DistSparseDNN

Distributed Sparse Deep Neural Network Inference

Dependencies

• C++17 or higher
• Pthread, OpenMP, and NUMActl (optional)
• Python3 and TensorFlow2 (only for creating customized sparse DNN)

Install

make

Uninstall

make clean

Test

bin/./app -i <input_ninstances input_nfeatures ncategories input_path> 
               -n <nneurons nmax_layers layers_path> 
               -c <input_compression_type[0-4] layer_compression_type[0-4]>
               -p <parallelism_type[0-4]>
               -h <hashing_type[0-3]>

Supported Features

Compression Types

0. Uncompressed Dense Column (UDC)
1. Compressed Sparse Column (CSC)
2. Doubly Compressed Sparse Column (DCSC) (not implemented yet)
3. Triply Compressed Sparse Column (TCSC) (not implemented yet)
4. Compressed Sparse Row (CSR)

Parallelism Types

0. Model parallelism (supported UDC and CSC)
1. Data parallelism (supported UDC, CSC, and CSR)
2. Data-then-model parallelism (supported CSC and CSR)
3. Manage-worker parallelism (supported CSC and CSR)
4. Work-stealing parallelism (supported CSC and CSR)

	Col-major Formats				Row-major Formats
Parallelism	UDC	CSC	DCSC	TCSC	UDR	CSR	DCSR	TCSR
Data	Yes	Yes				Yes
Model	Yes	Yes
Data-then-model		Yes				Yes
Manager-worker		Yes				Yes
Work-stealing		Yes				Yes

Hashing Types

0. No hashing
1. Input hashing
2. Layer hashing
3. Input and layer hashing

	Input		Layer
Hashing Type	Rows	Cols	Rows	Cols
None	No	No	No	No
Input	Yes	No	No	No
Layer	No	Yes	Yes	Yes
Input & Layer	Yes	Yes	Yes	Yes

Supported Multiplication Types

0. Dense matrix by dense matrix
1. Dense matrix by compressed matrix
2. Compressed matrix by dense matrix
3. Compressed matrix by compressed matrix
4. Compressed matrix by doubly compressed matrix (not implemented yet)
5. Compressed matrix by triply compressed matrix (not implemented yet)

	Col-major Formats		Row-major Formats
Multiplication Type (Input x Layer)	Input Format	Layer Format	Input Format	Layer Format
Dense x Dense	UDC	UDC
Dense x Compressed	UDC	CSC
Compressed x Dense	CSC	UDC
Compressed x Compressed	CSC	CSC	CSR	CSR
Dense x Doubly Compressed
Dense x Triply Compressed
Compressed x Doubly Compressed
Compressed x Triply Compressed

Datasets

• Radixnet Sparse DNN: Download the MNIST and DNN files from http://graphchallenge.mit.edu/data-sets
• Under Development Customized Sparse DNNs: Generate the sparse network using the sparse dnn generator program located under scripts/sparse_dnn_generator.py
  • Currently supported input datasets are MNIST, fashion MNIST, CIFAR-10, CIFAR-100, and IMDB.

Example Commands

For Radixnet Sparse DNN, e.g., for the smallest DNN use

mpirun -np 4 bin/./radixnet -i 60000 1024 0 data/radixnet/bin/MNIST -n 1024 120 data/radixnet/bin/DNN -c 1 1 -p 0 -h 3

For other datasets e.g., for inferring fashion MNIST on a sparse DNN with 30 layers each with 2048 neurons use

python3 scripts/sparse_dnn_generator.py fashion_mnist # Parameters are hardcoded
mpirun -np 4 bin/./fashion_mnist -i 60000 784 10 data/fashion_mnist/bin/ -n 2048 30 data/fashion_mnist/bin/ -c 1 1 -p 0 -h 3

Papers

• Mohammad Hasanzadeh Mofrad, Rami Melhem, Yousuf Ahmad and Mohammad Hammoud. “Accelerating Distributed Inference of Sparse Deep Neural Networks via Mitigating the Straggler Effect.” In proceedings of IEEE High Performance Extreme Computing (HPEC), Waltham, MA USA, 2020.
• Mohammad Hasanzadeh Mofrad, Rami Melhem, Yousuf Ahmad and Mohammad Hammoud. “Studying the Effects of Hashing of Sparse Deep Neural Networks on Data and Model Parallelisms.” In proceedings of IEEE High Performance Extreme Computing (HPEC), Waltham, MA USA, 2020.

Contact

Mohammad Hasnzadeh Mofrad ([email protected])