Ion Stoica
where else to find them
linkedin:ionstoica
what they talk about
all 12 in the filters
where they keep going back
- NSDI3 editions2013–2023
- Ray Summit2 editions2020–2021
all 19 in the filters
on a stage with
all 109 in the filters
who talks about the same things
- Robert Nishihara11 talks
- Vincent Caldeira16 talks
- Joe Spisak23 talks
- Chip Huyen10 talks
- Google DeepMind4 talks
- Ritesh Modi6 talks
| title | conference | date |
|---|---|---|
| Keynote: Reliability in AI (Agentic) Systems | Confidential Computing Summit 2026 | 2026-06-23 |
Study of 86 agent deployments across five domains finds failures are prevalent and no single root cause dominates; manual annotation of 150 traces plus an LLM… | ||
| Keynote Panel Discussion: Three Forces Making Confidential AI A Mandate | Confidential Computing Summit 2026 | 2026-06-23 |
Three forces make Confidential AI mandatory: Mythos‑class models make a malicious agent inside your environment probable, current agent architectures leak… | ||
| RLBoost: Harvesting Preemptible Cloud Resources for Cost-Efficient Reinforcement Learning | NSDI '26 | 2026-05-04 |
RLBoost harvests preemptible GPUs (cloud spot instances and spare cluster capacity) to run rollouts while keeping training on reserved full‑mesh GPUs, claiming… | ||
| Agentix: An Efficient Serving Engine for LLM Agents as General Programs | NSDI '26 | 2026-05-04 |
Agentix intercepts LLM calls from programs, tags requests with program and thread IDs, and maintains a process table of program-level stats to drive decisions… | ||
| Ray: A Distributed Compute Engine for AI | PyTorch Conference 2025 | 2025-10-22 |
Ray claims to be a general distributed compute engine built for modern AI, reporting 7x downloads last year and adoption by thousands of companies. It exposes… | ||
| SuperServe: Fine-Grained Inference Serving for Unpredictable Workloads | NSDI '25 | 2025-04-28 |
SuperServe implements SubNetAct and SlackFit to serve many weight‑shared sub-net models from a single pre-trained super-network by inserting control‑flow… | ||
| Starburst: A Cost-aware Scheduler for Hybrid Cloud | USENIX ATC 2024 | 2024-07-10 |
Starburst is a meta-scheduler that assigns per-job waiting budgets, uses compute-weight (or its practical approximation star-weight with preemption) plus… | ||
| Fairness in Serving Large Language Models | 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI '24) | 2024-07-10 |
Defines fairness as a token-cost function (input vs decode tokens) and implements a scheduler that updates per-client virtual counters at token granularity… | ||
| Can't Be Late: Optimizing Spot Instance Savings under Deadlines | 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) | 2024-04-16 |
Proposes a parameter-free policy Uniform Progress that switches to on-demand only when observed progress falls below the time-uniform expectation, and… | ||
| Cloudcast: High-Throughput, Cost-Aware Overlay Multicast in the Cloud | 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) | 2024-04-16 |
Cloudcast builds an overlay multicast system that spawns ephemeral waypoints (VM routers) and emits per-VM router programs to execute striped, multi-tree… | ||
| Keynote Q&A | Systems @Scale Summer 2023 | 2023-07-18 |
Ray splits into a core plus libraries model: use the libraries for common ML workflows, but implement novel distributed algorithms or custom hardware support… | ||
| Ray, a Unified Distributed Framework for the Modern AI Stack | Systems @Scale Summer 2023 | 2023-07-18 |
Ray presents a unified distributed framework that runs data ingestion, preprocessing, training, tuning and serving within one system, exposing only six core… | ||
| Sky Computing | USENIX ATC '23 | 2023-07-10 |
Sky Computing proposes an intercloud broker that maps application requirements to services across clouds; prototypes Sky Pilot (ML job broker using Ray) and… | ||
| Joint Keynote Address | OSDI 2023 | 2023-07-10 |
Claim: a portability layer is infeasible, so Sky runs apps across clouds by matching app requirements to available services via an intercloud broker… | ||
| ExoFlow: A Universal Workflow System for Exactly-Once DAGs | OSDI 2023 | 2023-07-10 |
ExoFlow implements a workflow runtime that decouples execution from recovery, providing exactly-once semantics via a separate recovery layer, task annotations… | ||
| Take Out the TraChe: Maximizing (Tra)nsactional Ca(che) Hit Rate | OSDI 2023 | 2023-07-10 |
Defines transactional hit rate and builds DeToX, a cache layer that groups keys by dependency (levels and interchangeable groups), scores groups by minimum… | ||
| Cilantro: Performance-Aware Resource Allocation for General Objectives via Online... | OSDI 2023 | 2023-07-10 |
Cilantro is a scheduler that replaces user resource requests with online learning from real-time job feedback to map resources to performance, decouples… | ||
| AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving | OSDI 2023 | 2023-07-10 |
