Home

Awesome

Filecoin Proving Subsystem

The Filecoin Proving Subsystem (or FPS) provides the storage proofs required by the Filecoin protocol. It is implemented entirely in Rust, as a series of partially inter-dependent crates – some of which export C bindings to the supported API.

There are currently several different crates:

The dependencies between those crates look like this:

       ┌────────────────────────────────────────────────────────────────────┐
       │                             filecoin-proofs                        │
       └─────┬────────────────────────────┬──────────────┬─────────────┬────┘
             │                            │              │             │
             │                            │              │             │
             │                            │              │             │
┌────────────▼──────────┐     ┌───────────▼──────────┐   │   ┌─────────▼───────────┐
│ storage-proofs-update ├─────▶ storage-proofs-porep │   │   │ storage-proofs-post │
└────────────┬──────────┘     └───────────┬──────────┘   │   └─────────┬───────────┘
             │                            │              │             │
             │                            │              │             │
             │                            │              │             │
       ┌─────▼────────────────────────────▼──────────────▼─────────────▼────┐
       │                           storage-proofs-core                      │
       └────────────────────────────────────────────────────────────────────┘

Things shared between crates, should go into storage-proofs-core. An exception is the storage-proofs-update, which needs the needs the stacked DRG from storage-proofs-porep. All crates are free to use other crates for the workspace like filecoin-hashers or fr32.

Security Audits

The rust-fil-proofs proofs code and the Filecoin Spec has undergone a proofs security audit performed by Sigma Prime and been deemed free of critical or major security issues. In addition to the security review, the document provides the summary of findings, vulnerability classifications, and recommended resolutions. All known issues have been resolved to date in both the code and the specification.

rust-fil-proofs has also undergone a SNARK proofs security audit performed by Dr. Jean-Philippe Aumasson and Antony Vennard and been deemed free of critical or major security issues. In addition to the security analysis, the document provides the audit goals, methodology, functionality descriptions and finally observations on what could be improved. All known issues have been resolved to date.

Design Notes

Earlier in the design process, we considered implementing what has become the FPS in Go – as a wrapper around potentially multiple SNARK circuit libraries. We eventually decided to use bellman – a library developed by Zcash, which supports efficient pedersen hashing inside of SNARKs. Having made that decision, it was natural and efficient to implement the entire subsystem in Rust. We considered the benefits (self-contained codebase, ability to rely on static typing across layers) and costs (developer ramp-up, sometimes unwieldiness of borrow-checker) as part of that larger decision and determined that the overall project benefits (in particular ability to build on Zcash’s work) outweighed the costs.

We also considered whether the FPS should be implemented as a standalone binary accessed from Filecoin nodes either as a single-invocation CLI or as a long-running daemon process. Bundling the FPS as an FFI dependency was chosen for both the simplicity of having a Filecoin node deliverable as a single monolithic binary, and for the (perceived) relative development simplicity of the API implementation.

If at any point it were to become clear that the FFI approach is irredeemably problematic, the option of moving to a standalone FPS remains. However, the majority of technical problems associated with calling from Go into Rust are now solved, even while allowing for a high degree of runtime configurability. Therefore, continuing down the same path we have already invested in, and have begun to reap rewards from, seems likely.

Install and configure Rust

NOTE: If you have installed rust-fil-proofs incidentally, as a submodule of lotus, then you may already have installed Rust.

The instructions below assume you have independently installed rust-fil-proofs in order to test, develop, or experiment with it.

Install Rust using rustup.

Build

NOTE: rust-fil-proofs can only be built for and run on 64-bit platforms; building will panic if the target architecture is not 64-bits.

Before building you will need OpenCL to be installed. On Ubuntu, this can be achieved with apt install ocl-icd-opencl-dev. Other system dependencies such as 'gcc/clang', 'wall' and 'cmake' are also required.

For the multicore sdr feature (enabled by default), you will also need to install the hwloc library. On Ubuntu, this can be achieved with apt install hwloc libhwloc-dev. For other platforms, please see the hwloc-rs Prerequisites section.

> cargo build --release --all

The hwloc dependency is optional and may be disabled. Disabling it will not allow the multicore sdr feature to be used. The fallback is single core replication, which is the default unless specified otherwise.

