geoffxy / habitat Goto Github PK

🔮 Execution time predictions for deep neural network training iterations across different GPUs.

License: Apache License 2.0

Shell 1.64% Python 70.00% CMake 1.48% C++ 26.17% Cuda 0.11% Dockerfile 0.39% C 0.22%

deep-learning deep-neural-networks neural-networks performance-prediction performance gpu

habitat's Introduction

Habitat: A Runtime-Based Computational Performance Predictor for Deep Neural Network Training

Habitat is a tool that predicts a deep neural network's training iteration execution time on a given GPU. It currently supports PyTorch. To learn more about how Habitat works, please see our research paper.

Running From Source

Currently, the only way to run Habitat is to build it from source. You should use the Docker image provided in this repository to make sure that you can compile the code.

Download the Habitat pre-trained models.
Run extract-models.sh under analyzer to extract and install the pre-trained models.
Run setup.sh under docker/ to build the Habitat container image.
Run start.sh to start a new container. By default, your home directory will be mounted inside the container under ~/home.
Once inside the container, run install-dev.sh under analyzer/ to build and install the Habitat package.
In your scripts, import habitat to get access to Habitat. See experiments/run_experiment.py for an example showing how to use Habitat.

Note: Habitat needs access to your GPU's performance counters, which requires special permissions if you are running with a recent driver (418.43 or later). If you encounter a CUPTI_ERROR_INSUFFICIENT_PRIVILEGES error when running Habitat, please follow the instructions here and in issue #5.

License

The code in this repository is licensed under the Apache 2.0 license (see LICENSE and NOTICE), with the exception of the files mentioned below.

This software contains source code provided by NVIDIA Corporation. These files are:

The code under cpp/external/cupti_profilerhost_util/ (CUPTI sample code)
cpp/src/cuda/cuda_occupancy.h

The code mentioned above is licensed under the NVIDIA Software Development Kit End User License Agreement.

We include the implementations of several deep neural networks under experiments/ for our evaluation. These implementations are copyrighted by their original authors and carry their original licenses. Please see the corresponding README files and license files inside the subdirectories for more information.

Research Paper

Habitat began as a research project in the EcoSystem Group at the University of Toronto. The accompanying research paper will appear in the proceedings of USENIX ATC'21. If you are interested, you can read a preprint of the paper here.

If you use Habitat in your research, please consider citing our paper:

@inproceedings{habitat-yu21,
  author = {Yu, Geoffrey X. and Gao, Yubo and Golikov, Pavel and Pekhimenko,
    Gennady},
  title = {{Habitat: A Runtime-Based Computational Performance Predictor for
    Deep Neural Network Training}},
  booktitle = {{Proceedings of the 2021 USENIX Annual Technical Conference
    (USENIX ATC'21)}},
  year = {2021},
}

habitat's People

Contributors

Stargazers

Watchers

Forkers

yzs-lab subeans baek-sohyeon liayan dxhpc kiruyamomochi machinelearningsystem rajveerb minxin-lai mit10000

habitat's Issues

Is it possible to provide the code generated by different model variants with different hyperparameter configurations

Networks with different kernel size, stride, padding, etc. configurations and their execution time label

Large Prediction Errors

Hi, I am reproducing the experiments in Habitat now. This is an interesting work and it's very convenient to run Habitat and process the results using the following two scripts.

bash habitat/experiments/gather_raw_data.sh  <target_device>
bash habitat/experiments/process_raw_data.sh

Due to the limitation of GPU resources, I can not access all GPU models listed in the paper and only test it on V100, P100 and T4. But the prediction error is quite large, compared to that shown in the paper. You can check the results here.

Basically, the setting I used follows habitat/docker/Dockerfile. Here are some of my experiment settings that may be different from yours:

CUDA driver version: 455.32.00,
I do not mount the user account on the host machine into the container

So,

Is there any hyper-parameter I need to tune to get a better prediction error ?
2.Can you share the cross-GPU prediction error between each pair of GPUs or just the output of habitat/experiments/process_raw_data.sh? Fig 3 in the paper only shows the results "averaged across all other “origin” GPUs".
Will the setting differences listed above affect the prediction error ? Or any other possible reasons ?

Thanks.

