USENIX Courses

Hydro - Surrogate-Based Hyperparameter Tuning Service in Datacenters

Innovative hyperparameter tuning service for deep learning models, optimizing efficiency and resource utilization in datacenters through surrogate models and adaptive resource management.

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk

Chardonnay - Fast and General Datacenter Transactions for On-Disk Databases

Innovative approach to optimize distributed on-disk database transactions using fast 2PC and lock-free read protocol, enabling efficient data prefetching and minimizing contention in datacenter environments.

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk

Replicating Persistent Memory Key-Value Stores with Efficient RDMA Abstraction

Efficient RDMA abstraction for replicating persistent memory key-value stores, improving throughput and latency while reducing write amplification in distributed systems.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

ServiceRouter - Hyperscale and Minimal Cost Service Mesh at Meta

Innovative service mesh solution for hyperscale environments, optimizing RPC routing with minimal hardware costs. Features include direct routing, sharded service support, and locality rings for geo-distributed load balancing.

• YouTube
• 17 minutes
• Self-Paced
• Conference Talk

ShRing - Networking with Shared Receive Rings

Innovative networking technique that shares receive rings among cores, reducing memory footprint and improving throughput for high-bandwidth scenarios, with significant latency reductions in NFV workloads.

• YouTube
• 14 minutes
• Self-Paced
• Conference Talk

ORC - Increasing Cloud Memory Density via Object Reuse with Capabilities

Explore ORC: a novel approach for efficient memory sharing in cloud environments, offering improved density and isolation through fine-grained object reuse and hardware-supported capabilities.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

Global Capacity Management With Flux

Automated system for optimizing capacity distribution across global datacenters, using RPC tracing and service models to efficiently rebalance resources for thousands of services and products.

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk

Effectively Scheduling Computational Graphs of Deep Neural Networks toward Their Domain-Specific Accelerators

Innovative approach for scheduling DNN computational graphs on domain-specific accelerators, optimizing resource utilization and enhancing performance through synergy between network and hardware architectures.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

Core Slicing - Closing the Gap Between Leaky Confidential VMs and Bare-Metal Cloud

Explores core slicing, a novel approach to enhance cloud security by enabling multiple untrusted guest OSes to run on shared bare-metal hardware, addressing vulnerabilities in confidential VMs.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

ExoFlow - A Universal Workflow System for Exactly-Once DAGs

ExoFlow: A universal workflow system enabling flexible recovery vs. performance tradeoffs within applications. Decouples execution from recovery, provides exactly-once semantics, and supports diverse workflow applications with optimized task communicatio…

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk

Flor - An Open High Performance RDMA Framework Over Heterogeneous RNICs

Explore Flor, an innovative framework enabling high-performance RDMA across diverse network interface cards, addressing bandwidth imbalances and enhancing datacenter application efficiency.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

AdaEmbed - Adaptive Embedding for Large-Scale Recommendation Models

Adaptive embedding system for large-scale recommendation models, reducing embedding size while maintaining accuracy. Improves deployment efficiency and model execution speed in industrial settings.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

Cilantro - Performance-Aware Resource Allocation for General Objectives via Online Feedback

Explore Cilantro, an innovative system for performance-aware resource allocation in clusters, using online learning to optimize various scheduling objectives and improve user utilities.

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk

Defcon - Preventing Overload with Graceful Feature Degradation

Explore Defcon, a system for improving large-scale Internet service availability through graceful feature degradation, enabling operators to manage overload conditions and enhance global product reliability.

• YouTube
• 15 minutes
• Self-Paced
• Conference Talk

NCC - Natural Concurrency Control for Strictly Serializable Datastores by Avoiding the Timestamp-Inversion Pitfall

Innovative concurrency control technique for datastores that leverages natural consistency to achieve minimal costs while guaranteeing strict serializability, avoiding common pitfalls and optimizing performance.

• YouTube
• 16 minutes
• Self-Paced
• Conference Talk