ML Goodput Measurement

Overview

ML Goodput Measurement is a library intended to be used with Cloud accelerators to log necessary information and query a job's Goodput and Badput Breakdown. It can be pip installed to import its modules, and retrieve information about a training job's overall productive Goodput and sources of Badput. The package exposes API interfaces to log useful information from the user application and query Goodput for the job run, gain insight into the productivity of ML workloads and utilization of compute resources.

The package also exposes Goodput Monitoring APIs which allow asynchronous query and export of the job's Goodput, Badput and Step Time Deviation to Tensorboard with configurable upload interval.

Components

The ML Goodput Measurement library consists of the following main components:

• GoodputRecorder

• GoodputCalculator

• GoodputMonitor

• GoodputCache

The GoodputRecorder exposes APIs to the client to export key timestamps while a training job makes progress, namely APIs that allow logging of productive step time and total job run time. The library will serialize and store this data in Google Cloud Logging.

The GoodputCalculator exposes APIs to compute Goodput based on the recorded data. Cloud Logging handles its internal operations asynchronously. The recommended way to compute Goodput is to run an analysis program separate from the training application, either on a CPU instance or on the users' development machine.

Under the hood, the GoodputCalculator uses a GoodputCache which is an internal component that locally caches pre-computations and useful logs such that repeated computations can be made inexpensive.

The GoodputMonitor exposes APIs to query and upload goodput and step time deviation data to Tensorboard asynchronously. It does this by instantiating a GoodputCaluclator under the hood.

Installation

To install the ML Goodput Measurement package, run the following command on the VM or machine you want to query or monitor your workload from:

pip install ml-goodput-measurement

Usage

The usage of this package requires the setup of a Google Cloud project with billing enabled to properly use Google Cloud Logging. If you don't have a Google Cloud project, or if you don't have billing enabled for your Google Cloud project, then do the following:

1. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

2. Make sure that billing is enabled for your Google Cloud project. Instructions can be found here

3. Enable the Cloud Logging API.

To run your training on Cloud accelerator, set up the environment by following instructions here.

To learn more about Google Cloud Logging, visit this page.

Access Scopes

You will need both read and write access scopes for cloud logging on both the GPU or TPU and CPU node pools. Full cloud logging access is granted by the following access scope during node pool creation:

• https://www.googleapis.com/auth/cloud-platform

XPK adds this access scope to the GPU, TPU and CPU node pools, so XPK is the recommended method to create clusters and node-pools in you intend to run your workloads on GKE.

Instructions on how to create clusters using XPK can be found here and how to create workloads using XPK can be found here.

NOTE: Access Scopes are immutable and workloads can only be migrated to new node pools with required access scopes. Access scopes on already created clusters cannot be updated.

Import

To use this package, import the goodput module:

from ml_goodput_measurement import goodput
from ml_goodput_measurement import monitoring

Define the name of the Google Cloud Logging logger.

Create a run-specific logger name where Cloud Logging entries can be written to and read from.

For example:

goodput_logger_name = f'goodput_{config.run_name}'

Create a GoodputRecorder object

Next, create a recorder object with the following parameters:

1. job_name: The full run name of the job.
2. logger_name: The name of the Cloud Logging logger object (created in the previous step).
3. logging_enabled: Whether or not this process has Cloud Logging enabled.

NOTE: For a multi-worker setup, please ensure that only one worker writes the logs to avoid the duplication. In JAX, for example, the check could be if jax.process_index() == 0

NOTE: logging_enabled defaults to False and Goodput computations cannot be completed if no logs are ever written.

For example:

goodput_recorder = goodput.GoodputRecorder(job_name=config.run_name, logger_name=goodput_logger_name, logging_enabled=(jax.process_index() == 0))

Record Data with GoodputRecorder

Record Job Start and End Time

Use the recorder object to record the job's overall start and end time.

For example:

def main(argv: Sequence[str]) -> None:
# Initialize configs…
goodput_recorder.record_job_start_time(datetime.datetime.now())
# Device Initialization and device scanning…
# Set up other things for the main training loop…
# Main training loop
train_loop(config)
goodput_recorder.record_job_end_time(datetime.datetime.now())

Record Step Time

Use the recorder object to record a step's start time using record_step_start_time(step_count):

For example:

def train_loop(config, state=None):
# Set up mesh, model, state, checkpoint manager…

# Initialize functional train arguments and model parameters…

# Define the compilation

for step in np.arange(start_step, config.steps):
  goodput_recorder.record_step_start_time(step)
  # Training step…

return state

Record Device Initialization, Training Preparation and Data Loading Time

• Use the recorder object to record Device Initialization time using record_tpu_init_start_time and record_tpu_init_end_time.
• Use the recorder object to record Training Preparation time using record_training_preparation_start_time and record_training_preparation_end_time.
• Use the recorder object to record Data Loading time using record_data_loading_start_time and record_data_loading_end_time.

