Benchmarks

Lotus comes with a benchmarking tool that can be used to test how long each resource-intensive sealing task takes. This guide describes how to install the benchmarking tool, and some basic operations.

Installation

  1. You must have the Lotus repository on your computer. If you do not have an existing copy of the repository, clone it from GitHub:

    git clone https://github.com/filecoin-project/lotus.git ~/lotus
    
    Cloning into '/root/lotus'...
    remote: Enumerating objects: 93, done.
    ...
    Resolving deltas: 100% (51531/51531), done.
    
  2. The lotus binary must be built and within the ~/lotus repository folder. If you just cloned the repository or have misplaced the lotus binary, build the project:

    cd ~/lotus
    make clean all && make install
    
    rm -rf  build/.filecoin-install build/.update-modules  lotus lotus-miner lotus-worker lotus-shed lotus-gateway lotus-seed lotus-pond lotus-townhall lotus-fountain lotus-chainwatch lotus-bench lotus-stats lotus-pcr lotus-health lotus-wallet testground
    make -C extern/filecoin-ffi/ clean
    ...
    install -C ./lotus /usr/local/bin/lotus
    install -C ./lotus-miner /usr/local/bin/lotus-miner
    install -C ./lotus-worker /usr/local/bin/lotus-worker
    
  3. Call make lotus-bench to build the Lotus benchmark binary:

    make lotus-bench
    
    rm -f lotus-bench
    go build -o lotus-bench ./cmd/lotus-bench
    ...
    go run github.com/GeertJohan/go.rice/rice append --exec lotus-bench -i ./build
    

    This will produce a lotus-bench binary in the current folder.

  4. You can now run lotus-bench against your system.

Usage

Use the self-documenting feature of the tool to explore the different commands.

./lotus-bench --help
NAME:
   lotus-bench - Benchmark performance of lotus on your hardware

USAGE:
   lotus-bench [global options] command [command options] [arguments...]

VERSION:
   1.15.4-dev+mainnet

COMMANDS:
   prove    Benchmark a proof computation
   sealing  Benchmark seal and winning post and window post
   simple   Run basic sector operations
   import   Benchmark chain import and validation
   help, h  Shows a list of commands or help for one command

GLOBAL OPTIONS:
   --help, -h     show help (default: false)
   --version, -v  print the version (default: false)

Benchmark

Sealing

Benchmark a sealing computation using lotus-bench sealing [command options] [arguments...]. For example:

./lotus-bench sealing
2022-04-13T18:37:41.141+0200    INFO    lotus-bench	lotus-bench/main.go:103	Starting lotus-bench
...
----
results (v28) SectorSize:(536870912), SectorNumber:(1)
seal: addPiece: 1.815989865s (281.9 MiB/s)
seal: preCommit phase 1: 23.512164789s (21.78 MiB/s)
seal: preCommit phase 2: 11.748902504s (43.58 MiB/s)
seal: commit phase 1: 18.391743ms (27.19 GiB/s)
seal: commit phase 2: 5.698575481s (89.85 MiB/s)
seal: verify: 1.536285ms
unseal: 26.136081682s  (19.59 MiB/s)

generate candidates: 101.562µs (4.808 TiB/s)
compute winning post proof (cold): 665.669815ms
compute winning post proof (hot): 624.667643ms
verify winning post proof (cold): 17.949129ms
verify winning post proof (hot): 1.26867ms

compute window post proof (cold): 446.549819ms
compute window post proof (hot): 450.785891ms
verify window post proof (cold): 8.572963ms
verify window post proof (hot): 1.517569ms

Available options:

OptionDescription
--storage-dir valuePath to the storage directory that will store sectors long term (default: “~/.lotus-bench”).
--sector-size valueSize of the sectors in bytes, i.e. 32GiB (default: “512MiB”).
--no-gpuDisable gpu usage for the benchmark run (default: false).
--miner-addr valuePass miner address (only necessary if using existing sectorbuilder) (default: “t01000”)
--benchmark-existing-sectorbuilder valuepass a directory to run post timings on an existing sectorbuilder
--json-outoutput results in json format (default: false)
--skip-commit2skip the commit2 (snark) portion of the benchmark (default: false)
--skip-unsealskip the unseal portion of the benchmark (default: false)
--save-commit2-input valueSave commit2 input to a file
--num-sectors value(default: 1)
--parallel value(default: 1)
--help, -hshow help (default: false)

Prove

Benchmark a proof computation using lotus-bench prove [command options] [arguments...]. For example:

./lotus-bench prove

Available options:

OptionsDescription
--no-gpuDisable gpu usage for the benchmark run (default: false).
--miner-addr valuePass miner address (only necessary if using existing sectorbuilder) (default: “t01000”).
--help, -hShow help (default: false).

Import

Benchmark chain import and validation using lotus-bench import command [command options] [arguments...]. For example:

./lotus-bench import analyze import.car

Available commands:

CommandDescription
analyzeAnalyze a .car file.
helpShow the help information.

