Schedule

Note this schedule is subject to change. Please check this page frequently.

Important the required readings must be completed before the class. Paper reviews should be submitted by 12pm on the day of the class.

Date Topic Readings Presenter Assignment
W1 Tue 09/01 Course Intro
  1. How to Read a Paper, Keshav, ACM SIGCOMM CCR, 2007
  2. How (and How Not) to Write a Good Systems Paper, Levin and Redell, ACM SIGOPS OSR, 1983
 Ryan Reviews not required
Thu 09/03 Background
  1. A Few Useful Things to Know about Machine Learning, Domingos, CACM, 2012
  2. Park: An Open Platform for Learning-Augmented Computer Systems, Mao et al., NeurIPS, 2019
Questions
  • Q: Which of Domingos's general lessons become harder when systems data is non-stationary, workload-dependent, and expensive to label?
  • Q: What must a platform such as Park expose about latency, safety, and resource costs so that an agent cannot optimize the wrong objective?
  • Q: For a new systems problem, what evidence would justify reinforcement learning over supervised learning or a well-tuned heuristic?
 Ryan Reviews not required
W2 Tue 09/08 Early ML for Systems
  1. Correlating Instrumentation Data to System States: A Building Block for Automated Diagnosis and Control, Cohen, Goldszmidt, Kelly, Symons, and Chase, OSDI, 2004
  2. Detecting Large-Scale System Problems by Mining Console Logs, Xu, Huang, Fox, Patterson, and Jordan, SOSP, 2009
Additional Reading
Questions
  • Q: The two papers learn from numerical instrumentation and unstructured logs, respectively. What information does each representation preserve or discard?
  • Q: What makes a diagnosis actionable rather than merely accurate, and how well does each paper meet that standard?
  • Q: How would these methods behave after a software update changes metrics, log templates, or the underlying failure distribution?
 Ryan
Thu 09/10 Early ML for Systems
  1. Automatic Misconfiguration Troubleshooting with PeerPressure, Wang, Platt, Chen, Zhang, and Wang, OSDI, 2004
  2. Capturing, Indexing, Clustering, and Retrieving System History, Cohen, Zhang, Goldszmidt, Symons, Kelly, and Fox, SOSP, 2005
Additional Reading
Questions
  • Q: PeerPressure compares a machine with a population of peers, while the history paper compares a state with prior executions. What assumptions make each comparison valid?
  • Q: The history paper represents system state as a signature. What information is lost, and when would that loss matter?
  • Q: How would either system handle a novel configuration or failure mode for which no useful peer or historical state exists?
 Project teams formed
W3 Tue 09/15 Operating Systems
  1. Learning-Based Memory Allocation for C++ Server Workloads, Maas, Andersen, Isard, Javanmard, McKinley, and Raffel, ASPLOS, 2020
  2. Learning to Walk: Architecting Learned Virtual Memory Translation, Zhao, Chen, Xu, Schatzberg, Hajinazar, Vakharwala, Anderson, and Skarlatos, MICRO, 2025
Additional Reading
Questions
  • Q: Where does the learned allocator policy sit in the allocation path, and how does the design keep inference overhead from dominating the decision?
  • Q: How does LVM make learned indexing practical in the address-translation path while supporting updates, multiple page sizes, and precise mappings?
Thu 09/17 Operating Systems
  1. Machine Learning for Load Balancing in the Linux Kernel, Chen, Banerjee, Kalbarczyk, and Iyer, APSys, 2020
  2. Wayfinder: Automated Operating System Specialization, Jung, Crăciunoiu, Karaolidis, Lefeuvre, Rubio, Huici, Rotsos, and Olivier, EuroSys, 2026
Additional Reading
Questions
  • Q: Compare the timescales of the two decisions: per-task load balancing versus workload-specific OS specialization. How does timescale change the acceptable training and inference cost?
  • Q: What safeguards are needed before either learned policy can influence a kernel, and where should fallback behavior live?
  • Q: How convincing is the evidence that the learned policy generalizes across workloads, machines, and software versions?
W4 Tue 09/22 Operating Systems
  1. Towards a Machine Learning-Assisted Kernel with LAKE, Fingler, Tarte, Yu, Szekely, Hu, Akella, and Rossbach, ASPLOS, 2023
  2. Velosiraptor: Code Synthesis for Memory Translation, Achermann, Chu, Mehri, Karimalis, and Seltzer, ASPLOS, 2025
Additional Reading
Questions
  • Q: Why is accelerator access from kernel context difficult, and which parts of LAKE's solution are specific to GPUs versus general to external inference services?