CUPTI_ERROR_INSUFFICIENT_PRIVILEGES in container

The default configuration on my OS and current directions in README may lead to a CUPTI_ERROR_INSUFFICIENT_PRIVILEGES when using CUPTI inside the container.

The example log is attached below:

/home/ubuntu/home/habitat/cpp/src/cuda/cupti_tracer.cpp:120: error: function cuptiActivityRegisterCallbacks(cuptiBufferRequested, cuptiBufferCompleted) failed with error CUPTI_ERROR_INSUFFICIENT_PRIVILEGES.
Traceback (most recent call last):
  File "run_experiment.py", line 246, in <module>
    main()
  File "run_experiment.py", line 238, in main
    run_dcgan_experiments(context)
  File "run_experiment.py", line 155, in run_dcgan_experiments
    context,
  File "run_experiment.py", line 85, in run_experiment_config
    threshold = compute_threshold(runnable, context)
  File "run_experiment.py", line 66, in compute_threshold
    runnable()
  File "run_experiment.py", line 150, in runnable
    iteration(*inputs)
  File "/home/ubuntu/home/habitat/experiments/dcgan/entry_point.py", line 41, in iteration
    netD.zero_grad()
  File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 1098, in zero_grad
    p.grad.detach_()
  File "/home/ubuntu/home/habitat/analyzer/habitat/tracking/operation.py", line 62, in hook
    kwargs,
  File "/home/ubuntu/home/habitat/analyzer/habitat/profiling/operation.py", line 45, in measure_operation
    record_kernels,
  File "/home/ubuntu/home/habitat/analyzer/habitat/profiling/operation.py", line 164, in _to_run_time_measurement
    if record_kernels else []
  File "/home/ubuntu/home/habitat/analyzer/habitat/profiling/kernel.py", line 34, in measure_kernels
    self._measure_kernels_raw(runnable, fname)
  File "/home/ubuntu/home/habitat/analyzer/habitat/profiling/kernel.py", line 48, in _measure_kernels_raw
    time_kernels = hc.profile(runnable)
RuntimeError: CUPTI_ERROR_INSUFFICIENT_PRIVILEGES

My solution:
Adding options nvidia "NVreg_RestrictProfilingToAdminUsers=0" to /etc/modprobe.d/nvidia-kernel-common.conf
and reboot.

Ref:

Broken pillow for torchvision in Dockerfile causes docker build failed

Currently, pip3 install torchvision==0.5.0 should fail due to the broken dependency of pillow, shown in the following CI building process:

https://github.com/yzs-lab/habitat/runs/4311964953?check_suite_focus=true#step:3:915

Corresponding logs are attached below:

The headers or library files could not be found for zlib,
    a required dependency when compiling Pillow from source.
    
    Please see the install instructions at:
       https://pillow.readthedocs.io/en/latest/installation.html
    
    Traceback (most recent call last):
      File "/tmp/pip-build-c0iq5ua_/pillow/setup.py", line 1024, in <module>
        zip_safe=not (debug_build() or PLATFORM_MINGW),
      File "/usr/lib/python3/dist-packages/setuptools/__init__.py", line 129, in setup
        return distutils.core.setup(**attrs)
      File "/usr/lib/python3.6/distutils/core.py", line 148, in setup
        dist.run_commands()
      File "/usr/lib/python3.6/distutils/dist.py", line 955, in run_commands
        self.run_command(cmd)
      File "/usr/lib/python3.6/distutils/dist.py", line 974, in run_command
        cmd_obj.run()
      File "/usr/lib/python3/dist-packages/setuptools/command/install.py", line 61, in run
        return orig.install.run(self)
      File "/usr/lib/python3.6/distutils/command/install.py", line 589, in run
        self.run_command('build')
      File "/usr/lib/python3.6/distutils/cmd.py", line 313, in run_command
        self.distribution.run_command(command)
      File "/usr/lib/python3.6/distutils/dist.py", line 974, in run_command
        cmd_obj.run()
      File "/usr/lib/python3.6/distutils/command/build.py", line 135, in run
        self.run_command(cmd_name)
      File "/usr/lib/python3.6/distutils/cmd.py", line 313, in run_command
        self.distribution.run_command(command)
      File "/usr/lib/python3.6/distutils/dist.py", line 974, in run_command
        cmd_obj.run()
      File "/usr/lib/python3/dist-packages/setuptools/command/build_ext.py", line 78, in run
        _build_ext.run(self)
      File "/usr/lib/python3.6/distutils/command/build_ext.py", line 339, in run
        self.build_extensions()
      File "/tmp/pip-build-c0iq5ua_/pillow/setup.py", line 790, in build_extensions
        raise RequiredDependencyException(f)
    __main__.RequiredDependencyException: zlib
    
