USENIX Courses

Accelerating Design Space Exploration for LLM Training Systems with Multi-experiment Parallel Simulation

Discover Multiverse, a GPU-based simulator that accelerates design space exploration for LLM training systems, achieving 43-73X speedup over CPU simulators with

• YouTube
• 16 minutes
• Self-Paced
• Free Video

MTP - Transport for In-Network Computing

Discover MTP, a novel transport protocol designed for in-network computing that handles message mutation, reordering, and delays while managing network device constraints through reliability and congestion control.

• YouTube
• 14 minutes
• Self-Paced
• Free Video

GPU-Disaggregated Serving for Deep Learning Recommendation Models at Scale

Discover how Prism eliminates GPU fragmentation in deep learning recommendation systems through resource disaggregation, reducing CPU/GPU waste by 53%/27% while enabling efficient scaling.

• YouTube
• 18 minutes
• Self-Paced
• Free Video

Region-based Content Enhancement for Efficient Video Analytics at the Edge

Discover how region-based content enhancement optimizes video analytics at the edge, achieving 10-19% accuracy improvement and 2-3× throughput over traditional frame-based methods.

• YouTube
• 14 minutes
• Self-Paced
• Free Video

Towards Energy Efficient 5G vRAN Servers

Discover how RENC system reduces 5G vRAN server CPU power consumption by 45% through intelligent frequency scaling while meeting strict real-time deadlines in cellular networks.

• YouTube
• 17 minutes
• Self-Paced
• Free Video

VEP - A Two-stage Verification Toolchain for Full eBPF Programmability

Discover VEP, an annotation-guided toolchain that overcomes eBPF verifier limitations, enabling full programmability through automated verification and flexible kernel security approaches.

• YouTube
• 17 minutes
• Self-Paced
• Free Video

AutoCCL - Automated Collective Communication Tuning for Accelerating Distributed and Parallel DNN Training

Discover AutoCCL, an automated tuning method that optimizes collective communication libraries for distributed DNN training, achieving up to 1.80× speedups without extra costs.

• YouTube
• 15 minutes
• Self-Paced
• Free Video

Making Serverless Pay-For-Use a Reality with Leopard

Discover how Leopard revolutionizes serverless billing with Nearly Pay-for-Use model, challenging current practices to achieve true resource-based pricing and 2x throughput improvements.

• YouTube
• 18 minutes
• Self-Paced
• Free Video

Rajomon - Decentralized and Coordinated Overload Control for Latency-Sensitive Microservices

Discover RAJOMON, a market-based overload control system using distributed rate-limiting and load shedding to protect microservice performance during traffic spikes and demand surges.

• YouTube
• 16 minutes
• Self-Paced
• Free Video

CEGS - Configuration Example Generalizing Synthesizer

Discover how CEGS automates network configuration synthesis using GNNs and LLMs to identify examples from manuals and generalize them to any topology without human involvement.

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

Runtime Protocol Refinement Checking for Distributed Protocol Implementations

Discover runtime bug detection in distributed protocol implementations using RPRC, a system that monitors deployed services like Etcd and Zookeeper to catch safety violations.

• YouTube
• 16 minutes
• Self-Paced
• Free Video

Mutant - Learning Congestion Control from Existing Protocols via Online Reinforcement Learning

Discover Mutant, an online reinforcement learning algorithm that adapts congestion control by learning from best-performing protocols, achieving lower delays and higher throughput.

• YouTube
• 17 minutes
• Self-Paced
• Free Video

ODRP - On-Demand Remote Paging with Programmable RDMA

Discover ODRP, a breakthrough system using programmable RDMA to offload memory management to NICs, achieving better memory utilization with minimal CPU usage and low performance overhead.

• YouTube
• 14 minutes
• Self-Paced
• Free Video

eTran - Extensible Kernel Transport with eBPF

Discover how eTran revolutionizes kernel transport design using eBPF to enable agile implementation of diverse network protocols with significant performance improvements over existing solutions.

• YouTube
• 15 minutes
• Self-Paced
• Free Video

BFTBrain - Adaptive BFT Consensus with Reinforcement Learning

Discover how BFTBrain uses reinforcement learning to adaptively switch between Byzantine fault-tolerant protocols in real-time, improving throughput by up to 119% under dynamic conditions.

• YouTube
• 17 minutes
• Self-Paced
• Free Video