For example:

def train_loop(config, state=None):
goodput_recorder.record_tpu_init_start_time()
# Set up mesh, model, state, checkpoint manager…
goodput_recorder.record_tpu_init_end_time()
goodput_recorder.record_training_preparation_start_time()
# Set up training set, initialize functional train arguments and model parameters…
# Define the compilation
# Set up any metrics collectors
goodput_recorder.record_training_preparation_end_time()

for step in np.arange(start_step, config.steps):
  goodput_recorder.record_data_loading_start_time()
  example_batch = load_next_batch(data_iterator, example_batch, config)
  goodput_recorder.record_data_loading_end_time()
  goodput_recorder.record_step_start_time(step)
  # Training step…

return state

Retrieve Goodput with GoodputCalculator

In order to retrieve the Goodput of a job run, all you need to do is instantiate a GoodputCalculator object with the job's run name and the Cloud Logging logger name used to record data for that job run. Then call the get_job_goodput API to get the computed Goodput for the job run.

It is recommended to make the get_job_goodput calls for a job run from an instance that runs elsewhere from your training machine.

Create a GoodputCalculator object

Create the calculator object:

goodput_logger_name = f'goodput_{config.run_name}' # You can choose your own logger name.
goodput_calculator = goodput.GoodputCalculator(job_name=config.run_name, logger_name=goodput_logger_name)

If you want to enable Pathways, turn on the using_pathways flag:

goodput_logger_name = f'goodput_{config.run_name}' # You can choose your own logger name.
goodput_calculator = goodput.GoodputCalculator(job_name=config.run_name, logger_name=goodput_logger_name, using_pathways=True)

Retrieve Goodput

Finally, call the get_job_goodput API to retrieve Goodput for the entire job run. This API takes an optional parameter include_badput_breakdown. which defaults to False.

The returned result is a tuple of the job’s Goodput at query-time, a dictionary mapping various sources of Badput and their corresponding percentages and the last recorded step. If include_badput_breakdown is not set, an empty dictionary for Badput is returned.

If you are only interested in Goodput:

total_goodput, _, _ = goodput_calculator.get_job_goodput()
print(f"Total job goodput: {total_goodput:.2f}%")

Retrieve Badput Breakdown

Badput breakdown is dictionary representation of various sources of Badput mapped to its corresponding value. Badput is the percentage of time spent by the job doing work that is not training to the total lifetime of the job. This includes time spent doing Device initialization, training preparation, checkpoint loading, compilation or re-compilation, data loading, checkpoint saving and time lost due to disruptions.

Following Badput Breakdown buckets are supported by the library at this time:

# Supported Badput Types
class BadputType(enum.Enum):
 """The type of Badput."""
  TPU_INITIALIZATION = 1
  TRAINING_PREP = 2
  PROGRAM_STARTUP = 3
  DATA_LOADING = 4
  UNPRODUCTIVE_CHECKPOINT_SAVE_TIME = 5
  UNPRODUCTIVE_CHECKPOINT_RESTORE_TIME = 6
  WASTED_PROGRESS_FROM_DISRUPTION = 7
  OTHER = 8

Badput Breakdown Details

• Accelerator Initialization Time (TPU_INITIALIZATION)

This is the time spent on device discovery, slice initialization, device driver re-initialization and reset, security setup, initialization of pre-mapped buffers and more.

• Training Preparation Time (TRAINING_PREP)

This is the time spent on the creation of checkpoint managers, checkpoint loading, running mesh and model optimizers and more.

• Program Startup Time (PROGRAM_STARTUP)

This is the time spent on framework specific function transformations (such as JAX tracing), compilation tasks, runtime initialization etc.

• Data Loading Time (DATA_LOADING)

This is the time spent on loading each batch of data for the training at a step to continue. This should be a small contribution to Badput if parallel data loading is used.

• Checkpointing Time (UNPRODUCTIVE_CHECKPOINT_SAVE_TIME, UNPRODUCTIVE_CHECKPOINT_RESTORE_TIME)

This is the time spent on saving a checkpoint and restoring a checkpoint.

Depending on the type of checkpointing technology used by the program, there could be unproductive time while saving a checkpoint. When checkpointing is synchronous, the save operation will block training progress until it is complete.

During asynchronous checkpointing, the model parameters or weights have to be transferred from the device memory to the host memory which is a blocking operation on the device. After the transfer, the device can proceed with model training while the CPU saves the checkpoint to storage in the background. The first blocking operation contributes to unproductive checkpoint time.

If auto checkpointing is used, the checkpoint save operation is initiated upon detection of a planned disruption signal. The save operation in type of checkpointing is synchronous resulting in time lost to Badput.

• Wasted Progress due to Disruption (WASTED_PROGRESS_FROM_DISRUPTION)

Based on checkpointing frequency, a disruption may result in time lost in the form of wasted progress, i.e. time that was spent on productive training but lost after restart.

