Projects

A cost-effective in-memory object cache built entirely atop ephemeral serverless functions, using erasure coding to keep data available.

A scalable, locality-enhanced serverless parallel framework that decentralizes DAG scheduling across AWS Lambda executors.

Scalable and fast provisioning of custom serverless container runtimes, deployed at Alibaba Cloud Function Compute.

Enables fundamental cacheability for distributed deep learning training by caching the most important training samples without losing accuracy.

An elastic serverless cloud storage system that combines serverless functions and object storage for cost-effective, scalable storage.

An adaptive-learning, priority OS scheduler that learns workload intelligence in user space to schedule serverless functions in the kernel.

An elastic, high-performance metadata service for large-scale distributed file systems, built on serverless functions.

A smart OS scheduler for short-job-dominant serverless function workloads that reduces turnaround time under contention.

A model-aware, lossless data reduction pipeline for large-scale LLM storage, with BitX, a 10+ GB/s compressor for safetensors files.

A GPU-efficient Jupyter Notebooks-as-a-Service platform that allocates GPUs on demand for interactive deep learning training workloads.

Algorithmic complexity attacks that expose performance vulnerabilities in dynamic learned indexes.

I/O caching and scheduling for federated learning that prioritizes the most important client samples to speed up training.

A study of the storage compressibility of pre-trained ML models, with a near-lossless compressor for effective model storage reduction.

A compact, tensor-centric storage system that exploits fine-grained tensor-level relationships across model hubs for lossless model storage reduction.

A workload-aware LLM serving framework that scales GPU memory with runtime layer swapping and KV cache resizing for bursty traffic.

A new function orchestrator that reduces cold-start cost in production FaaS platforms.

A highly efficient Jupyter notebook platform offering free serverless GPUs to everyone at UVA.

A stall-free offloading engine for LLM training that decouples GPU and CPU parameter updates. Integrated into DeepSpeed.