Zhaoqi(Roy) Zhang

PhD Candidate at Tufts University

Talks and presentations

Accurate and Predictable Scoring for Hardware Contention in Microservices

December 18, 2025
Industry Research Talk, Tufts University

This talk presents a practical approach to quantifying and forecasting hardware contention in microservices so mitigation can happen before SLO violations. It motivates why contention remains underappreciated in containerized systems, introduces a general contention score (0-1) designed around four properties (correlation with latency degradation, sharp rise, slow decay, and predictability), and describes the Gordion pipeline: workload/stress generation (wrk2 + stress-ng), multi-layer instrumentation (within-service gRPC timing plus perf counters), score labeling, and ML-based next-step prediction. The evaluation highlights both usefulness and deployability, including competitive prediction quality and low runtime overhead near saturation (about 1% mean and 8% P90 at 500 RPS). Slides: (Industry) Accurate and Predictable Scoring for Hardware Contention in Microservices.

Automatically Identifying Errant Upstream Contention Triggers in Microservice-based Applications

December 20, 2024
Industry Research Talk, Tufts University

This talk presents OrthoView, a tracing-based method for diagnosing errant upstream contention triggers in microservice applications, where user-visible slowdowns are caused or sustained by problematic upstream request behaviors rather than only local service issues. The presentation frames the problem through victim/survivor differentiation under SLO violations, then walks through a stepwise pipeline: heavy-hitter edge detection on critical paths, temporal-overlap aggressor discovery, and backtrace feature scoring/ranking to identify likely problematic contention triggers and assess whether they are truly errant. It also covers implementation and evaluation on a large TrainTicket testbed under controlled bursty-load scenarios, highlighting initial effectiveness and practical next steps around tracing-overhead control and sampling strategies. Slides: CACTI-CGM-postICPE.

(Guest Lecture) TA’s Journey of Solving OS Research Questions

November 04, 2024
Operating System Course, Tufts University

This guest lecture walks through a systems-research journey from distributed tracing in user space to debugging invisible kernel- and hardware-level bottlenecks in microservices, highlighting why request context is lost at kernel boundaries and how that limits cross-layer diagnosis. The talk compares three context-propagation approaches—extra syscalls, eBPF-based propagation/maps/ring buffers, and customized syscall hooks inspired by Zpoline—and discusses their trade-offs in overhead, maintainability, and observability, with the practical takeaway that efficient context propagation plus low-overhead kernel-to-user data paths are key to scalable vertical tracing. Slides: TA’s journey of solving OS research questions.

Fun talks

(Fun Talk) Math in Basketball

October 07, 2024
Internal Group Meeting, Tufts University

This talk explains how basketball analytics evolved from basic box-score numbers to advanced metrics such as BPM, EPM, RAPTOR, and LEBRON, then uses case studies to show both the power and limits of statistics in practice: the three-point revolution (expected value), Moneyball-style decision making around player value (including the James Harden trade context), and a Scottie Pippen example showing why metric changes must be interpreted with team context rather than used blindly. The closing takeaway is to cross-validate across multiple metrics and real game context instead of trusting any single stat. Slides: Math in Basketball.

(Fun Talk) Fighting Game Explained

September 30, 2024
Internal Group Meeting, Tufts University

This talk explains fighting games from a few complementary angles: what makes something a fighting game, how hitboxes and hurtboxes plus frame data define “fair” timing between players, why moves behave like rock-paper-scissors (hit/block/dodge/throw with high/mid/low variants), how combos emerge from framedata and recovery, and how the deeper “non-Nash equilibrium” of character-specific options creates mind-game and strategy (learning the opponent’s pattern, setting up baits, and punishing whiffs). Slides: Fighting Games Explained.