Bring reference_hardware.json inline with machine used for weights (#5196)
Since `May 2023` after https://github.com/paritytech/substrate/pull/13548 optimization, `Blake2256` is faster with about 30%, that means that there is a difference of ~30% between the benchmark values we ask validators to run against and the machine we use for generating the weights.So if all validators, just barely pass the benchmarks our weights are potentially underestimated with about ~20%, so let's bring this two in sync. Same thing happened when we merged https://github.com/paritytech/polkadot-sdk/pull/2524 in `Nov 2023` SR25519-Verify became faster with about 10-15% ## Results Generated on machine from here: https://github.com/paritytech/devops/pull/3210 ``` +----------+----------------+--------------+-------------+-------------------+ | Category | Function | Score | Minimum | Result | +============================================================================+ | CPU | BLAKE2-256 | 1.00 GiBs | 783.27 MiBs |✅ Pass (130.7 %) | |----------+----------------+--------------+-------------+-------------------| | CPU | SR25519-Verify | 637.62 KiBs | 560.67 KiBs |✅ Pass (113.7 %) | |----------+----------------+--------------+-------------+-------------------| | Memory | Copy | 12.19 GiBs | 11.49 GiBs |✅ Pass (106.1 %) | ``` Discovered and discussed here: https://github.com/paritytech/polkadot-sdk/pull/5127#issuecomment-2258423469 ## Downsides Machines that barely passed the benchmark will suddenly find themselves bellow the benchmark, but since that is just an warning and everything else continues as before it shouldn't be too impactful and should give the validators the necessary information that they need to become compliant, since they actually aren't when compared with the used weights. --------- Signed-off-by: Alexandru Gheorghe <[email protected]>
Please register or sign in to comment