A paper published on arXiv on June 19, 2026 describes ARGUS, which combines CPU, framework, and GPU-kernel data to diagnose slowness, reporting continuous overhead below 2%.
A paper published on arXiv on June 19, 2026 describes ARGUS, a system for diagnosing slowness in large-scale training. Its approach combines information from CPU stacks, framework semantics, and GPU kernel data. The text reports continuous overhead below 2% and compression of kernel events by about 3,700 times.
The work highlights potential value for engineers running distributed training who want to identify stragglers without keeping detailed profiling active at all times. To check the scope, method, and findings, consult the original paper on arXiv and compare its metrics with the conditions and definitions stated by the authors.