This repo contains code for the paper, MemMAP: Compact and Generalizable Meta-LSTM Models for Memory Access Prediction. It includes code for experiments on Specialized, Concatenated, MAML-DCLSTM, and C-MAML-DCLSTM, cited as:

@inproceedings{srivastava2020memmap,
  title={MemMAP: Compact and Generalizable Meta-LSTM Models for Memory Access Prediction},
  author={Srivastava, Ajitesh and Wang, Ta-Yang and Zhang, Pengmiao and De Rose, Cesar Augusto F and Kannan, Rajgopal and Prasanna, Viktor K},
  booktitle={Pacific-Asia Conference on Knowledge Discovery and Data Mining},
  pages={57--68},
  year={2020},
  organization={Springer}
}

The trace uses the PARSEC benchmark(https://parsec.cs.princeton.edu/), generated using Pin tool, see example Memory Reference Trace (https://software.intel.com/sites/landingpage/pintool/docs/97503/Pin/html/)

In data folder, two sample raw traces are given to run the scripts.

• python: 3.x
• TensorFlow v1.0+
• Keras v1.0+
• Pytorch: 0.4+
• NVIDIA GPU

The specialized model uses doubly compressed LSTM discribed in paper:

@inproceedings{srivastava2019predicting,
  title={Predicting memory accesses: the road to compact ML-driven prefetcher},
  author={Srivastava, Ajitesh and Lazaris, Angelos and Brooks, Benjamin and Kannan, Rajgopal and Prasanna, Viktor K},
  booktitle={Proceedings of the International Symposium on Memory Systems},
  pages={461--470},
  year={2019}
}

First, cd Specialized,

Then run the script use python3 Specialized.py bodytrack_1_1M.out 20, where the first argument is the trace file name in folder ../data/, the second argument is the training epochs. The length of training and testing sequences are both hardcoded as 200k.

cd ./Concatenated_Rerun

First, run python3 concatenate_data.py [num] to concatenate the traces, where the argument num is the number of traces.

• To run the sample traces directly, set num=2.

• To run all 13 applications, set num=13, the trace list variable TRACE_FILE_NAMES in "Concatenate_data.py" should be modified accordingly.

• The processed training and testing data is saved in folder ./data_combine.

Then, run python3 ./gen_conc_np.py [num] to generate training set and testing set in numpy files, and tokenization in pickles.

python3 ./conc_model_train.py [num] [epoch], where num is the number of traces and epoch is the training epochs.

cd ./Meta-DCLSTM

Run python3 prep_meta.py [len], the argument is the length of a sequence. This script does two thins: 1) converting the raw trace into deltas and save in data_dt folder, and 2) tokenize and binarize the deta traces, save processed files in np_file folder and pk_file folder.

Run python3 meta-DCLSTM.py [len] [num], where len is the sequqnce length, set 200000; num is the number of tasks, for PARSEC benchmark, the task number = application number = 13. If using the given sample, then num=2