To disable multicore sdr so that hwloc is not required, you can build proofs like this:

> cargo build --release --all --no-default-features --features opencl

Note that the multicore-sdr feature is omitted from the specified feature list, which removes it from being used by default.

There is experimental support for CUDA behind the cuda feature (disabled by default). You will need to install nvcc. On Ubuntu, this can be achieved with apt install nvidia-cuda-toolkit. To enable CUDA support, you can build proofs like this:

> cargo build --release --all --features cuda

It now builds it with both, CUDA and OpenCL support, CUDA will then be preferred at runtime, but can be disabled with the FIL_PROOFS_GPU_FRAMEWORK environment variable (see more information in the GPU usage section below).

Building for Arm64

In order to build for arm64 the current requirements are

Example for building filecoin-proofs

$ rustup +nightly target add aarch64-unknown-linux-gnu
$ cargo +nightly build -p filecoin-proofs --release --target aarch64-unknown-linux-gnu

Test

> cargo test --all

Benchmarks

The main benchmarking tool is called benchy. benchy has several subcommands, including merkleproofs, prodbench, winning_post, window_post and window_post_fake (uses fake sealing for faster benching). Note that winning_post now has a --fake option for also running fake sealing for faster benching. You can run them with various configuration options, but some examples are below:

> cargo run --release --bin benchy -- merkleproofs --size 2KiB
> cargo run --release --bin benchy -- winning-post --size 2KiB
> cargo run --release --bin benchy -- winning-post --size 2KiB --fake
> cargo run --release --bin benchy -- window-post --size 2KiB
> cargo run --release --bin benchy -- window-post-fake --size 2KiB --fake
> cargo run --release --bin benchy -- prodbench
#
# Synthetic PoRep examples (for both 2KiB test and 32GiB production sector sizes)
#
# Preserve a cache using synthetic-porep (2KiB sectors)
> cargo run --release --bin benchy -- porep --size 2KiB --cache /tmp/cache-2k-synthetic --preserve-cache --api-features synthetic-porep
# Preserve a cache using synthetic-porep (32GiB sectors)
> cargo run --release --bin benchy -- porep --size 32GiB --cache /tmp/cache-32g-synthetic --preserve-cache --api-features synthetic-porep
#
# After a preserved cache is generated, this command tests *only* synthetic proof generation (2KiB sectors)
> cargo run --release --bin benchy -- porep --size 2KiB --cache /tmp/cache-2k-synthetic --api-features synthetic-porep --skip-precommit-phase1 --skip-precommit-phase2 --skip-commit-phase1 --skip-commit-phase2
# After a preserved cache is generated, this command tests *only* synthetic proof generation (32GiB sectors)
> cargo run --release --bin benchy -- porep --size 32GiB --cache /tmp/cache-32g-synthetic --api-features synthetic-porep --skip-precommit-phase1 --skip-precommit-phase2 --skip-commit-phase1 --skip-commit-phase2

Window PoSt Bench usages

The Window PoSt bench can be used a number of ways, some of which are detailed here.

First, you can run the benchmark and preserve the working directory like this:

cargo run --release --bin benchy -- window-post --size 2KiB --cache window-post-2KiB-dir --preserve-cache

Then if you want to run the benchmark again to test commit-phase2, you can quickly run it like this:

cargo run --release --bin benchy -- window-post --size 2KiB --skip-precommit-phase1 --skip-precommit-phase2 --skip-commit-phase1 --cache window-post-2KiB-dir

Alternatively, if you want to test just GPU tree building, you can run it like this:

cargo run --release --bin benchy -- window-post --size 2KiB --skip-precommit-phase1 --skip-commit-phase1 --skip-commit-phase2 --cache window-post-2KiB-dir

Note that some combinations of arguments will cause destructive changes to your cached directory. For larger benchmark sector sizes, it is recommended that once you create an initial cache, that it be saved to an alternate location in the case that it is corrupted by a test run. For example, the following run sequence will be guaranteed to corrupt your cache:

# Do NOT run this sequence.  For illustrative purposes only:
# Generate clean cache
cargo run --release --bin benchy -- window-post --size 2KiB --cache window-post-2KiB-dir --preserve-cache
# Skip all stages except the first
cargo run --release --bin benchy -- window-post --size 2KiB --skip-precommit-phase2 --skip-commit-phase1 --skip-commit-phase2 --cache broken-cache-dir

The reason this fails is because new random piece data is generated (rather than loaded from disk from a previous run) in the first step, and then we attempt to use it in later sealing steps using data from previously preserved run. This cannot work.

