[CS] Updated: Jiya Su MS PresentationJun 26, 2025

[via cs]

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

Date: Thursday, June 26, 2025

Time:  9 am CST

Remote Location: https://urldefense.com/v3/__https://uchicago.zoom.us/j/96782882677?pwd=kKSVv7ROBwYlPcPebaUHzOXmX8vpwV.1__;!!BpyFHLRN4TMTrA!6D4GO2ItPPFP0TjRAsVOIFWVVUFHjsQ1n4T6Jf8oqCOQRMKn8h0QRW4hXP6dYf2KSq-0FmXP5rmB_T39EMY$

Location: JCL 298

Title: Understanding Network Communication Requirements for the UpDown System

Abstract: The UpDown System is a novel fine-grained, event-driven architecture designed for irregular graph computation, that has demonstrated scalable performance to 16,384 nodes. Due to its unique event-driven mechanism, UpDown exhibits distinct computation and DRAM access patterns compared to conventional architectures such as CPUs and GPUs, thus, UpDown has different network communication characteristics. Understanding the network traffic pat- terns generated by UpDown applications—and how parameters such as latency and injection bandwidth impact performance—is critical for system designers to identify appropriate net- work configurations.

In this thesis, we simulate three irregular graph applications—Triangle Counting (TC), Breadth-First Search (BFS), and PageRank (PR)—on 1 to 512 UpDown nodes under varying network latencies and injection bandwidth constraints, and we project performance trends for a 16,384-node system. Results show that UpDown applications generate uniform traffic across nodes, with packet sizes ranging from 16 to 88 bytes. These applications produce approximately 0.5 to 2 bytes of network traffic per instruction, requiring 2–8 TB/s of injec- tion bandwidth per node. Moreover, system performance remains stable under constrained injection bandwidth and increased network latencies of up to 4000 ns (8 times that of the hardware designed), indicating strong resilience to both transient bandwidth fluctuations and network designs that deliver longer communication delays.

Advisors: Andrew Chien

Committee Members:  Andrew A. Chien, David Gleich, and Nick Feamster