  • Q: Velosiraptor obtains guarantees through synthesis, whereas LAKE executes learned models. Where should a systems designer draw the boundary between learning and synthesis?
  • Q: Both systems move complex computation away from a critical path. What remains on the path, and what failures can still propagate into the running system?
Thu 09/24 No Class: Work on Proposal
  1. Producing Wrong Data Without Doing Anything Obviously Wrong!, Mytkowicz, Diwan, Hauswirth, and Sweeney, ASPLOS, 2009
  2. Dos and Don'ts of Machine Learning in Computer Security, Arp, Quiring, Pendlebury, Warnecke, Pierazzi, Wressnegger, Cavallaro, and Rieck, USENIX Security, 2022
 Project proposal due
W5 Tue 09/29 Project Proposal Pitches
Thu 10/01 Storage Systems
  1. LinnOS: Predictability on Unpredictable Flash Storage with a Light Neural Network, Hao, Toksoz, Li, Halim, Hoffmann, and Gunawi, OSDI, 2020
  2. Improving Storage Systems Using Machine Learning, Akgun, Aydin, Burford, McNeill, Arkhangelskiy, and Zadok, ACM Transactions on Storage, 2023
Additional Reading
Questions
  • Q: LinnOS makes an inference on every I/O. What is its inference budget, and how does the design stay within it?
  • Q: Using the survey's taxonomy, which storage decisions are especially suitable for learning, and which should remain conventional mechanisms?
  • Q: What evidence would show that LinnOS's model remains useful under device aging, workload drift, and hardware replacement?
W6 Tue 10/06 Database
  1. The Case for Learned Index Structures, Kraska, Beutel, Chi, Dean, and Polyzotis, SIGMOD, 2018
  2. Bao: Making Learned Query Optimization Practical, Marcus, Negi, Mao, Tatbul, Alizadeh, and Kraska, SIGMOD, 2021
Additional Reading
Questions
  • Q: In what sense is a B-tree already a model, and what does reframing an index that way enable?
  • Q: The learned index replaces a data structure, while Bao steers a conventional query optimizer. Which integration strategy has the safer failure mode?
  • Q: How do updates, workload drift, and online exploration challenge the assumptions behind these two systems?
Thu 10/08 Cache
  1. HALP: Heuristic Aided Learned Preference Eviction Policy for YouTube Content Delivery Network, Song, Chen, Sarda, Altınbüken, Brevdo, Coleman, Ju, Jurczyk, Schooler, and Gummadi, NSDI, 2023
  2. Learning-Augmented Heuristics: Simple Yet Smart, Robust and Interpretable Cache Eviction, Xia, Nixon, Marthen, Bhandari, and Yang, OSDI, 2026
Additional Reading
Questions
  • Q: HALP is deployed in a production CDN, while LAH emphasizes robustness and interpretability. What evidence from each paper matters most for deciding whether a learned cache policy should ship?
  • Q: Where do these designs place learning relative to the eviction data path, and what latency or state overhead follows from that choice?
  • Q: How should a cache policy be evaluated when traces are non-stationary and a small hit-rate gain may hide worse tail behavior?
W7 Tue 10/13 Networking
  1. Learning in situ: A Randomized Experiment in Video Streaming, Yan, Ayers, Zhu, Fouladi, Hong, Zhang, Levis, and Winstein, NSDI, 2020
  2. Mutant: Learning Congestion Control from Existing Protocols via Online Reinforcement Learning, Pappone, Sacco, and Esposito, NSDI, 2025
Additional Reading
Questions
  • Q: What does the randomized production experiment reveal that simulation or trace replay cannot?
  • Q: Mutant selects among existing congestion-control protocols rather than generating control actions directly. How does that choice change the safety and exploration problem?
  • Q: How should these systems account for fairness to competing flows when optimizing user-facing performance?
Thu 10/15 Cloud Resource Management
  1. Cilantro: Performance-Aware Resource Allocation for General Objectives via Online Feedback, Bhardwaj, Kandasamy, Biswal, Guo, Hindman, Gonzalez, Jordan, and Stoica, OSDI, 2023
  2. Autothrottle: A Practical Bi-Level Approach to Resource Management for SLO-Targeted Microservices, Wang, Li, Liang, Wu, and Yan, NSDI, 2024
Additional Reading
Questions
  • Q: Cilantro learns resource-to-performance mappings, while Autothrottle separates application-wide learning from fast per-service control. What does each decomposition buy?