There is also a bench called gpu-cpu-test:

> cargo run --release --bin gpu-cpu-test

Some results are displayed at the command line, or alternatively written as JSON files. Logging can be enabled using the RUST_LOG=trace option (see more Logging options in the Logging section below).

Note: On macOS you need gtime (brew install gnu-time), as the built in time command is not enough.

Regression Testing

Within the filecoin-proofs crate there is a regression suite. The idea is to record some generated proofs at various proof release versions, so that future versions/revisions can always ensure that it can properly verify historical proofs as expected.

By default, there is a test that verifies all known regression records that exist within the source tree.

In order to generate a new set of regression records, the feature flag persist-regression-proofs must be used.

When the feature is used and all of the filecoin-proofs tests are run (including the ignored tests), the following files are written to disk:

filecoin-proofs/tests/seal_regression_records.json

Once the new files are generated with a given proof version, they should be renamed appropriately and added to the repository and then referenced for verification during routine testing in the filecoin-proofs/tests/regression.rs source (see the const values at the top and go from there).

Logging

For better logging with backtraces on errors, developers should use expects rather than expect on Result<T, E> and Option<T>.

The crate use log for logging, which by default does not log at all. In order to log output crates like fil_logger can be used.

For example

fn main() {
    fil_logger::init();
}

and then when running the code setting

> RUST_LOG=filecoin_proofs=info

will enable all logging.

For advanced/verbose/debug logging, you can use the code setting

> RUST_LOG=trace

Settings

Further down in this README, various settings are described that can be adjusted by the end-user. These settings are summarized in rust-fil-proofs.config.toml.sample and this configuration file can be used directly if copied to ./rust-fil-proofs.config.toml. Alternatively, each setting can be set by using environment variables of the form "FIL_PROOFS_<setting name here>", in all caps. For example, to set rows_to_discard to the value 2, you would set FIL_PROOFS_ROWS_TO_DISCARD=2 in your environment.

Any configuration setting that is not specified has a reasonable default already chosen.

To verify current environment settings, you can run:

cargo run --bin settings

Parameter File Location

Filecoin proof parameter files are expected to be located in /var/tmp/filecoin-proof-parameters. If they are located in an alternate location, you can point the system to that location using an environment variable

FIL_PROOFS_PARAMETER_CACHE=/path/to/parameters

If you are running a node that is expected to be using production parameters (i.e. the ones specified in the parameters.json file within this repo), you can optionally verify your on-disk parameters using an environment variable

FIL_PROOFS_VERIFY_PRODUCTION_PARAMS=1

By default, this verification is disabled.

Optimizing for either speed or memory during replication

While replicating and generating the Merkle Trees (MT) for the proof at the same time there will always be a time-memory trade-off to consider, we present here strategies to optimize one at the cost of the other.

Speed

One of the most computationally expensive operations during replication (besides the encoding itself) is the generation of the indexes of the (expansion) parents in the Stacked graph, implemented through a Feistel cipher (used as a pseudorandom permutation). To reduce that time we provide a caching mechanism to generate them only once and reuse them throughout replication (across the different layers).

FIL_PROOFS_SDR_PARENTS_CACHE_SIZE=2048

This value is defaulted to 2048 nodes, which is the equivalent of 112KiB of resident memory (where each cached node consists of DEGREE (base + exp = 6 + 8) x 4 byte elements = 56 bytes in length). Given that the cache is now located on disk, it is memory mapped when accessed in window sizes related to this variable. This default was chosen to minimize memory while still allowing efficient access to the cache. If you would like to experiment with alternate sizes, you can modify the environment variable

Increasing this value will increase the amount of resident RAM used.

Lastly, the parent's cache data is located on disk by default in /var/tmp/filecoin-parents. To modify this location, use the environment variable

FIL_PROOFS_PARENT_CACHE=/path/to/parent/cache

Using the above, the cache data would be located at /path/to/parent/cache/filecoin-parents.

