[CS] REMINDER: Tianshuo Su MS Presentation/Oct 4,2024

[via cs]

This is an announcement of Tianshuo Su's MS Presentation
===============================================
Candidate: Tianshuo Su

Date: Friday, October 4th, 2024

Time: : 3:30pm CT 

Title: Scalable Data Abstractions for Irregular Parallel Applications

Abstract: Scaling is difficult in large-scale irregular applications, such as graph processing and sparse computations. These applications employ fine-grained data access with skewed patterns and work distribution. These dimensions of irregularity make efficient parallelization and work distribution difficult.
Our thesis is that efficient, scalable data abstractions can be constructed for skewed, irregular inputs on fine-grained parallel architectures. These abstractions, Scalable HashTable (SHT) and Parallel Graph Abstraction (PGA), are broadly applicable beyond graphs.

The design and evaluation are done using the UpDown system architecture, which pro-vides thousands of fine-grained event-driven processors per node. We demonstrate the scalability of SHT on UpDown using half-node (1024 processors) to 8-node (16,384 processors) weak scaling simulations. The SHT shows 14.8×, 17.1×, and 17.2× speedup for insert, update, and get experiments. PGA insert edge experiment shows up to 16.1× speedup on the same setup.

The SHT and PGA also show that the average instruction cost of operations ranges from100s to 300s in cycles, comparable to high-performance implementations on multicore CPUs.


Advisors: Andrew Chien

Committee Members: Andrew Chien, Hank Hoffmann, Aaron Elmore, David Gleich