SuperBench Config File
YAML format configuration file is an efficient method to take full advantage of SuperBench.
You can put it in any place and specify the path to config file through -c /path/to/config.yaml
in sb
CLI.
This document covers schema of SuperBench configuration YAML file. You can learn YAML basics from Learn YAML in Y minutes. SuperBench configuration supports most of the YAML features, including anchors and aliases, merge key, etc.
#
ConventionsHere lists syntax conventions used in this document:
The schema and example are in YAML format.
In YAML mappings which use a colon
:
to markkey: value
pair.The left side of colon is a literal keyword defined in configuration, if it is surrounded by
${}
, like${name}
, then the key is a string that can be defined by user.The right side of colon is a data type, which may be Python built-in types (like
string
,dict
), or a rich structure defined in this document (first character capitalized).The notation
[ datatype ]
indicates a YAML sequence of the mentioned data type. For example,[ string ]
is a list of strings.The notation
|
indicates there are multiple optional data types. For example,string | [ string ]
means either a string or a list of strings is allowed.
#
Configuration SchemaThe configuration file describes all benchmarks running by SuperBench. There will be one or more benchmarks, each benchmark has its own settings and parameters. One benchmark may have one or more modes, which indicates how to run benchmarks in all given machines.
Here is an overview of SuperBench configuration structure:
- Schema
- Example
version: stringsuperbench: enable: string | [ string ] monitor: enable: bool sample_duration: int sample_interval: int var: ${var_name}: dict benchmarks: ${benchmark_name}: Benchmark
version: v0.11superbench: enable: benchmark_1 monitor: enable: false sample_duration: 10 sample_interval: 1 var: var_1: value benchmarks: benchmark_1: enable: true modes: - name: local
version
#
Version of the configuration file.
Lower version sb
CLI may not understand higher version config.
superbench
#
SuperBench configuration for all benchmarks.
superbench.enable
#
Enable which benchmark to run, could be one or multiple benchmarks' name.
If not specified, will use ${benchmark_name}.enable
in each benchmark as default.
- value from: benchmark names defined in
superbench.benchmarks
- default value:
null
superbench.monitor
#
Enable monitor to collect system metrics periodically, currently only support CUDA platform. There are three settings:
enable
#
Whether enable the monitor module or not.
sample_duration
#
Calculate the average metrics during sample_duration seconds, such as CPU usage and NIC bandwidth.
sample_interval
#
Do sampling every sample_interval seconds.
superbench.var
#
User-defined variables to be used in the configuration. Leveraging YAML anchors and aliases, common settings can be defined here to avoid config duplication.
Here is a usage example:
superbench: var: common_param: ¶m num_warmup: 16 num_steps: 128 batch_size: 128 benchmarks: model-benchmarks:foo: models: - resnet50 parameters: *param model-benchmarks:bar: models: - vgg19 parameters: *param
The above configuration equals to the following:
superbench: benchmarks: model-benchmarks:foo: models: - resnet50 parameters: num_warmup: 16 num_steps: 128 batch_size: 128 model-benchmarks:bar: models: - vgg19 parameters: num_warmup: 16 num_steps: 128 batch_size: 128
superbench.benchmarks
#
Mappings of ${benchmark_name}: Benchmark
.
There are three types of benchmarks, micro-benchmark, model-benchmark, and docker-benchmark. Each benchmark has its own unique name listed in docs.
${benchmark_name}
can be one of the followings:
${benchmark_unique_name}
, it can be the exact same as benchmark's own unique name;${benchmark_unique_name}:${annotation}
, or if there's a need to run one benchmark with different settings, an annotation separated by:
can be appended after benchmark's unique name.
See Benchmark
Schema for benchmark definition.
Benchmark
Schema#
Definition for each benchmark, here is an overview of Benchmark
configuration structure:
- Schema
- Example
#
Micro-Benchmark${benchmark_name}: enable: bool timeout: int modes: [ Mode ] frameworks: [ enum ] parameters: run_count: int duration: int log_raw_data: bool ${argument}: bool | str | int | float | list
#
Model-Benchmarkmodel-benchmarks:${annotation}: enable: bool timeout: int modes: [ Mode ] frameworks: [ enum ] models: [ enum ] parameters: run_count: int duration: int log_raw_data: bool num_warmup: int num_steps: int sample_count: int batch_size: int precision: [ enum ] model_action: [ enum ] pin_memory: bool ${argument}: bool | str | int | float | list
#
Micro-Benchmarkkernel-launch: enable: true timeout: 120 modes: - name: local proc_num: 8 prefix: CUDA_VISIBLE_DEVICES={proc_rank} parallel: yes parameters: num_warmup: 100 num_steps: 2000000 interval: 2000
#
Model-Benchmarkmodel-benchmarks:resnet: enable: true timeout: 1800 modes: - name: torch.distributed proc_num: 8 node_num: 1 frameworks: - pytorch models: - resnet50 - resnet101 - resnet152 parameters: duration: 0 num_warmup: 16 num_steps: 128 batch_size: 128 precision: - float32 - float16 model_action: - train
enable
#
Enable current benchmark or not, can be overwritten by superbench.enable
.
- default value:
true
timeout
#
Set the timeout value in seconds, the benchmarking will stop early if timeout is triggered.
- default value: none
modes
#
A list of modes in which the benchmark runs. Currently only one mode is supported for each benchmark.
See Mode
Schema for mode definition.
frameworks
#
A list of frameworks in which the benchmark runs. Some benchmarks can support multiple frameworks while others only support one.