Available options:

OptionDescription
--start-tipset valuestart validation at the given tipset key; in format cid1,cid2,cid3…
--end-tipset valuehalt validation at the given tipset key; in format cid1,cid2,cid3…
--genesis-tipset valuegenesis tipset key; in format cid1,cid2,cid3…
--start-height valuestart validation at given height; beware that chain traversal by height is very slow (default: 0)
--end-height valuehalt validation after given height; beware that chain traversal by height is very slow (default: 0)
--batch-seal-verify-threads valueset the parallelism factor for batch seal verification (default: 4)
--repodir valueset the repo directory for the lotus bench run (defaults to /tmp)
--syscall-cache valueread and write syscall results from datastore
--export-tracesshould we export execution traces (default: true)
--no-importshould we import the chain? if set to true chain has to be previously imported (default: false)
--global-profile(default: true)
--only-import(default: false)
--use-pebble(default: false)
--use-native-badger(default: false)
--car valuepath to CAR file; required for import; on validation, either a CAR path or the –head flag are required
--head valuetipset key of the head, useful when benchmarking validation on an existing chain store, where a CAR is not available; if both –car and –head are provided, –head takes precedence over the CAR root; the format is cid1,cid2,cid3…
--help, -hshow help (default: false)
--version, -vprint the version (default: false)

Single task benchmark

Sometimes you may only want to test the performance of a single task without running through the whole sealing task pipeline. For this, you can use the lotus-bench simple command.

Available options:

OptionDescription
--sector-sizeSpecify the sector-size (supports: 2K, 512MiB, 32GiB, 64GiB) (default: 512MiB)
--miner-addrpass miner address (only necessary if using existing sectorbuilder) (default: “t01000”)

Create sector file

Before we can run sealing benchmarks, we need to create an unsealed sector file. You can specify which sector-size you want to create with the –sector-size flag.

./lotus-bench simple addpiece --sector-size <size> /dev/zero /your/path/unsealed

Note that the unsealed in /your/path/ is the file that will be created by the command, and not a directory.

Together with the performance, the command will output a piece CID and the number of bytes in the created unsealed sector. You will need both of these to perform the PreCommit 1 benchmark.

AddPiece 1m26.991655711s (376.7 MiB/s)
baga6ea4seaqao7s73y24kcutaosvacpdjgfe5pw76ooefnyqw4ynr3d2y6x2mpq 34359738368

Sealing tasks

PreCommit1:

To run a single PreCommit 1, you will need the CID and amount of bytes in the unsealed sector created in the AddPiece step. You will also need to specify where the path to the unsealed file is, and where the sealed sector and the cache file will be placed.

./lotus-bench simple precommit1 --sector-size <size> /your/path/unsealed /your/path/sealed /your/path/cache [piece cid] [piece size]

Together with the performance, the command will create a PreCommit1 encoding that you will need if you want to perform the PreCommit 2 benchmark.

PreCommit1 3h10m38.942613688s (2.865 MiB/s)
eyJfbG90dXNfU2VhbFJhbmRvbW5lc3MiOi[...]==

PreCommit2:

To run a single PreCommit 2 you will need to specify the path of the sealed sector, the cache file, and the PreCommit1 encoding from the previous step.

./lotus-bench simple precommit2 --sector-size <size> /your/path/sealed /your/path/cache [pc1 encoding]

Together with the performance, the command will output the commD and commR that you will need if you want to perform the Commit 1 benchmark.

seal: preCommit phase 2: 14m17.347652736s (38.22 MiB/s)
d:baga6ea4seaqdsvqopmj2soyhujb72jza76t4wpq5fzifvm3ctz47iyytkewnubq r:bagboea4b5abcbztu2gpgzz746m537wntioqm5mjnfay5dwsugfqyshv4zljmnwyb

Commit1:

To run a single Commit1 you will need to specify the path of the sealed sector, the cache file, and the commD and commR output from the PreCommit2 step. You will also need to specify the path for the json output from the Commit1 step.

./lotus-bench simple commit1 --sector-size <size> /your/path/sealed /your/path/cache <commD> <commR> /your/path/c1.json

Together with the performance, the command will save a .json file in the path you specified. This file is needed to perform the Commit2 step.

Commit1 352.760466ms (90.71 GiB/s)

Commit2:

To run a single Commit2 you only need to specify the .json file from the Commit1 step.

./lotus-bench simple commit2 /your/path/c1.json

Together with the performance, the command will output the final proof for the sector.

Commit2 21m45.831527979s (25.09 MiB/s)
proof: 972929647be634d708e071bb0834d28e45[...]==

PoSt tasks

WindowPoSt:

To benchmark a single WindowPoSt you will need to specify the path of the sealed sector, the cache file, and the commR output you got in the PreCommit2 step. Specifying the correct sector number is only needed if you are using an existing sectorbuilder and want to benchmark a real sector. Else you can specify any integer as the sector number, it won´t affect the performance.

