Writing large number of metrics to Google Cloud Storage via Cloud Storage FUSE is very slow

Abierto
#6,903 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
38/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Estancado
Stack tecnológico
google-cloud, python

Línea de trabajo

Comienza en tensorboard.summary.writer.event_file_writer.EventFileWriter e inspecciona cómo su escritor de archivos genérico gestiona fs_supports_append. Reproduce la comparación con una ruta de Google Cloud Storage FUSE y, después, verifica que escribir muchos eventos evita el retraso demostrado y mantiene el comportamiento normal del sistema de archivos local; el payload no especifica ningún archivo de prueba.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

This can be trivially reproduced given that you have an access to some Google Cloud project in which you have permissions to run Vertex AI custom training jobs and create Google Cloud Storage (GCS) buckets.

  1. In Vertex AI, submit some fake custom training job (using e.g. python:3.10 Docker image and command: ["sleep", "3600"]) with interactive shell enabled (if you're submitting this job via Python's SDK, simply pass enable_web_access=True to CustomJob's .submit() method).
  2. Once you'd get this job in "Training" state, open two terminals by clicking "Launch web terminal" in its GUI - we will use them for Python and Bash, respectively.
  3. In GCS, create some bucket of your choice and make sure that you can write to it.
  4. In the first terminal, run pip install tensorboard and then launch Python interpreter.
  5. Copy & paste this code to Python's shell, replacing <NAME_OF_YOUR_BUCKET> with the name of the bucket that you've created in step (3):
from time import time

from tensorboard.compat.proto import event_pb2
from tensorboard.compat.proto.summary_pb2 import Summary
from tensorboard.summary.writer.event_file_writer import EventFileWriter

path = "/gcs/<NAME_OF_YOUR_BUCKET>/"

def get_fake_metric(i):
    summary = Summary(value=[Summary.Value(tag=f"val_metric{i}", simple_value=0.123)])
    event = event_pb2.Event(summary=summary)
    event.wall_time = time()
    event.step = 1
    return event

  1. In the second terminal (Bash), cd to /gcs/<NAME_OF_YOUR_BUCKET>/ and make sure that there are no existing event files there.
  2. We will start with the default writer - paste this to Python's shell:
writer = EventFileWriter(path)
  1. Switch to Bash and run the following (we will be using this to monitor the size of event files - at the beginning you should see one such file having 88 bytes):
while true; do date '+%H:%M:%S'; ls -l | grep event; sleep 1; done
  1. Switch to Python and run:
for i in range(200): writer.add_event(get_fake_metric(i))
  1. Switch to Bash and observe how your event file is growing - it should reach slightly above 10 KB. Once it will stop growing, terminate your monitoring loop and note how long it took (for me, it was 2m23s). Delete this file afterwards.
  2. Now, we will repeat this experiment with slightly modified writer. Switch to Python and paste the following:
writer = EventFileWriter(path)
assert writer._general_file_writer.fs_supports_append
writer._general_file_writer.fs_supports_append = False
  1. Switch to Bash and run the same monitoring loop as in step (8).
  2. Switch to Python and run the same loop as in step (9), followed by writer.flush() because otherwise you'd have to wait extra 2 minutes.
  3. Switch to Bash and notice that your event file reached its expected size almost instantly.

Q: What does this experiment show?
A: It shows that EventFileWriter is not aware of Cloud Storage FUSE filesystem, which leads to suboptimal performance (and to make it clear: this monkey-patching of fs_supports_append doesn't mean that this fs doesn't support append - it's just a hack meant to foolEventFileWriter).

Q: But writing events to disk is asynchronous, so even if it is slow, why should we care?
A: Because some training frameworks (e.g. PyTorch Lightning) explicitly flush metrics at the end of the validation loop (which is quite reasonable thing to do, I'd say), and with hundreds of them (600+ in my case), training is stuck for non-negligible time after each validation (7-8 minutes in my case) - and the more frequent the validations, the more badly it accumulates w.r.t. total training time and GPU utilization. And besides, good engineers simply can't stand seeing that writing such relatively small amount of data takes so much time ;)

Lenguaje dominante
TypeScript
Estrellas
7.2k
Forks
1.7k
Merge medio
1 d 13 h
PR fusionados (30 d)
3

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de tensorflow/tensorboard

Todos los issues de tensorflow/tensorboard

Issues similares

Más issues de TypeScript

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.