CPU: AMD Ryzen 7 3800XT 8-Core Processor
Client: geth
##### VKT proof benchmark #####
Setup: **in-memory** tree with 1000000 random key-values...
For 1 random-key values of the tree:
Generating proof took 39ms:
Collected 8 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 0ms
The rest (Multiproof + nits) took 39ms
Serializing proof took 0ms
Deserializing proof took 0ms
For 1000 random-key values of the tree:
Generating proof took 210ms:
Collected 7307 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 107ms
The rest (Multiproof + nits) took 103ms
Serializing proof took 7ms
Deserializing proof took 45ms
For 2000 random-key values of the tree:
Generating proof took 357ms:
Collected 14352 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 189ms
The rest (Multiproof + nits) took 167ms
Serializing proof took 11ms
Deserializing proof took 80ms
For 4000 random-key values of the tree:
Generating proof took 650ms:
Collected 28377 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 343ms
The rest (Multiproof + nits) took 307ms
Serializing proof took 20ms
Deserializing proof took 156ms
For 8000 random-key values of the tree:
Generating proof took 1292ms:
Collected 56284 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 645ms
The rest (Multiproof + nits) took 646ms
Serializing proof took 36ms
Deserializing proof took 309ms
For 16000 random-key values of the tree:
Generating proof took 2706ms:
Collected 111399 polynomials evaluations to prove
Collecting those polys evals (comm, evals, etc) took 1180ms
The rest (Multiproof + nits) took 1526ms
Serializing proof took 68ms
Deserializing proof took 598ms
Interpretation notes:
- A VKT with 1 million nodes is generated (important assumption about depth!)
- For each case, we generate a proof for X random key-values, so this is a bit pessimistic since, in an EVM block execution, there should be some “locality” behavior in the access which reduces the number of openings.
- In each case, it’s detailed
Collected XXXX polynomials evaluations to proveshowing how many openings this proof must prove in each case. - As mentioned in the output, the VKT is in-memory. A client will have to optimize to try making loading nodes from disk to have the least amount of overhead as possible, but probably we need to account for some extra overhead from that.
Extra notes:
- If proofs should also include post-state, then that should be accounted as extra overhead.
In this benchmark, we forget about the VKT and just benchmark creating a multiproof for X number of polynomial openings.
##### Raw polynomials multiproof benchmark #####
For 1 polynomials:
Proving:
Proof serialization duration: 0.11ms
Proof creation duration: 84ms
Duration per step:
Generate challenge r and powers : 0.02ms
Calculate t, g(x) and D : 30.59ms
Calculate h(x) and E : 20.41ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 33.73ms
Verification:
Proof deserialization duration: 0.11ms
Total duration: 5ms
Duration per step:
Generate challenge r and powers : 0.01ms
Calculating helper_scalars r^i/(t-z_i) : 0.01ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.06ms
Compute E : 0.32ms
Compute E-D and verify IPA : 4.71ms
For 1000 polynomials:
Proving:
Proof serialization duration: 0.11ms
Proof creation duration: 50ms
Duration per step:
Generate challenge r and powers : 2.33ms
Calculate t, g(x) and D : 13.89ms
Calculate h(x) and E : 4.11ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 30.27ms
Verification:
Proof deserialization duration: 0.11ms
Total duration: 10ms
Duration per step:
Generate challenge r and powers : 3.47ms
Calculating helper_scalars r^i/(t-z_i) : 0.06ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.05ms
Compute E : 2.81ms
Compute E-D and verify IPA : 3.81ms
For 2000 polynomials:
Proving:
Proof serialization duration: 0.13ms
Proof creation duration: 60ms
Duration per step:
Generate challenge r and powers : 5.52ms
Calculate t, g(x) and D : 19.88ms
Calculate h(x) and E : 4.03ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 31.38ms
Verification:
Proof deserialization duration: 0.13ms
Total duration: 14ms
Duration per step:
Generate challenge r and powers : 5.89ms
Calculating helper_scalars r^i/(t-z_i) : 0.08ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.05ms
Compute E : 4.74ms
Compute E-D and verify IPA : 3.87ms
For 4000 polynomials:
Proving:
Proof serialization duration: 0.11ms
Proof creation duration: 78ms
Duration per step:
Generate challenge r and powers : 11.09ms
Calculate t, g(x) and D : 30.85ms
Calculate h(x) and E : 4.03ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 32.32ms
Verification:
Proof deserialization duration: 0.11ms
Total duration: 21ms
Duration per step:
Generate challenge r and powers : 9.28ms
Calculating helper_scalars r^i/(t-z_i) : 0.17ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.03ms
Compute E : 8.11ms
Compute E-D and verify IPA : 3.72ms
For 8000 polynomials:
Proving:
Proof serialization duration: 0.11ms
Proof creation duration: 113ms
Duration per step:
Generate challenge r and powers : 25.60ms
Calculate t, g(x) and D : 53.76ms
Calculate h(x) and E : 4.03ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 30.55ms
Verification:
Proof deserialization duration: 0.11ms
Total duration: 32ms
Duration per step:
Generate challenge r and powers : 16.05ms
Calculating helper_scalars r^i/(t-z_i) : 0.31ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.03ms
Compute E : 13.42ms
Compute E-D and verify IPA : 2.88ms
For 16000 polynomials:
Proving:
Proof serialization duration: 0.11ms
Proof creation duration: 198ms
Duration per step:
Generate challenge r and powers : 64.94ms
Calculate t, g(x) and D : 99.44ms
Calculate h(x) and E : 4.06ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 30.49ms
Verification:
Proof deserialization duration: 0.11ms
Total duration: 56ms
Duration per step:
Generate challenge r and powers : 32.02ms
Calculating helper_scalars r^i/(t-z_i) : 0.72ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.03ms
Compute E : 21.26ms
Compute E-D and verify IPA : 2.92ms
For 128000 polynomials:
Proving:
Proof serialization duration: 0.12ms
Proof creation duration: 3000ms
Duration per step:
Generate challenge r and powers : 2221.08ms
Calculate t, g(x) and D : 743.75ms
Calculate h(x) and E : 4.04ms
Calculate (h-g)(x) and E-D : 0.00ms
IPA for (h-g)(x) and E-D on t : 31.76ms
Verification:
Proof deserialization duration: 0.12ms
Total duration: 402ms
Duration per step:
Generate challenge r and powers : 260.35ms
Calculating helper_scalars r^i/(t-z_i) : 17.53ms
g_2(t) = SUM y_i*(r^i/(t-z_i))=SUM y_i*helper_scalars : 0.04ms
Compute E : 121.05ms
Compute E-D and verify IPA : 3.59msInterpretation notes:
- The current implementation tries to parallelize a decent amount of work, so core number is quite important.
Other notes:
- It’s planned to run these benchmarks in less powerful CPUs, like in a Rock5B.
- When the new Kaustinen network is working, we can try logging “real-ish” e2e benchmark provings to avoid a bit these synthetic setup. That can account for extra overhead at the geth layer (e.g: node loading from disk if needed).