AlpaServe uses model parallelism not just to fit big models but to statistically multiplex GPUs across many models, trading partition overhead for burst… | ||
| SkyPilot: An Intercloud Broker for Sky Computing | NSDI 2023 | 2023-04-17 |
SkyPilot is an intercloud broker that automatically matches jobs to concrete zones and instance types across AWS, GCP, Azure and others, taking into account… | ||
| SHEPHERD: Serving DNNs in the Wild | NSDI 2023 | 2023-04-17 |
SHEPHERD is a model serving system that combines a two-level design planner that partitions hundreds-to-thousands of request streams into groups and an online… | ||
| Skyplane: Optimizing Transfer Cost and Throughput Using Cloud-Aware Overlays | NSDI 2023 | 2023-04-17 |
Skyplane is a system for fast, low-cost bulk transfers between cloud object stores that probes cloud network throughput, models routing with mixed-integer… | ||
| Keynote – Ion Stoica | Data @Scale Spring 2022 | 2022-05-18 |
Claims that Ray consolidates streaming, training, tuning and serving into one system and replaces stitching disparate frameworks, with a minimalist six-call… | ||
| Ekya: Continuous Learning of Video Analytics Models on Edge Compute Servers | 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) | 2022-04-04 |
Edge servers run compressed video models that drift with live data; Ekya continuously retrains selected models and schedules GPU time to maximize mean… | ||
| NetHint: White-Box Networking for Multi-Tenant Data Centers | 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) | 2022-04-04 |
NetHint exposes three lightweight hints — a hierarchical virtual topology, per-link residual bandwidth and number of competing flows, and a 100 ms update… | ||
| Gauss: Program Synthesis by Reasoning Over Graphs | OOPSLA 2021 | 2021-10-17 |
Gauss synthesizes table-transformation programs (Pandas and R) from partial outputs plus user intent graphs that record cell/row/column relations (example… | ||
| [Opening Keynote] Making Distributed Computing Easy | Ray Summit 2021 | 2021-06-22 |
AI workloads outpace single machines — compute demand grows 35x every 18 months while per‑core performance inches up, so training and serving go distributed… | ||
| Twenty Years After: Hierarchical Core-Stateless Fair Queueing | NSDI ’21 | 2021-04-12 |
Implements Core-Stateless Fair Queueing on programmable switches and generalizes it to Hierarchical CSFQ that approximates hierarchical fair queuing while… | ||
| TEGRA: Efficient Ad-Hoc Analytics on Evolving Graphs | NSDI ’21 | 2021-04-12 |
TEGRA exposes evolving graphs via a Timelapse API and stores versions with a persistent adaptive radix tree, enabling snapshot reads without replaying deltas… | ||
| Caerus: NIMBLE Task Scheduling for Serverless Analytics | NSDI ’21 | 2021-04-12 |
Caerus is a task-level scheduler that implements the NIMBLE algorithm to schedule serverless analytics tasks by computing per-task launch times. It models… | ||
| Ownership: A Distributed Futures System for Fine-Grained Tasks | NSDI ’21 | 2021-04-12 |
Ownership claims a decentralized runtime that shards metadata by object owner, trades global persistence for simple recovery, and achieves 1ms latency per task… | ||
| Keynote: Programming the Cloud as Easily as Your Laptop | Ray Summit 2020 | 2020-09-30 |
Claims that training compute grows ~35x every 18 months and model size grew ~20x (200x in the last two years), while single‑core speed and even GPUs/TPUs… | ||
| DistCache: Provable Load Balancing for Large-Scale Storage Systems with Distributed Caching | USENIX ATC '19 | 2019-07-10 |
DistCache splits a single hot-cache into many small nodes, allocating hot items with two independent hash functions and routing queries with the… | ||
| Shuffling, Fast and Slow: Scalable Analytics on Serverless Infrastructure | NSDI 2019 | 2019-02-26 |
Locus implements a hybrid shuffle that buffers hot partitions in fast cache (Redis) and spills to cheap object storage (S3), guided by a performance/cost model… | ||
| Confluo: Distributed Monitoring and Diagnosis Stack for High-speed Networks | NSDI 2019 | 2019-02-26 |
Confluo claims an end-host stack that captures packet-header telemetry at line rate on 10Gbps links using a single core by storing fixed-size header fields in… | ||
| NetChain: Scale-Free Sub-RTT Coordination | NSDI '18 | 2018-04-09 |
NetChain moves the replicated coordination KV store into programmable switches (PISA) to cut client-visible latency to about 9.7 microseconds and to push… | ||
| Ray:A Distributed Computing Platform for Reinforcement Learning | The Artificial Intelligence Conference 2017 | 2017-09-18 |
The show you’re about to hear is part of a series of shows recorded in San Francisco at the Artificial Intelligence Conference. In this episode, I talk with… | ||
| Effective Straggler Mitigation: Attack of the Clones | NSDI 2013 | 2013-04-02 |
Builds a system called Dolly that proactively launches task‑level clones for small interactive jobs, then uses a probabilistic delay assignment to avoid… | ||
| nothing matches | ||