Polynomial allocation benchmark jump in AztecIvc #1072
Comments
ledwards2225
added a commit
to AztecProtocol/aztec-packages
that referenced
this issue
Aug 14, 2024
Introduces more in depth testing of the AztecIvc scheme, including tests that mimic the benchmark setup as well as more in depth tests for the inter-circuit databus consistency checks. Other noteworthy components of this PR: - Introduces utility class `PrivateFunctionExecutionMockCircuitProducer` for consistently generating mock circuits (shared between the benchmark and this new test suite). Subclass `MockDatabusProducer` incorporates nontrivial databus interactions into these circuits for the first time. - Removes the old `accumulate()` method on AztecIvc (which simultaneously added recursive verifiers and performed prover work) in favor of two methods that separate the logic but combine to the same effect: `complete_kernel_circuit_logic()` and a new `accumulate()` (which now only performs the prover work). Confusing jump in benchmark explained in [this](AztecProtocol/barretenberg#1072) issue: ``` ----------------------------------------------------------------------------------------- Benchmark Time CPU Iterations UserCounters... ----------------------------------------------------------------------------------------- AztecIVCBench/FullStructured/6 38855 ms 36066 ms 1 Arithmetic::accumulate=4.03742M DEBUG:ProverPolynomials()=12 DEBUG:ProverPolynomials()(t)=3.68002G function ms % sum construct_circuits(t) 4474 11.61% ProverInstance(Circuit&)(t) 6787 17.62% ProtogalaxyProver::fold_instances(t) 21788 56.55% Decider::construct_proof(t) 1655 4.30% ECCVMProver(CircuitBuilder&)(t) 235 0.61% ECCVMProver::construct_proof(t) 2585 6.71% TranslatorProver::construct_proof(t) 824 2.14% Goblin::merge(t) 177 0.46% ```
AztecBot
pushed a commit
that referenced
this issue
Aug 15, 2024
Introduces more in depth testing of the AztecIvc scheme, including tests that mimic the benchmark setup as well as more in depth tests for the inter-circuit databus consistency checks. Other noteworthy components of this PR: - Introduces utility class `PrivateFunctionExecutionMockCircuitProducer` for consistently generating mock circuits (shared between the benchmark and this new test suite). Subclass `MockDatabusProducer` incorporates nontrivial databus interactions into these circuits for the first time. - Removes the old `accumulate()` method on AztecIvc (which simultaneously added recursive verifiers and performed prover work) in favor of two methods that separate the logic but combine to the same effect: `complete_kernel_circuit_logic()` and a new `accumulate()` (which now only performs the prover work). Confusing jump in benchmark explained in [this](#1072) issue: ``` ----------------------------------------------------------------------------------------- Benchmark Time CPU Iterations UserCounters... ----------------------------------------------------------------------------------------- AztecIVCBench/FullStructured/6 38855 ms 36066 ms 1 Arithmetic::accumulate=4.03742M DEBUG:ProverPolynomials()=12 DEBUG:ProverPolynomials()(t)=3.68002G function ms % sum construct_circuits(t) 4474 11.61% ProverInstance(Circuit&)(t) 6787 17.62% ProtogalaxyProver::fold_instances(t) 21788 56.55% Decider::construct_proof(t) 1655 4.30% ECCVMProver(CircuitBuilder&)(t) 235 0.61% ECCVMProver::construct_proof(t) 2585 6.71% TranslatorProver::construct_proof(t) 824 2.14% Goblin::merge(t) 177 0.46% ```
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
For unknown reasons, this seemingly unrelated PR more than doubles the time it takes to allocate polynomials in the$2^{19}$ . (I confirmed this explicitly but its also evident from the fact that none of the other bench metrics increase). The
aztec_ivc_bench. Both master and branch are constructing polys of sizeclient_ivc_benchwas not affected. I can't explain this. It's possible it's simply an accounting error in the bench timings but I can't see how that could be either. I'm leaving this an open question for now since AztecIvc is not fully integrated outside of bberg, and this pattern of instantiating memory for all polynomials at once is about to change.Before:
After:
The text was updated successfully, but these errors were encountered: