderBuilder` pattern:

```rust
// `new` requires a type that implements `CallApiAt`. 
let mut block_builder = BlockBuilderBuilder::new(client)
                // Then you need to specify the hash of the parent block the block will be build on top of
		.on_parent_block(at)
                // The block builder also needs the block number of the parent block. 
                // Here it is fetched from the given `client` using the `HeaderBackend`
                // However, there also exists `with_parent_block_number` for directly passing the number
		.fetch_parent_block_number(client)
		.unwrap()
                // Enable proof recording if required. This call is optional.
		.enable_proof_recording()
                // Pass the digests. This call is optional.
                .with_inherent_digests(digests)
		.build()
		.expect("Creates new block builder");
```

---------

Co-authored-by: Sebastian Kunert <skunert49@gmail.com>
Co-authored-by: command-bot <>
2023-11-03 19:06:31 +01:00
..
2023-09-04 12:02:32 +03:00

The benchmark overhead command

Each time an extrinsic or a block is executed, a fixed weight is charged as "execution overhead". This is necessary since the weight that is calculated by the pallet benchmarks does not include this overhead. The exact overhead to can vary per Substrate chain and needs to be calculated per chain. This command calculates the exact values of these overhead weights for any Substrate chain that supports it.

How does it work?

The benchmark consists of two parts; the BlockExecutionWeight and the ExtrinsicBaseWeight. Both are executed sequentially when invoking the command.

BlockExecutionWeight

The block execution weight is defined as the weight that it takes to execute an empty block. It is measured by constructing an empty block and measuring its executing time. The result are written to a block_weights.rs file which is created from a template. The file will contain the concrete weight value and various statistics about the measurements. For example:

/// Time to execute an empty block.
/// Calculated by multiplying the *Average* with `1` and adding `0`.
///
/// Stats [NS]:
///   Min, Max: 3_508_416, 3_680_498
///   Average:  3_532_484
///   Median:   3_522_111
///   Std-Dev:  27070.23
///
/// Percentiles [NS]:
///   99th: 3_631_863
///   95th: 3_595_674
///   75th: 3_526_435
pub const BlockExecutionWeight: Weight =
    Weight::from_parts(WEIGHT_REF_TIME_PER_NANOS.saturating_mul(3_532_484), 0);

In this example it takes 3.5 ms to execute an empty block. That means that it always takes at least 3.5 ms to execute any block. This constant weight is therefore added to each block to ensure that Substrate budgets enough time to execute it.

ExtrinsicBaseWeight

The extrinsic base weight is defined as the weight that it takes to execute an empty extrinsic. An empty extrinsic is also called a NO-OP. It does nothing and is the equivalent to the empty block form above. The benchmark now constructs a block which is filled with only NO-OP extrinsics. This block is then executed many times and the weights are measured. The result is divided by the number of extrinsics in that block and the results are written to extrinsic_weights.rs.

The relevant section in the output file looks like this:

 /// Time to execute a NO-OP extrinsic, for example `System::remark`.
/// Calculated by multiplying the *Average* with `1` and adding `0`.
///
/// Stats [NS]:
///   Min, Max: 67_561, 69_855
///   Average:  67_745
///   Median:   67_701
///   Std-Dev:  264.68
///
/// Percentiles [NS]:
///   99th: 68_758
///   95th: 67_843
///   75th: 67_749
pub const ExtrinsicBaseWeight: Weight =
    Weight::from_parts(WEIGHT_REF_TIME_PER_NANOS.saturating_mul(67_745), 0);

In this example it takes 67.7 µs to execute a NO-OP extrinsic. That means that it always takes at least 67.7 µs to execute any extrinsic. This constant weight is therefore added to each extrinsic to ensure that Substrate budgets enough time to execute it.

Invocation

The base command looks like this (for debugging you can use --release):

cargo run --profile=production -- benchmark overhead --dev

Output:

# BlockExecutionWeight
Running 10 warmups...
Executing block 100 times
Per-block execution overhead [ns]:
Total: 353248430
Min: 3508416, Max: 3680498
Average: 3532484, Median: 3522111, Stddev: 27070.23
Percentiles 99th, 95th, 75th: 3631863, 3595674, 3526435
Writing weights to "block_weights.rs"

# Setup
Building block, this takes some time...
Extrinsics per block: 12000

# ExtrinsicBaseWeight
Running 10 warmups...
Executing block 100 times
Per-extrinsic execution overhead [ns]:
Total: 6774590
Min: 67561, Max: 69855
Average: 67745, Median: 67701, Stddev: 264.68
Percentiles 99th, 95th, 75th: 68758, 67843, 67749
Writing weights to "extrinsic_weights.rs"

The complete command for Polkadot looks like this:

cargo run --profile=production -- benchmark overhead --chain=polkadot-dev --wasm-execution=compiled --weight-path=runtime/polkadot/constants/src/weights/

This will overwrite the the block_weights.rs and extrinsic_weights.rs files in the Polkadot runtime directory. You can try the same for Rococo and to see that the results slightly differ. 👉 It is paramount to use --profile=production and --wasm-execution=compiled as the results are otherwise useless.

Output Interpretation

Lower is better. The less weight the execution overhead needs, the better. Since the weights of the overhead is charged per extrinsic and per block, a larger weight results in less extrinsics per block. Minimizing this is important to have a large transaction throughput.

Arguments

  • --chain / --dev Set the chain specification.
  • --weight-path Set the output directory or file to write the weights to.
  • --repeat Set the repetitions of both benchmarks.
  • --warmup Set the rounds of warmup before measuring.
  • --wasm-execution Should be set to compiled for correct results.
  • --mul
  • --add
  • --metric
  • --weight-path
  • --header

License: Apache-2.0