Skip to content

Latest commit

 

History

History
284 lines (169 loc) · 10.6 KB

README.md

File metadata and controls

284 lines (169 loc) · 10.6 KB
Raw

Unit Tests Formatting Static Analysis

A Jupyter Kernel for Performance Engineering

This is the JUmPER Kernel that enables you to:

  1. Monitor Jupyter cells and measure system metrics like cpu, gpu, I/O or memory utilization.

  2. Instrument and trace or profile Jupyter cells with Score-P.

For binding to Score-P, the kernel uses the Score-P Python bindings. Monitoring does not rely on Score-P and you can use it without a Score-P installation.

Table of Content

Installation

To install the kernel and required dependencies for supporting the monitoring features:

pip install jumper-kernel
python -m jumper.install

You can also build the kernel from source via:

pip install .

The kernel will then be installed in your active python environment. You can select the kernel in Jupyter as jumper.

For using the Score-P features of the kernel you need a proper Score-P installation. Note: this is not required for the monitoring feature of system metrics.

pip install scorep

From the Score-P Python bindings:

You need at least Score-P 5.0, build with --enable-shared and the gcc compiler plugin. Please make sure that scorep-config is in your PATH variable. For Ubuntu LTS systems there is a non-official ppa of Score-P available: https://launchpad.net/~andreasgocht/+archive/ubuntu/scorep .

Usage

Monitoring

Every cell that is executed will be monitored by a parallel running process that collects system metrics for CPU, Memory, IO and if available GPU. Besides that, Jumper forwards the execution of that code to the default Python kernel.

The frequency for performance monitoring can be set via the JUMPER_REPORT_FREQUENCYenvironment variable.

%env JUMPER_REPORT_FREQUENCY=2

Additionally, the number of reports required to store performance data can be defined by the JUMPER_REPORTS_MIN environment variable.

%env JUMPER_REPORTS_MIN=2

The performance data is recorded in-memory and the kernel provides several magic commands to display and interact with the data:

%%display_code_history

Shows the history of the code of monitored cells with index and timestamp.

%%display_code_for_index

Shows the code for the cell of the selected index.

%%display_graph_for_last

Shows the performance display for the last monitored cell.

%%display_graph_for_index [index]

Shows the performance display for the cell of the selected index.

%%display_graph_for_all

Shows the accumulated performance display for all monitored cells.

%%perfdata_to_variable [varname]

Exports the performance data to a variable

%%perfdata_to_json [filename]

Exports the performance data and the code to json files.

Score-P Instrumentation

Configuring Score-P in Jupyter

Set up your Score-P environment with %env line magic, e.g.:

%env SCOREP_ENABLE_TRACING=1
%env SCOREP_TOTAL_MEMORY=3g

For a documentation of Score-P environment variables, see: Score-P Measurement Configuration.

%%scorep_python_binding_arguments

Set the Score-P Python bindings arguments. For a documentation of arguments, see Score-P Python bindings.

%%marshalling_settings

Set marshaller/serializer used for persistence and mode of communicating persistence between notebook and subprocess. Currently tested marshallers: dill, cloudpickle, parallel_marshall; modes of communication: disk, memory. If no arguments were provided, will print current configuration. Use:

%%marshalling_settings
MARSHALLER=[dill,cloudpickle]
MODE=[disk,memory]

When using persistence in disk mode, user can also define directory to which serializer output will be saved with SCOREP_KERNEL_PERSISTENCE_DIR environment variable.

%%execute_with_scorep

Executes a cell with Score-P, i.e. it calls python -m scorep <cell code>

Multi-Cell Mode

You can also treat multiple cells as one single cell by using the multi cell mode. Therefore you can mark the cells in the order you wish to execute them.

%%enable_multicellmode

Enables the multi-cell mode and starts the marking process. Subsequently, "running" cells will not execute them but mark them for execution after %%finalize_multicellmode.

%%finalize_multicellmode

Stop the marking process and executes all the marked cells. All the marked cells will be executed with Score-P.

%%abort_multicellmode

Stops the marking process, without executing the cells.

