This project is inspired by PARROT by Google Research, simulating and evaluating the model on the Wikipedia CDN trace.

Dataset

The simulation uses The Wikipedia CDN trace dataset in 2018 and 2019 provided by Evan et al, which is collected with the support of the Wikimedia Foundation and is publicly released on sunnyszy/lrb.

ML architecture

The model architecture is based on the attention mechanism.

Training model

Train the model to capture access patterns from the given dataset, where the model checkpoint file will be saved in every training.save_frequency learning step.

python main.py

Configuration

You can configure the model hyperparameters and the training process, including learning rates or batch size. configuration.py includes the below fields.

Field	Description
experiment.base_dir	The base directory where experiment-related files and results will be saved.
experiment.name	A name or identifier for the experiment.
dataset.filepath	The filepath to the dataset used for training and evaluation.
dataset.window_size	The size of the sliding window used to extract subsequences from the dataset.
dataset.capacity	The maximum capacity of the cache used in the simulation.
dataset.access_history_len	The length of the access history used in the cache policy model.
dataset.scorer_type	The type of scorer used in the cache policy model. (belady, lru, learned, random, mixture)
model.obj_id_embedder.type	The type of object ID embedder used in the cache policy model. (dynamic_vocab, logarithmic, positional)
model.obj_id_embedder.max_vocab_size	The maximum vocabulary size for the object ID embedder.
model.obj_id_embedder.embedding_dim	The dimension of embeddings for object IDs.
model.obj_size_embedder.type	The type of object size embedder used in the cache policy model. (dynamic_vocab, logarithmic, positional)
model.obj_size_embedder.embedding_dim	The dimension of embeddings for object sizes.
model.obj_size_embedder.max_size	The maximum object size considered in the model.
model.obj_size_embedder.max_vocab_size	The maximum vocabulary size for the object size embedder.
model.cache_lines_embedder	The type of cache lines embedder used in the cache policy model.
model.positional_embedder.type	The type of positional embedder used in the cache policy model.
model.positional_embedder.embedding_dim	The dimension of embeddings for positional information.
model.lstm_hidden_size	The size of the LSTM hidden state in the cache policy model.
model.max_attention_history	The maximum number of attention history steps considered in the model.
dagger_schedule.type	The type of schedule used for DAgger updates during training. (linear, constant)
dagger_schedule.initial	The initial DAgger rate.
dagger_schedule.final	The final DAgger rate.
dagger_schedule.num_steps	The total number of training steps.
dagger_schedule.update_frequency	The frequency of DAgger updates during training.
training.learning_rate	The learning rate used for training the cache policy model.
training.batch_size	The batch size used in training.
training.sequence_length	The length of training sequences used in training.
training.collection_multiplier	The multiplier used to determine the number of training sequences collected.
training.total_steps	The total number of training steps.
training.save_frequency	The frequency at which model checkpoints are saved.
training.evaluation_frequency	The frequency at which model evaluation is performed.
training.evaluation_size	The size of the evaluation dataset.
training.log_loss_frequency	The frequency at which training loss is logged.
training.checkpoint_dir	The directory where model checkpoints are saved.

Evaluation

Load the trained model from checkpoint files, and evaluate cache hit rates on the validation dataset.

python evaluate.py

External Code and Licenses

This project includes code from the Google Research project, which is provided under the Apache License 2.0. Any modifications made to the original code are documented in the respective files.

• Original Code Repository: google-research/cache_replacement
• License: Apache License 2.0