- accepted values:
[ onnxruntime | pytorch | tf1 | tf2 | none ]
- default value:
[ none ]
models
#
A list of models to run, only supported in model-benchmark.
- accepted values:
# pytorch framework[ alexnet | densenet121 | densenet169 | densenet201 | densenet161 | googlenet | inception_v3 | mnasnet0_5 | mnasnet0_75 | mnasnet1_0 | mnasnet1_3 | mobilenet_v2 | resnet18 | resnet34 | resnet50 | resnet101 | resnet152 | resnext50_32x4d | resnext101_32x8d | wide_resnet50_2 | wide_resnet101_2 | shufflenet_v2_x0_5 | shufflenet_v2_x1_0 | shufflenet_v2_x1_5 | shufflenet_v2_x2_0 | squeezenet1_0 | squeezenet1_1 | vgg11 | vgg11_bn | vgg13 | vgg13_bn | vgg16 | vgg16_bn | vgg19_bn | vgg19 | bert-base | bert-large | gpt2-small | gpt2-medium | gpt2-large | gpt2-xl | llama2-7b | llama2-13b | llama2-70b ]
- default value:
[ ]
parameters
#
Parameters for benchmark to use, varying for different benchmarks.
There are four common parameters for all benchmarks:
- run_count: how many times does user want to run this benchmark, default value is 1.
- duration: the elapsed time of benchmark in seconds. It can work for all model-benchmark. But for micro-benchmark, benchmark authors should consume it by themselves.
- log_raw_data: log raw data into file instead of saving it into result object, default value is
False
. Benchmarks who have large raw output may want to set it asTrue
, such asnccl-bw
/rccl-bw
. - log_flushing: real-time log flushing, default value is
False
.
For Model-Benchmark, there are some parameters that can control the elapsed time.
- duration: the elapsed time of benchmark in seconds.
- num_warmup: the number of warmup steps, should be positive integer.
- num_steps: the number of test steps.
If duration > 0
and num_steps > 0
, then benchmark will take the least as the elapsed time. Otherwise only one of them will take effect.
Mode
Schema#
Definition for each benchmark mode, here is an overview of Mode
configuration structure:
- Schema
- Example
name: enumproc_num: intnode_num: intenv: dictmca: dictprefix: strparallel: bool
name: localproc_num: 8prefix: CUDA_VISIBLE_DEVICES={proc_rank}parallel: yes
name
#
Mode name to use. Here lists available modes:
local
: run benchmark as local process.torch.distributed
: launch benchmark through PyTorch DDP, each process will run on one GPU.mpi
: launch benchmark through MPI, the benchmark implementation could leverage MPI interface.
Some attributes may only be suitable for specific mode.
local | torch.distributed | mpi | |
---|---|---|---|
proc_num | ✓ | ✓ | ✓ |
node_num | ✘ | ✓ | ✓ |
prefix | ✓ | ✘ | ✘ |
env | ✓ | ✓ | ✓ |
mca | ✘ | ✘ | ✓ |
parallel | ✓ | ✘ | ✘ |
pattern | ✘ | ✘ | ✓ |
- accepted values:
local | torch.distributed | mpi
- default value:
local
proc_num
#
Process number to run per node. Each process will run an individual benchmark, how processes communicate depends on the mode.
- default value:
1
node_num
#
Node number to run in the mode. Defaults to all nodes provided by host file in the run.
Will be ignored in local
mode.
For example, assuming you are running model benchmark on 4 nodes,
proc_num: 8, node_num: 1
will run 8-GPU distributed training on each node,
while proc_num: 8, node_num: null
will run 32-GPU distributed training on all nodes.
- default value:
null
prefix
#
Command prefix to use in the mode, in Python formatted string.
Available variables in formatted string include:
proc_rank
proc_num
So prefix: CUDA_VISIBLE_DEVICES={proc_rank}
will be expressed as CUDA_VISIBLE_DEVICES=0
, CUDA_VISIBLE_DEVICES=1
, etc.
env
#
Environment variables to use in the mode, in a flatten key-value dictionary. The value needs to be string, any integer or boolean values need to be enclosed in quotes.
Formatted string is also supported in value, available variables include:
proc_rank
proc_num
mca
#
MCA (Modular Component Architecture) frameworks, components, or modules to use in MPI,
in a flatten key-value dictionary.
Only available for mpi
mode.
parallel
#
Whether run benchmarks in parallel (all ranks at the same time) or in sequence (one rank at a time).
Only available for local
mode.
- default value:
yes
pattern
#
Pattern variables to run benchmarks with nodes in specified traffic pattern combination, in a flatten key-value dictionary.
Only available for mpi
mode.
Available variables in formatted string includes:
type(str)
: the traffic pattern type, required.- accepted values:
all-nodes
,pair-wise
,k-batch
,topo-aware
- accepted values:
mpi_pattern(bool)
: generate pattern config file in./output/mpi_pattern.txt
for diagnosis, required.batch(int)
: the scale of batch, required ink-batch
pattern.ibstat(str)
: the path of ibstat output, wil be auto-generated in./output/ibstat_file.txt
if not specified, optional intopo-aware
patternibnetdiscover(str)
: the path of ibnetdiscover outputibnetdiscover_file.txt
, required intopo-aware
pattern.min_dist(int)
: minimum distance of VM pair, required intopo-aware
pattern.max_dist(int)
: maximum distance of VM pair, required intopo-aware
pattern.