A standalone Aleo prover build upon snarkOS and snarkVM, with multi-threading optimization.
It's called "light" because it won't spin up a full node, but instead will only run the prover part.
This prover only supports operators using my modified code as it relies on the custom messaging protocol to work properly.
Install the dependencies:
rust (>= 1.56)
clang
libssl-dev
pkg-config
Run cargo build --release to build the binary.
Please refer to the usage help (target/release/aleo-prover --help):
prover 0.2.6
Standalone prover.
USAGE:
aleo-prover [FLAGS] [OPTIONS] --address <address> --pool <pool>
FLAGS:
-d, --debug Enable debug logging
-h, --help Prints help information
--new-address Generate a new address
-V, --version Prints version information
OPTIONS:
-a, --address <address> Prover address (aleo1...)
-g, --cuda <cuda>... Indexes of GPUs to use (starts from 0)
Specify multiple times to use multiple GPUs
Example: -g 0 -g 1 -g 2
Note: Pure CPU proving will be disabled as each GPU job requires one CPU thread as well
-j, --cuda-jobs <jobs> Parallel jobs per GPU, defaults to 1
Example: -g 0 -g 1 -j 4
The above example will result in 8 jobs in total
-o, --log <log> Output log to file
-p, --pool <pool> Pool address:port
-t, --threads <threads> Number of threads
Use your own address as the prover address, not the pool's address.
You can specify the number of threads to use by --threads option. The prover will use all available threads by default, but it won't use 100% of them all the time. You can try adjusting the number to see if using more threads gives better proof rate.
You can enable debug logging by --debug option.
When starting the prover, there will be a line of log showing the structure of the thread pools.
About the thread pool configuration
The prover will check if the number of threads can be divided by 12, 10 or 8. It will then create multiple thread pools with that number of threads per pool. Otherwise, the prover will create thread pools with 6 threads per pool, using as many threads as possible up to the specified number.
It's recommended to enable debug logging to observe the time needed for each proof.
Before 0.2.3 the prover could use 4 threads per pool in worse case scenario which might make one proof take longer than 10 seconds to generate. This might make the prover completely miss some short-interval blocks, so in 0.2.3 it's been changed to 6 threads per pool.
GPU support is added in version 0.2.0.
To enable GPU support, use cargo build --release --features enable-cuda when building the binary. Obviously you will need to install the CUDA runtime.
Use -g option to enable GPU support. To use multiple GPUs, use -g multiple times with different GPU indexes.
Use -j option to specify the number of jobs per GPU. The default is 1.
Every GPU job will use a CPU thread as well, so it's really a "GPU accelerated prover" instead of a "GPU prover", as only the scalar multiplication on BLS12-377 curve is GPU accelerated.
snarkVM would load programs to all GPUs in the system but the prover will only use the specified GPUs. It wastes some GPU memory, unfortunately.
GPU proving should be slightly faster (~20%).
You can now generate new Aleo addresses by using --new-address option.
You can use domain names when specifying the pool address.
Stopped the prover from sending stale shares to pool. Note that you might see the speed reported by the prover to drop a little. However, the actual valid shares are not affected.
GPU proving should be slightly faster (~5-10%).
Changed the thread pool configuration for CPU proving.
Promoted the thread pool configuration log to info level.
Added log file support.
Added extra check for -g option.
Removed OpenCL dependency.
Added GPU support.
Initial release.
GPL-3.0-or-later