  • Q: How many SLO violations or inefficient allocations must an online controller tolerate while it learns, and are the papers' evaluation protocols realistic about that cost?
  • Q: Where should uncertainty estimates, guardrails, and heuristic fallbacks enter these control loops?
W8 Tue 10/20 Fall Study Break
Thu 10/22 No Class: Project Hack Day Progress #1 report due
W9 Tue 10/27 Cloud Resource Management
  1. Resource Central: Understanding and Predicting Workloads for Improved Resource Management in Large Cloud Platforms, Cortez, Bonde, Muzio, Russinovich, Fontoura, and Bianchini, SOSP, 2017
  2. SelfTune: Tuning Cluster Managers, Karthikeyan, Natarajan, Somashekar, Zhao, Bhagwan, Fonseca, Racheva, and Bansal, NSDI, 2023
Additional Reading
Questions
  • Q: Resource Central separates prediction from the resource managers that consume predictions. What does that architectural split buy, and what does it cost?
  • Q: SelfTune learns tuning policies for cluster managers. How does it transfer knowledge without making unsafe assumptions about a new workload or deployment?
  • Q: When is predicting workload behavior first preferable to learning the resource-management action directly?
Thu 10/29 Cloud Operations
  1. Gandalf: An Intelligent, End-to-End Analytics Service for Safe Deployment in Large-Scale Cloud Infrastructure, Li, Cheng, Hsieh, Dang, Huang, Singh, Yang, Lin, Wu, Levy, and Chintalapati, NSDI, 2020
  2. Hardware Lifecycle-Aware Power Planning in Commercial Hyperscale Datacenters, Li, Piga, Su, Torres, Stojkovic, Yadwadkar, John, and Dhanotia, OSDI, 2026
Additional Reading
Questions
  • Q: What is the cost asymmetry between false positives and false negatives in safe deployment and power planning, and how does each design reflect it?
  • Q: PowerSight must predict new hardware behavior before production telemetry exists. Which features and validation evidence make that extrapolation credible?
  • Q: Why do both systems combine learned models with domain rules or established operational processes instead of using an end-to-end model?
W10 Tue 11/03 Diagnosis
  1. Sage: Practical and Scalable ML-Driven Performance Debugging in Microservices, Gan, Liang, Dev, Lo, and Delimitrou, ASPLOS, 2021
  2. Automatic Root Cause Analysis via Large Language Models for Cloud Incidents, Chen et al., EuroSys, 2024
Additional Reading
Questions
  • Q: Sage uses causal structure, while RCACopilot uses an LLM over incident evidence. What does each approach assume about observability and prior knowledge?
  • Q: Root-cause analysis is difficult to evaluate. What ground truth and success criteria would make the reported results convincing to an on-call engineer?
  • Q: Where should deterministic checks or human review constrain an LLM diagnosis before it can trigger remediation?
Thu 11/05 Project Progress Presentations
W11 Tue 11/10 Testing
  1. Snowcat: Efficient Kernel Concurrency Testing Using a Learned Coverage Predictor, Gong, Peng, Altınbüken, Fonseca, and Maniatis, SOSP, 2023
  2. KernelGPT: Enhanced Kernel Fuzzing via Large Language Models, Yang, Zhao, and Zhang, ASPLOS, 2025
Additional Reading
Questions
  • Q: Snowcat predicts coverage even though exact coverage can eventually be measured. Why is an approximation useful in this feedback loop?
  • Q: KernelGPT asks an LLM to generate syscall specifications and then hands them to a conventional fuzzer. Where is the trust boundary, and how is generated output validated?
  • Q: How do the two systems avoid reinforcing a narrow model of the input space and missing qualitatively different bugs?
Thu 11/12 Program Analysis and Verification
  1. KNighter: Transforming Static Analysis with LLM-Synthesized Checkers, Yang, Zhao, Xie, Li, and Zhang, SOSP, 2025
  2. Neuro-Symbolic Proof Generation for Scaling Systems Software Verification, He, Li, Sun, Yao, Chen, Ma, and Su, OSDI, 2026
Additional Reading
Questions
  • Q: KNighter has an LLM synthesize a checker, while the proof-generation system has an LLM propose steps to a verifier. Which design degrades more gracefully when the model is wrong?