./lotus-bench simple window-post --sector-size <size> /your/path/to/sealed-sector-file /your/path/to/sector-cache-folder <commR> [sector num]

Together with the performance, the command will output the proof for the WindowPoSt.

Vanilla 24.765454ms (20.19 GiB/s)
Proof 1.6556735s (309.2 MiB/s)
tfhJc8rPBfUEe/b1GFOPCD9pd[...]=

WinningPoSt:

To benchmark a single WinningPoSt you will need to specify the path of the sealed sector, the cache file, and the commR output you got in the PreCommit2 step. Specifying the correct sector number is only needed if you are using an existing sectorbuilder and want to benchmark a real sector. Else you can specify any integer as the sector number, it won´t affect the performance.

./lotus-bench simple winning-post --sector-size <size> /your/path/to/sealed-sector-file /your/path/to/sector-cache-folder <commR> [sector num]

Together with the performance, the command will output the proof for the WinningPoSt.

Vanilla 58.265476ms (8.581 GiB/s)
Proof 5.05608944s (101.3 MiB/s)
uEK4wgt2qG20ymd0Kee7Z+M[...]=

SnapDeal tasks

Unsealed update sector:

Before we can run any SnapDeal sealing benchmarks, we need to create an new unsealed sector file that we are going to update.

./lotus-bench simple addpiece --sector-size <size> /dev/urandom /your/path/new-unsealed

Note that the new-unsealed in /your/path/ is the file that will be created by the command, and not a directory.

Together with the performance, the command will output a piece CID and the number of bytes in the created new-unsealed sector. You will need both of these to perform the ReplicaUpdate benchmark.

AddPiece 1m26.991655711s (376.7 MiB/s)
baga6ea4seaqmbj3lw5365pwbqfh6pmf2rkc65t2ovobqt6zbtjn2frx4uvkyumy 34359738368

ReplicaUpdate:

To run a single ReplicaUpdate benchmark and create an updated sealed file, you will need to specify the path of the sealed sector, the cache, and the new-unsealed sector. You will also need to specify the path for the updated sector, the updated cache file, and the piece CID and piece size from the add-piece step.

./lotus-bench simple addpiece --sector-size <size> /your/path/sealed /your/path/cache /your/path/sealed /your/path/new-unsealed /your/path/updated /your/path/updated-cache <piece cid> <piece size>

Together with the performance, the command will output the commD and commR of the updated sector that you will need if you want to perform the ProveReplicaUpdate 1 & 2 benchmarks.

ReplicaUpdate 17.796460491s (28.77 MiB/s)
d:baga6ea4seaqgfbyx6wi2zuff5c5njfhrpzrdbv7rgyulnl7ypashh7z62hercha r:bagboea4b5abcazijqwde3greyittndi6qvg6euem6gmc5qt6l6a3t2dbevijlsaa

ProveReplicaUpdate1:

To run a single ProveReplicaUpdate1 benchmark you will need to specify the path of the original sealed sector and its cache file, together with the updated sector and its updated cache file. You will also have to specify the original sectors sector-key (The commR), and the updated sectors commR and commD. Lastly, you will need to specify the path for the json output from this benchmark.

./lotus-bench simple provereplicaupdate1 --sector-size <size> /your/path/sealed /your/path/cache /your/path/updated /your/path/updated-cache <sectorKey> <ReplicaUpdate commR> <ReplicaUpdate commD> /your/path/PRU1.json

Together with the performance, the command will save a .json file in the path you specified. This file is needed to perform the ProveReplicaUpdate 2 step.

ProveReplicaUpdate1 1.480227714s (345.9 MiB/s)

ProveReplicaUpdate2:

To run a single ProveReplicaUpdate2 you need to specify the path for the original sectors sector-key, and the updated sectors commRand commD together with the .json file from the ProveReplicaUpdate 1 step.

./lotus-bench simple provereplicaupdate2 --sector-size <size> <sectorKey> <ReplicaUpdate commR> <ReplicaUpdate commD> /your/path/PRU1.json

Together with the performance, the command will output the final proof for the SnapDeal sector.

ProveReplicaUpdate2 21m45.831527979s (25.09 MiB/s)
p: ixvQ/W+Npy9Q0xQ2/K7hLjeM0if/yPYi1TY[...]==

GPUs

It is recommended to configure and use the CUDA architecture for Lotus.

Testing whether the GPU is used

First, to watch GPU utilization run nvtop in one terminal, then in a separate terminal, run a sealing benchmark to simulate sealing of a sector of small size:

./lotus-bench sealing --sector-size=2KiB

This process uses a small amount of GPU, and generally takes only a couple of minutes to complete. If you do not see any activity in nvtop from lotus during the entire process, it is likely something is misconfigured with your GPU.

Edit this page on GitHub