Representative Projects

Systems for detecting, tolerating, and recovering from complex failures.

OrderLab builds practical tools for reliable AI training, dependable distributed systems, production debugging, and operating-system support. This page highlights a representative slice of our recent and foundational work.

Most of our lab's code repositories are hosted on GitHub.

AI training reliability Distributed systems Fault injection Runtime checking OS support

Featured Projects

Recent systems and tools that represent the lab's current research directions.

TrainCheck

OSDI 2025

TrainCheck

TrainCheck catches silent errors in deep learning training by learning semantic invariants from sample pipelines and turning them into proactive runtime checks.

Paper Software
TrainVerify figure

SOSP 2025

TrainVerify

TrainVerify uses equivalence-based verification to check the parallelization logic of distributed LLM training and eliminate subtle correctness bugs.

Paper Software
Phoenix figure

SOSP 2025

Phoenix

Phoenix introduces OS-level optimistic recovery and partial process state preservation to improve software availability after failures.

Paper Software
Atropos figure

SOSP 2025

Atropos

Atropos mitigates application resource overload with targeted task cancellation, preserving useful work while shedding harmful load.

Paper Software
T2C figure

OSDI 2025

T2C

T2C derives semantic checkers from existing system tests, transforming developer-written test logic into runtime checks for silent distributed-system failures.

Paper Software
Xinda figure

NSDI 2025

Xinda

Xinda combines automated slow-fault testing with lightweight adaptive detection to improve slow-fault tolerance in modern distributed systems.

Paper Software

Benchmark

DistDebug-Bench

DistDebug-Bench evaluates code-level diagnosis of distributed-system failures, providing a benchmark for debugging tools on realistic distributed failures.

Project Software

More Representative Systems

Selected projects across failure exposure, runtime checking, observability, recovery, and OS resource management.

SOSP 2024

Anduril

Feedback-driven fault injection that quickly reproduces a target fault-induced production failure in distributed systems.

Paper Software

NSDI 2024

Legolas

State-guided fault injection that infers abstract states from code to expose partial-failure bugs in large distributed systems.

Paper Software

SOSP 2023

pBox

Application-level mechanisms for pushing performance isolation boundaries into applications.

Paper Software

OSDI 2022

OathKeeper

Runtime checking for silent semantic violations in large distributed systems, motivated by failures that escape generic detectors.

Paper Software

OSDI 2022

Orbit

Operating-system support for safe and efficient auxiliary execution, enabling speculative work without compromising the main process.

Paper Software

EuroSys 2021

Arthas

System support for understanding and dealing with hard faults in persistent memory systems.

Paper Software

OSDI 2020

Violet

Automated reasoning for detecting specious configuration in large systems with symbolic execution.

Paper Software

NSDI 2020

OmegaGen

Techniques for understanding, detecting, and localizing partial failures in large system software.

Paper
LeaseOS

ASPLOS 2019

LeaseOS

Lease-based mobile OS resource management that mitigates app energy bugs while preserving useful work.

Paper Software
Panorama

OSDI 2018

Panorama

Program-analysis and instrumentation support for in situ observability, helping requesters detect gray failures in cloud systems.

Paper Software
DefDroid

MobiSys 2016

DefDroid

A defensive mobile OS approach that reacts to disruptive app behavior at runtime without compromising app usability.

Paper Project