PKAS and Metis at HPDC 2026
Two papers with GRC authors appeared at the 35th International Symposium on High-Performance Parallel and Distributed Computing (HPDC 2026), held in Cleveland from July 13 to 16. PKAS, which Jie Ye presented on July 14, schedules LLM inference by predicting how much KV cache memory each request will need. Metis, in the REX-IO workshop, uses LLM agents to explain what HPC I/O functions do.
PKAS: admitting requests with the future in mind
LLM inference engines keep a KV cache of attention states so they do not recompute them for every new token. The cache grows with sequence length and batch size and easily outgrows GPU memory. To keep GPUs busy, engines such as vLLM use continuous batching: they mix new requests, which process a prompt (prefill), with ongoing requests that generate tokens (decode).
The schedulers that decide which new requests to admit look only at the memory free for the next step. They do not account for the KV cache those requests will need as they decode. When the cache overflows, the engine preempts requests and later recomputes them, which hurts both throughput and latency.
PKAS, short for Predictive KV Cache-Aware Scheduling, runs a low-overhead simulation of future KV cache use and combines it with a lightweight prediction of output length to decide whether a candidate request can be admitted safely. The team built PKAS into vLLM's scheduling interface and compared it with vLLM's default first-come, first-served scheduler and with a learning-to-rank scheduler. The tests ran on Argonne's Polaris system, with Llama 3.1 8B and a mixture-of-experts coding model, on two long-context datasets (NarrativeQA and codeU, with average inputs of 24,000 and 31,500 tokens) and on ShareGPT conversations. PKAS reached up to 7.34 times higher throughput and 8 times lower latency, with the largest gains on long-context workloads where KV cache pressure is highest.
The authors are Ye, Avinash Maurya and Bogdan Nicolae of Argonne National Laboratory, Krishna Teja Chitty-Venkata of Red Hat, Anthony Kougkas and Xian-He Sun. The work is part of DeepIO, GRC's research on I/O for scientific AI.
Metis: agents that explain I/O behavior
Explaining why an HPC application's I/O is slow requires knowing what each I/O function means and what role it plays at runtime, across every layer of the software stack. Experts still do much of this characterization by hand, which does not scale and is hard to reproduce as I/O libraries grow more complex.
Metis uses agentic LLMs to combine different sources of evidence about each function, and measures how far the agents agree. Compared with a baseline built on Model Context Protocol (MCP) tools, Metis generalized better to function categories it had not seen (0.90 versus 0.35), finished in 27.7 seconds instead of 84.5, and characterized 41.8 functions per minute instead of 14.28. Under federated scaling, its verifier sustained 330,000 to 1.18 million functions per second. The authors applied it to the behavior of the HDF5 library.
Neeraj Rajesh of GRC wrote Metis with Karim Youssef and Hariharan Devarajan of Lawrence Livermore National Laboratory and Sarah Neuwirth of Johannes Gutenberg University Mainz.
Update (July 2026): On July 29 the lab released WisIO 0.2.0, its open-source tool that connects I/O bottlenecks in workflow traces to their root causes. Every Python version that 0.1.1 supported had reached end of life, so the release moves to Python 3.10 through 3.13. Small analysis views now run in pandas instead of Dask, which cut one DFTracer analysis from 68 to 19 seconds. Each reported bottleneck now states why it is severe, and a new web app lets users upload a trace and read the findings.