Best Paper Award (First Prize)
DLIO is a benchmark suite that emulates the I/O behavior of scientific deep learning workloads, based on profiling on Argonne's Theta supercomputer. It helps application developers and system architects identify I/O bottlenecks and guide optimizations that reduce training time.
Karsten Schwan Best Paper Award
LABIOS uses a label-based data representation and asynchronous I/O to support conflicting scientific and Big Data workloads within a single distributed storage system. It combines heterogeneous storage resources, storage elasticity, and data provisioning for in-situ analytics.
378 publications
We present GPUH5, a GPU-first I/O library for persistent kernels.
We present GDSight, which attributes each NVMe command to the operation that caused it, whether it entered cuFile or fell back to POSIX.
StageOnce coordinates these speculative path predictions through a node-local daemon.
ToolFuse addresses this cold-wave gap by coalescing active calls at a gateway between agents and tool handlers.
We present Vigil, which couples deterministic science-aware triggers with multimodal LLM confirmation through a trigger-render-reason pipeline.
Scientific simulations produce increasingly large datasets that surpass practical limits for frequent input and output operations.
We present the Agent Error Simulator (AES), a benchmark framework that evaluates LLM agent fault tolerance through controlled, declarative fault injection.
We propose PKAS, a Predictive KV Cache-Aware Scheduling algorithm to mitigate this inefficiency by reducing preemptions.
This paper presents a workflow-centric methodology to characterize and analyze I/O using workflow-specific patterns and metrics.
We present an empirical study across two scientific workflows, genomics and climate tracking, on a production HPC cluster.
We propose a speculative dispatch mechanism to accelerate multi-agent orchestration.
LLM coding agents are increasingly used to generate scientific workflows, but evaluations stop at output correctness.
To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM's PagedAttention.
To fill this gap, in this paper we study the impact of KV caching.
Modern HPC workloads involve large data transfers that can become bottlenecks.