gym-cache 1.0.2

gym environment simulating file cache.

gym-cache

OpenAI based Gym environments for training RL caching agent

install it: pip install gym-cache

import it like this:

import gym

gym.make('gym_cache:Cache-v0')

observation space has following variables:

• six tokens (integers)
• file size [kB]
• how full is the cache at that moment

There are two discrete action environments (Cache-v0 and Cache-large-v0) and one continuous action environment (Cache-continuous-v0).

Data extractions and preprocessing

This is a two step procedure:

• extract raw data data/extract_data.py - change PQ, date range
• process raw data data/process_data.py - tokenizes filenames, generates unique fileIDs, sorts by access time.

Processed data should be copied to the directory where actor runs. It is a parque file (.pa) with one dataframe:

• index - access time (sorted)
• six tokens derived from the filename ('1', '2', '3', '4', '5', '6')
• filesize ('kB')
• unique file identifier ('fID')

Rewards

• always negative and correspond to cost to get the file - if it was cached it will be smaller
• files are cached irrespectively from what action actor performed for the file
• cleanup: environment memorizes actions. on cleanup it first deletes files judged not to be needed again (action 0 in discrete environments or smaller values in continues environment). If multiple files have the same action value, LRU one is removed first.

Possible technical implementation in XCache server

• There are additional containers in the pod.
  • environment container
    • recieves gstream pfc, and disk info
    • recalculates new state, reward, tokenizes recieved gstream info.
    • memorizes last state, actors actions for each file
    • triggers cleanup at lower HWM then xcache itself. Loops through memorized paths and removes ones least probably needed.
  • redis db - used by environment container to store actor responses
  • actor container

Miscalenious

To change environments:

• clone github repository
• make changes
• install locally:
  python -m pip install --user -e . or python setup.py bdist_wheel python -m pip install dist\gym_cache-1.0.1-py3-none-any.whl
• to upload to pypi repository

  create %USER%.pypirc file first.

  python setup.py bdist_wheel python -m twine upload dist*