  • Q: A proof assistant can reject an invalid step. What equivalent validation exists for an LLM-generated static-analysis checker?
  • Q: If most but not all analysis or proof work is automated, where does the remaining human effort concentrate?
W12 Tue 11/17 LLMs for Systems Development
  1. QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a Neural-Symbolic Approach, Dong et al., OSDI, 2025
  2. ECO: An AI-Driven Code Efficiency Optimizer for Warehouse-Scale Computers, Lin et al., OSDI, 2026
Additional Reading
Questions
  • Q: QiMeng-Xpiler uses symbolic repair, while ECO relies on testing, execution feedback, self-review, and deployment monitoring. What classes of error can each validation strategy catch?
  • Q: How should these systems trade optimization aggressiveness against semantic correctness and regression risk?
  • Q: How much of each system's gain comes from the LLM itself versus search, profiling, domain-specific transformations, and feedback?
Thu 11/19 LLMs for Autonomous System Operations
  1. AIOpsLab: A Holistic Framework to Evaluate AI Agents for Enabling Autonomous Clouds, Chen, Shetty, Somashekar, Ma, Simmhan, Mace, Bansal, Wang, and Rajmohan, arXiv, 2025
  2. Xpert: Empowering Incident Management with Query Recommendations via Large Language Models, Jiang, Zhang, He, Yang, Ma, Qin, Kang, Dang, Rajmohan, Lin, and Zhang, ICSE, 2024
Additional Reading
Questions
  • Q: What aspects of real production operations does AIOpsLab reproduce well, and which organizational or environmental factors must it necessarily omit?
  • Q: Xpert recommends queries rather than taking corrective actions. How does recommendation change the acceptable error rate and the role of the human operator?
  • Q: What metrics should evaluate an operations agent's consistency, recovery behavior, and cost—not just whether it eventually completes a task?
W13 Tue 11/24 LLMs for Systems Research
  1. Principles and Methodologies for Serial Performance Optimization, Park, Guan, Cheng, and Kim, OSDI, 2025
  2. Barbarians at the Gate: How AI Is Upending Systems Research, Cheng et al., arXiv, 2025
Additional Reading
Questions
  • Q: The optimization paper reduces techniques to removal, replacement, and reordering. Is that taxonomy predictive, or only a useful retrospective description?
  • Q: How should SysGPT be evaluated for genuinely useful optimization ideas rather than plausible restatements of patterns in its training literature?
  • Q: If AI makes design-space exploration and implementation cheaper, what becomes the scarce resource in systems research?
 Progress #2 report due
Thu 11/26 Thanksgiving Recess
W14 Tue 12/01 Pitfalls
  1. On the Use of ML for Blackbox System Performance Prediction, Fu, Gupta, Mittal, and Ratnasamy, NSDI, 2021
  2. Is Machine Learning Necessary for Cloud Resource Usage Forecasting?, Christofidi, Papaioannou, and Doudali, SoCC, 2023
Additional Reading
Questions
  • Q: Both required papers question whether sophisticated ML beats simple baselines. What baseline and tuning-budget rules should become mandatory in learned-systems evaluations?
  • Q: Which observed prediction failures are fundamental limits of the available signal, and which are consequences of a poor model or evaluation design?
  • Q: When does a negative result generalize into a design principle, and when is it only evidence about one workload or deployment?
Thu 12/03 Open Challenges
  1. AutoSys: The Design and Operation of Learning-Augmented Systems, Liang, Xue, Yang, Zhou, Zhu, Li, Wang, Chen, Zhang, Liu, and Dai, USENIX ATC, 2020
  2. Measuring Agents in Production, Pan et al., ICML, 2026
Additional Reading
Questions
  • Q: Which production lessons in AutoSys are failures of learning algorithms, and which are failures of interfaces, monitoring, or operational process?
  • Q: Production agents are commonly short-horizon, human-supervised, and evaluated by people. Should research benchmarks reproduce those constraints or push beyond them?
  • Q: What safety property should an autonomous operations agent satisfy when its telemetry can be manipulated, and where could that property be enforced?
W15 Tue 12/08 No Class: Project Hack Day
Thu 12/10 Final Project Presentations Slides due 24h before
W16 Fri 12/18 Final Project Report Due Final report + code due