stefandeml / bellman_benchmarks

bellman scaling benchmarks

bellman scaling benchmarks

This repository summarises our analysis on the scaling behaviour of the bellman zkSNARK backend for larger constraint circuits.

We use the blake2s hashing circuit as a basis and conduct runs for several pre-image sizes:

Groth16

NumConstraints	ByteInput	MinMemoryPerExecutorGB	RunTimeParamGenSec	RunTimeProverSec	ParameterMB
64K	192	0.1	9.5	0.46	32
128K	384	0.18	18.95	0.83	63
256K	768	0.27	43.8	2.6	150
512K	1536	0.758	87.6	5.3	300
1M	3072	1.5	174.5	10.57	601
2M	6144	2.9	348.7	20	1200
4M	12288	5.8	696.3	40.4	2400
8M	24576	11.5	1400.8	82.43	4700
16M	49152	23.1	2789.5	167.1	9400

Sonic

NumConstraints	Byte input	MinMemoryPerExecutorGB	RunTimeParamGenSec	RunTimeProverSec	RunTimeCreateAdvice	RunTimeGenAggregate5	RunTimeVerify1Proof	RunTimeVerify5Naive	RunTimeVerify5Aggregated
64K	192	0.5	84	4	1.2	6.5	147ms	525ms	170ms
128K	384	0.9	170	7.9	2.4	12.5	277ms	988ms	303ms
256K	768	1.8	341	14.8	4.2	23.3	539ms	1.98s	563ms
512K	1536	3.6	683	28.9	8.2	42.8	1	3.9	1.1
1M	3072	7.1	1365	57	15.9	83.6	2.1	7.8	2.1
2M	6144	14.3	2735	112	31	162	4.2	15.6	4.2

Some additional runs with varying batch size:

NumConstraints	Byte input	BatchSize	MinMemoryPerExecutorGB	RunTimeParamGenSec	RunTimeProverSec	RunTimeCreateAdvice	RunTimeGenAggregate	RunTimeVerify1	RunTimeVerifyBatch	RunTimeVerifyBatchAggregated
128K	384	3	0.9	170	7.9	2.4	9	285ms	646ms	300ms
128K	384	5	0.9	170	7.9	2.4	12.5	281ms	1	300ms
128K	384	10	0.9	170	7.9	2.4	20.32	281ms	1.87	295ms
128K	384	20	0.9	170	7.9	2.4	36	281ms	3.62	295ms
128K	384	40	0.9	170	7.9	2.4	68.7	281ms	7.1	300ms

We use the following system config to run the benchmark program:

 $ inxi -C
CPU: 6 core Intel Core i5-8400 (-MCP-) cache: 9216 KB
clock speeds: max: 4000 MHz 1: 800 MHz 2: 800 MHz 3: 800 MHz 4: 800 MHz 5: 800 MHz 6: 800 MHz