When there is a disruption, Badput is expected to accumulate in each of the following buckets after restart:

• Accelerator Initialization
• Training Preparation
• Program Startup
• Wasted Progress due to Disruption

If you are interested in retrieving Badput Breakdown along with Goodput:

goodput, badput_breakdown, last_step = goodput_calculator.get_job_goodput(include_badput_breakdown=True)
print(f"Last step recorded: {last_step}")
print(f"Goodput: {goodput:.2f}%")
print(f"Badput due to TPU initialization: {badput_breakdown[goodput.BadputType.TPU_INITIALIZATION]:.2f}%")
print(f"Badput due to training preparation: {badput_breakdown[goodput.BadputType.TRAINING_PREP]:.2f}%")
print(f"Badput due to program startup: {badput_breakdown[goodput.BadputType.PROGRAM_STARTUP]:.2f}%")
print(f"Badput due to data loading: {badput_breakdown[goodput.BadputType.DATA_LOADING]:.2f}%")
print(f"Badput due to disruption and wasted progress: {badput_breakdown[goodput.BadputType.WASTED_PROGRESS_FROM_DISRUPTION]:.2f}%")

Interval Query Goodput and Badput

If you are interested in retrieving Goodput and Badput of the workload within a specific window of time, the GoodputCalculator exposes the get_job_goodput_interval API which computes metrics between the start and end of this window.

This API also returns the last step recorded for the job. the total job time in this window and the number of disruptions within the interval window.

IMPORTANT: Use this API if you know the exact window of time within the workload's total run time that you are interested in.

IMPORTANT: Do NOT use this API if your workload has been manually disrupted.

IMPORTANT: Do NOT use this API if you have accidentally re-used a previous run_name.

# Example usage
start_time_str = "2024-12-16 1:05:00"
start_time_utc = convert_pst_to_utc(start_time_str)
end_time_str = "2024-12-17 2:00:00"
end_time_utc = convert_pst_to_utc(end_time_str)
current_goodput, badput_breakdown, last_step, total_time, disruptions = goodput_calculator.get_job_goodput_interval(start_time_utc, end_time_utc)

Monitor Goodput with GoodputMonitor

In order to monitor the Goodput of a job run on Tensorboard, all you need to do is instantiate a GoodputMonitor object with the job's run name, cloud logger name and Goodput monitoring configurations (as described below). Then call the start_goodput_uploader API to asynchronously query and upload measured Goodput to the specified Tensorboard directory.

Create a GoodputMonitor object

Create a GoodputMonitor object with the following parameters:

1. job_name: The full run name of the job.
2. logger_name: The name of the Cloud Logging logger object (created in the previous step).
3. tensorboard_dir: The directory to write TensorBoard data to.
4. upload_interval: The time interval at which to query and upload data to TensorBoard.
5. monitoring_enabled: Whether or not monitoring is enabled. If the application is interested in monitoring Goodput, it should set this value to True. Only one worker should enable monitoring.
6. include_badput_breakdown: Whether to query and upload badput breakdown data to Tensorboard.

NOTE: Please ensure that only one worker enables monitoring of Goodput. In JAX, for example, the check could be if jax.process_index() == 0

For example:

goodput_logger_name = f'goodput_{config.run_name}' # You can choose your own logger name.
goodput_monitoring_enabled = config.monitor_goodput and jax.process_index() == 0 # Check for configs whether or not the enable monitoring.

goodput_monitor = monitoring.GoodputMonitor(
      job_name=config.run_name,
      logger_name=logger_name,
      tensorboard_dir=config.tensorboard_dir,
      upload_interval=config.goodput_upload_interval_seconds,
      monitoring_enabled=True,
      include_badput_breakdown=True,
    )

If you want to enable Pathways, turn on the pathway_enabled flag:

goodput_logger_name = f'goodput_{config.run_name}' # You can choose your own logger name.
goodput_monitoring_enabled = config.monitor_goodput and jax.process_index() == 0 # Check for configs whether or not the enable monitoring.

goodput_monitor = monitoring.GoodputMonitor(
      job_name=config.run_name,
      logger_name=logger_name,
      tensorboard_dir=config.tensorboard_dir,
      upload_interval=config.goodput_upload_interval_seconds,
      monitoring_enabled=True,
      include_badput_breakdown=True,
      pathway_enabled=True
    )

If you want to monitor Step Time Deviation, configure the GoodputMonitor as follows:

goodput_logger_name = f'goodput_{config.run_name}' # You can choose your own logger name.
goodput_monitoring_enabled = config.monitor_goodput and jax.process_index() == 0 # Check for configs whether or not the enable monitoring.

goodput_monitor = monitoring.GoodputMonitor(
      job_name=config.run_name,
      logger_name=logger_name,
      tensorboard_dir=config.tensorboard_dir,
      upload_interval=config.goodput_upload_interval_seconds,
      monitoring_enabled=True,
      include_badput_breakdown=True,
      include_step_deviation=True,
      configured_ideal_step_time=None # Optional, the library will compute ideal step time if it is not provided
    )

Start asynchronous "query and upload" of Goodput

Call the start_goodput_uploader API to spin off a thread which continuously queries and uploads Goodput.

goodput_monitor.start_goodput_uploader()

Start asynchronous "query and upload" of Step Time Deviation

Call the start_step_deviation_uploader API to spin off a thread which continuously queries and uploads step time deviation.

goodput_monitor.start_step_deviation_uploader()

Visualize on Tensorboard

1. Make sure you have tensorboard-plugin-profile, tensorflow and tensorboard packages installed
2. Follow instructions here to start the Tensorboard server