Two papers from ICSA researchers accepted to be presented at IISWC'19
2019 IEEE International Symposium on Workload Characterization
2019 IISWC takes place in Orlando, Florida, USA on November 3-5, 2019.
In their paper "A Closer Look at Lightweight Graph Reordering", Priyank Faldu and Boris Grot along with a collaborator from Oracle Labs characterize lightweight graph reordering techniques that improve cache locality. Their work finds that structure preservation should be considered a first order objective in order to achieve high performance from vertex reordering for real world graphs. Based on this observation they propose a novel lightweight technique, which employs a coarse-grain reordering to largely preserve graph structure while reducing the cache footprint of hot vertices.
The other paper titled "Performance Aware Convolutional Neural Network Channel Pruning for Embedded GPUs" by V. Radu et al. exposes the pitfall of neural network channel pruning (a popular neural network compression technique to make models suitable for running on smaller devices), which can lead to major slowdowns, up to 2×, when done without considering the characteristics of target inference device (libraries and hardware). We show surprising performance patterns of pruned networks running on embedded GPUs, which can instruct the best level of pruning.