    During handling of the above exception, another exception occurred:
    
    Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/tmp/pip-build-c0iq5ua_/pillow/setup.py", line 1037, in <module>
        raise RequiredDependencyException(msg)
    __main__.RequiredDependencyException:
    
    The headers or library files could not be found for zlib,
    a required dependency when compiling Pillow from source.
    
    Please see the install instructions at:
       https://pillow.readthedocs.io/en/latest/installation.html
    
    
    
    ----------------------------------------
Command "/usr/bin/python3 -u -c "import setuptools, tokenize;__file__='/tmp/pip-build-c0iq5ua_/pillow/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, __file__, 'exec'))" install --record /tmp/pip-8eakyb7g-record/install-record.txt --single-version-externally-managed --compile" failed with error code 1 in /tmp/pip-build-c0iq5ua_/pillow/
The command '/bin/sh -c pip3 install   torch==1.4.0   torchvision==0.5.0   pandas==1.1.2   tqdm==4.49.0' returned a non-zero code: 1

How can I make habitat support more devices

Hi, Habitat greatly attracts me. I want to use it to do more predictions. How can I make it support other GPUs?

How Habitat measures the execution time associated with the operation’s backward pass?

Hi! Thanks for your perfect job.

It's easy to understand to measure the execution time in the forward pass.
But in the backward pass, how Habitat does? I think it is an undoubtedly different processor, right?

@geoffxy
Hope for your reply soon!

Fail to build the image

Hi, I am following the steps here to reproduce habitat. When running setup.sh to build the image, the following error occurs

Step 14/19 : RUN gpg --keyserver ha.pool.sks-keyservers.net --recv-keys B42F6819007F00F88E364FD4036A9C25BF357DD4
---> Running in d42ae3b13a05
gpg: WARNING: unsafe permissions on homedir '/root/.gnupg'
gpg: keybox '/root/.gnupg/pubring.kbx' created
gpg: keyserver receive failed: No name
The command '/bin/sh -c gpg --keyserver ha.pool.sks-keyservers.net --recv-keys B42F6819007F00F88E364FD4036A9C25BF357DD4' returned a non-zero code: 2

Does it mean the keyserver ha.pool.sks-keyservers.net is not accessible now?

I wonder whether it is necessary to duplicate the user account on the host machine into the container. With a root account in the container, I can access everything mounted from the host machine. What problem does it cause?

Looking forward to your reply. Thanks.

error: function cuptiProfilerBeginSession(&begin_session_params) failed with error CUPTI_ERROR_UNKNOWN

Hi @geoffxy,

Great work here. I am quite interested in your project and try to reproduce from my side. Hower hit the error in the titel, I suspect that it may be caused by incompetible between CUPTI and NVIDIA driver version, I am wondering if could share you experiment setup here, mostly the host side, are you still using 18.04, what the nvidia driver version, did you use nvidia-docker2 or nvidia-container-runtime? what is your docker version?

As mine, I am using 18.04 as host, driver 470.103.01, nvidia-docker2, docker 20.10.12.

Thanks,
Liang

I wonder what the meaning of varing kernel is.

Hi I am reading Habitat research paper.

I wonder what the meaning of varing kernel is.
I thought the GPU kernel is a collection of instructions that run in parallel, is that right?

Can you give me an example of this phrase?
'some DNN operations are implemented using different GPU kernels on different GPUs '

Thank you for taking the time to read.

CMake Error at CMakeLists.txt:22 (pybind11_add_module):

when running "install-dev.sh", hit below error:

CMake Error at CMakeLists.txt:22 (pybind11_add_module):
Unknown CMake command "pybind11_add_module".

-- Configuring incomplete, errors occurred!