Alternatively, use FIL_PROOFS_CACHE_DIR=/path/to/parent/cache, in which the parent cache will be located in $FIL_PROOFS_CACHE_DIR/filecoin-parents. Note that if you're using FIL_PROOFS_CACHE_DIR, it must be set through the environment and cannot be set using the configuration file. This setting has no effect if FIL_PROOFS_PARENT_CACHE is also specified.

If you are concerned about the integrity of your on-disk parent cache files, they can be verified at runtime when accessed for the first time using an environment variable

FIL_PROOFS_VERIFY_CACHE=1

If they are inconsistent (compared to the manifest in storage-proofs/porep/parent-cache.json), they will be automatically re-generated at runtime. If that cache generation fails, it will be reported as an error.

FIL_PROOFS_USE_MULTICORE_SDR

When performing SDR replication (Precommit Phase 1) using only a single core, memory access to fetch a node's parents is a bottleneck. Multicore SDR uses multiple cores (which should be restricted to a single core complex for shared cache) to assemble each nodes parents and perform some prehashing. This setting is not enabled by default but can be activated by setting FIL_PROOFS_USE_MULTICORE_SDR=1.

Best performance will also be achieved when it is possible to lock pages which have been memory-mapped. This can be accomplished by running the process as a non-root user, and increasing the system limit for max locked memory with ulimit -l. Alternatively, the process can be run as root, if its total locked pages will fit inside physical memory. Otherwise, the OOM-killer may be invoked. Two sector size's worth of data (for current and previous layers) must be locked -- along with 56 * FIL_PROOFS_PARENT_CACHE_SIZE bytes for the parent cache.

Default parameters have been tuned to provide good performance on the AMD Ryzen Threadripper 3970x. It may be useful to experiment with these, especially on different hardware. We have made an effort to use sensible heuristics and to ensure reasonable behavior for a range of configurations and hardware, but actual performance or behavior of multicore replication is not yet well tested except on our target. The following settings may be useful, but do expect some failure in the search for good parameters. This might take the form of failed replication (bad proofs), errors during replication, or even potentially crashes if parameters prove pathological. For now, this is an experimental feature, and only the default configuration on default hardware (3970x) is known to work well.

FIL_PROOFS_MULTICORE_SDR_PRODUCERS: This is the number of worker threads loading node parents in parallel. The default is 3 so the producers and main thread together use a full core complex (but no more). FIL_PROOFS_MULTICORE_SDR_PRODUCER_STRIDE: This is the (max) number of nodes for which a producer thread will load parents in each iteration of its loop. The default is128. FIL_PROOFS_MULTICORE_SDR_LOOKAHEAD: This is the size of the lookahead buffer into which node parents are pre-loaded by the producer threads. The default is 800.

GPU Usage

The column hashed tree 'tree_c' can optionally be built using the GPU with noticeable speed-up over the CPU. To activate the GPU for this, use the environment variable

FIL_PROOFS_USE_GPU_COLUMN_BUILDER=1

Similarly, the 'tree_r_last' tree can also be built using the GPU, which provides at least a 2x speed-up over the CPU. To activate the GPU for this, use the environment variable

FIL_PROOFS_USE_GPU_TREE_BUILDER=1

Note that both of these GPU options can and should be enabled if a supported GPU is available.

Advanced GPU Usage

When using the GPU to build 'tree_r_last' (using FIL_PROOFS_USE_GPU_TREE_BUILDER=1), an experimental variable can be tested for local optimization of your hardware.

FIL_PROOFS_MAX_GPU_TREE_BATCH_SIZE=Z

The default batch size value is 700,000 tree nodes.

When using the GPU to build 'tree_c' (using FIL_PROOFS_USE_GPU_COLUMN_BUILDER=1), two experimental variables can be tested for local optimization of your hardware. First, you can set

FIL_PROOFS_MAX_GPU_COLUMN_BATCH_SIZE=X

The default value for this is 400,000, which means that we compile 400,000 columns at once and pass them in batches to the GPU. Each column is a "single node x the number of layers" (e.g. a 32GiB sector has 11 layers, so each column consists of 11 nodes). This value is used as both a reasonable default, but it's also measured that it takes about as much time to compile this size batch as it does for the GPU to consume it (using the 2080ti for testing), which we do in parallel for maximized throughput. Changing this value may exhaust GPU RAM if set too large, or may decrease performance if set too low. This setting is made available for your experimentation during this step.

