Parrot: Efficient Serving of LLM-based Applications with Semantic Variable
Discover Parrot, a system enhancing LLM-based applications with Semantic Variables. Learn how this innovation optimizes performance and enables efficient programming of AI agents and co-pilots.
Motor: Enabling Multi-Versioning for Distributed Transactions on Disaggregated Memory
Innovative system for efficient distributed transactions on disaggregated memory, improving throughput and reducing latency through multi-versioning and RDMA-based protocols.
Fairness in Serving Large Language Models
Explore innovative scheduling algorithms for fair and efficient LLM inference services, addressing challenges of unpredictable request lengths and unique batching characteristics on parallel accelerators.
Nomad: Non-Exclusive Memory Tiering via Transactional Page Migration
Innovative memory management strategy for tiered systems, improving performance under pressure through non-exclusive page retention and transactional migration.
USHER: Holistic Interference Avoidance for Resource Optimized ML Inference
Optimizes ML inference systems by maximizing GPU resource utilization and minimizing interference, resulting in higher goodput and cost-efficiency for deep learning model deployment.
Caravan - Practical Online Learning of In-Network ML Models with Labeling Agents
Practical online learning system for in-network ML models, addressing automatic labeling and efficient performance monitoring. Improves model adaptability to dynamic network environments.
Automatically Reasoning About How Systems Code Uses the CPU Cache
Technique for automatic analysis of CPU cache usage in systems code. Enables developers to identify performance issues and security vulnerabilities without extensive benchmarking.
ChameleonAPI: Automatic and Efficient Customization of Neural Networks for ML Applications
Automatic customization of neural networks for ML applications, reducing incorrect decisions by 43% compared to commercial APIs without changing application code.
Validating the eBPF Verifier via State Embedding
Innovative technique for validating eBPF verifier correctness using state embedding, uncovering critical Linux kernel security bugs and improving overall system reliability.
nnScaler: Constraint-Guided Parallelization Plan Generation for Deep Learning Training
Innovative framework for efficient deep learning training parallelization, using primitives and constraints to generate optimized execution plans, outperforming existing solutions for popular DNN models.
ServiceLab: Preventing Tiny Performance Regressions at Hyperscale through Pre-Production Testing
Explore Meta's ServiceLab platform for detecting minute performance regressions in large-scale systems, addressing challenges in noisy cloud environments through statistical analysis and machine factor studies.
SquirrelFS - Using the Rust Compiler to Check File-System Crash Consistency
SquirrelFS: A novel approach to building crash-safe file systems using Rust's typestate pattern for compile-time enforcement, offering performance comparable to state-of-the-art alternatives.
μSlope - High Compression and Fast Search on Semi-Structured Logs
Innovative system for compressing and searching semi-structured log data, offering superior compression ratios and faster search speeds compared to existing solutions for managing large-scale log datasets.
Chop Chop: Byzantine Atomic Broadcast to the Network Limit
Innovative Byzantine Atomic Broadcast system achieving high throughput and low latency through novel batching technique. Demonstrates superior performance in geo-distributed deployments for various applications.
Anvil: Verifying Liveness of Cluster Management Controllers
Innovative framework for verifying liveness in cluster management controllers, ensuring continuous reconciliation to desired states despite failures and asynchrony in cloud environments.
IntOS: Persistent Embedded Operating System for Multi-threaded Intermittent Computing
Innovative embedded OS for battery-less devices, enabling efficient multi-threaded computing with crash consistency and simplified programming through novel replay-and-bypass approach.
Harvesting Memory-bound CPU Stall Cycles in Software with MSH
Innovative software system for efficiently harvesting memory-bound CPU stall cycles, offering configurable latency and high concurrency to improve datacenter workload performance.
CyberStar - Simple, Elastic and Cost-Effective Network Functions Management in Cloud Networks
Innovative cloud-scale network function service leveraging elastic compute for massive traffic and diverse requests. Achieves 100Gbps bandwidth and scales to millions of connections within seconds.
QDSR: Accelerating Layer-7 Load Balancing by Direct Server Return with QUIC
Innovative load balancing scheme using QUIC and Direct Server Return to enhance cloud service performance, increasing request processing capacity and significantly reducing latency in high-load scenarios.
SuperBench - Improving Cloud AI Infrastructure Reliability with Proactive Validation
Proactive validation system enhancing cloud AI infrastructure reliability by mitigating hidden degradation, featuring comprehensive benchmarks and intelligent component evaluation for optimal performance.
Every Mapping Counts in Large Amounts - Folio Accounting
Innovative folio accounting technique for Linux enhances memory management, improving performance in fork and unmap operations. Demonstrates significant speedup in fork-intensive workloads like Redis.