This repo contains code accompaning the manuscript, "C-MemMAP: Clustering-driven Compact, Adaptable, and Generalizable Meta-LSTM Models for Memory Access Prediction"
- python: 3.x
- TensorFlow v1.0+
- Keras v1.0+
- Pytorch: 0.4+
- NVIDIA GPU
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/)
Each application is rerun for three times and acquires inconsistent configuration traces: T1, T2, and T3.
data_dt folder provides the deferentiated deltas from the raw traces T1, T2, and T3.
First, cd ./Specialized_Rerun
Then run python3 ./Preprocessing.py 200000, where the argument is the length of deltas sequences.
Preprocessing.py tokenize and binarize the sequence for doubly-compress LSTM training and testing. For each application, T1 and T2 delta traces are concatenated as training set and T3 delta trace is the testing set.
python3 Train2Test1.py 200000 20, where argv[1] is the length of sequences and argv[2] is the training epochs.
cd ./Delegated_Model_Clustering
Run the DM clustering use script python3 ./Delegated_Model_Clustering.py
The script uses the DCLSTM models in folder Specialized_rerun_model. The model weights of the DCLSTM models are concatenated, dimension reduced using PCA, and clustered using k-means.
cd ./Concatenated_Rerun
Run python3 ./prep_concac.py 200000, where the argument is the length of deltas sequences.
python3 Train_all_Test_each.py 200000 20, where argv[1] is the length of sequences and argv[2] is the training epochs.
cd ./Meta-DCLSTM
Run python3 ./prep_meta.py 200000, where the argument is the length of deltas sequences.
python3 meta-DCLSTM.py 200000, where the argument is the length of sequences.
Use the clustering result from Delegated Model Clustering, train one Meta-DCLSTM model for each cluster of applications.