Hints:

  • The %%execute_with_scorep command has no effect in the multi cell mode.

  • There is no "unmark" command available but you can abort the multicellmode by the %%abort_multicellmode command. Start your marking process again if you have marked your cells in the wrong order.

  • The %%enable_multicellmode, %%finalize_multicellmode and %%abort_multicellmode commands should be run in an exclusive cell. Additional code in the cell will be ignored.

Write Mode

Analogous to %%writefile command in IPykernel, you can convert a set of cells to the Python script which is to be executed with Score-P Python bindings (with settings and environment described in auxillary bash script).

%%start_writefile [scriptname]

Enables the write mode and starts the marking process. Subsequently, "running" cells will not execute them but mark them for writing into a python file after %%end_writefile. scriptname is jupyter_to_script.py by default.

%%end_writefile

Stops the marking process and writes the marked cells in a Python script. Additionally, a bash script will be created for setting the Score-P environment variables, Pyhton bindings arguments and executing the Python script.

Hints:

  • Recording a cell containing %%scorep_python_binding_arguments will add the Score-P Python bindings to the bash script.

  • Code of a cell which is not to be executed with Score-P (not inside the multicell mode and without %%execute_with_scorep) will be framed with with scorep.instrumenter.disable() in the Python script to prevent instrumentation.

  • Other cells will be recorded without any changes, except for dropping all magic commands.

  • %%abort_multicellmode will be ignored in the write mode and will not unmark previous cells from instrumentation.

Presentation of Performance Data

For the monitoring data, use the build-in magic commands or build your own visualizations after exporting the data to a variable or json via the introduced magic commands.

To inspect the Score-P collected performance data, use tools as Vampir (Trace) or Cube (Profile).

Limitations

Serialization Type Support

For the execution of a cell, the kernel uses the default IPython kernel. For a cell with Score-P it starts a new Python process. Before starting this process, the state of the previous executed cells is persisted using dill (https://github.com/uqfoundation/dill) or cloudpickle (https://github.com/cloudpipe/cloudpickle/releases). However:

dill cannot yet pickle these standard types: frame, generator, traceback

Similar yields for cloudpickle. Use the %%marshalling_settings magic command to switch between both depending on your needs.

Overhead

When dealing with big data structures, there might be a big runtime overhead at the beginning and the end of a Score-P cell. This is due to additional data saving and loading processes for persistency in the background. However this does not affect the actual user code and the Score-P measurements.

Future Work

The kernel is still under development. The following is on the agenda:

  • Provide perfmonitors for multi node setups
  • Config for default perfmonitor to define collected metrics

PRs are welcome.

Citing

If you publish some work using the kernel, we would appreciate if you cite the following paper:

Werner, E., Manjunath, L., Frenzel, J., & Torge, S. (2021, October).
Bridging between Data Science and Performance Analysis: Tracing of Jupyter Notebooks.
In The First International Conference on AI-ML-Systems (pp. 1-7).
https://dl.acm.org/doi/abs/10.1145/3486001.3486249

Additionally, please refer to the Score-P Python bindings, published here:

Gocht A., Schöne R., Frenzel J. (2021)
Advanced Python Performance Monitoring with Score-P.
In: Mix H., Niethammer C., Zhou H., Nagel W.E., Resch M.M. (eds) Tools for High Performance Computing 2018 / 2019. Springer, Cham.
https://doi.org/10.1007/978-3-030-66057-4_14

or

Gocht-Zech A., Grund A. and Schöne R. (2021)
Controlling the Runtime Overhead of Python Monitoring with Selective Instrumentation
In: 2021 IEEE/ACM International Workshop on Programming and Performance Visualization Tools (ProTools)
https://doi.org/10.1109/ProTools54808.2021.00008

Contact

[email protected]

Acknowledgments

This work was supported by the German Federal Ministry of Education and Research (BMBF, SCADS22B) and the Saxon State Ministry for Science, Culture and Tourism (SMWK) by funding the competence center for Big Data and AI "ScaDS.AI Dresden/Leipzig