The second variable that may affect overall 'tree_c' performance is the size of the parallel write buffers when storing the tree data returned from the GPU. This value is set to a reasonable default of 262,144, but you may adjust it as needed if an individual performance benefit can be achieved. To adjust this value, use the environment variable

FIL_PROOFS_COLUMN_WRITE_BATCH_SIZE=Y

Note that this value affects the degree of parallelism used when persisting the column tree to disk, and may exhaust system file descriptors if the limit is not adjusted appropriately (e.g. using ulimit -n). If persisting the tree is failing due to a 'bad file descriptor' error, try adjusting this value to something larger (e.g. 524288, or 1048576). Increasing this value processes larger chunks at once, which results in larger (but fewer) disk writes in parallel.

When the library is built with both CUDA and OpenCL support, you can choose which one to use at run time. Use the environment variable:

FIL_PROOFS_GPU_FRAMEWORK=cuda

You can set it to opencl to use OpenCL instead. The default value is cuda, when you set nothing or any other (invalid) value.

CUDA kernels are compiled and build time. By default, they are built for recent architectures, Turing (sm_75 and Ampere (sm_80, sm_86). This increases the overall build time by several minutes. You can reduce it by compiling it only for the specific architecture you need. For example if you only need the CUDA kernels to work on the Turing architecture, you can set on all dependencies that use CUDA kernels:

BELLMAN_CUDA_NVCC_ARGS="--fatbin --gpu-architecture=sm_75 --generate-code=arch=compute_75,code=sm_75"
NEPTUNE_CUDA_NVCC_ARGS="--fatbin --gpu-architecture=sm_75 --generate-code=arch=compute_75,code=sm_75"

Memory

At the moment the default configuration is set to reduce memory consumption as much as possible so there's not much to do from the user side. We are now storing Merkle trees on disk, which were the main source of memory consumption. You should expect a maximum RSS between 1-2 sector sizes, if you experience peaks beyond that range please report an issue (you can check the max RSS with the /usr/bin/time -v command).

Advanced Storage Tuning

With respect to the 'tree_r_last' cached Merkle Trees persisted on disk, a value is exposed for tuning the amount of storage space required. Cached merkle trees are like normal merkle trees, except we discard some number of rows above the base level. There is a trade-off in discarding too much data, which may result in rebuilding almost the entire tree when it's needed. The other extreme is discarding too few rows, which results in higher utilization of disk space. The default value is chosen to carefully balance this trade-off, but you may tune it as needed for your local hardware configuration. To adjust this value, use the environment variable

FIL_PROOFS_ROWS_TO_DISCARD=N

Note that if you modify this value and seal sectors using it, it CANNOT be modified without updating all previously sealed sectors (or alternatively, discarding all previously sealed sectors). A tool is provided for this conversion, but it's considered an expensive operation and should be carefully planned and completed before restarting any nodes with the new setting. The reason for this is because all 'tree_r_last' trees must be rebuilt from the sealed replica file(s) with the new target value of FIL_PROOFS_ROWS_TO_DISCARD in order to make sure that the system is consistent.

Adjusting this setting is NOT recommended unless you understand the implications of modification.

Generate Documentation

First, navigate to the rust-fil-proofs directory.

> git clone https://github.com/filecoin-project/rust-fil-proofs.git
> cd rust-fil-proofs

For documentation corresponding to the latest source, you should clone rust-fil-proofs yourself.

Now, generate the documentation:

> cargo doc --all --no-deps

View the docs by pointing your browser at: …/rust-fil-proofs/target/doc/proofs/index.html.

API Reference

The FPS is accessed from lotus via FFI calls to its API, which is the union of the APIs of its constituents:

The source of truth defining the FPS APIs is a separate repository of Rust source code. View the source directly:

The above referenced repository contains the consumer facing API and it provides a versioned wrapper around the rust-fil-proofs repository's internal APIs. End users should not be using the internal APIs of rust-fil-proofs directly, as they are subject to change outside of the formal API provided.

To generate the API documentation locally, follow the instructions to generate documentation above. Then navigate to:

Contributing

See Contributing

License

The Filecoin Project is dual-licensed under Apache 2.0 and MIT terms: