Faster runs, with a narrower claim
Best-effort execution was associated with markedly faster results in the reported cluster benchmarks. At 64 processes, the condition showed a 2.1-fold speedup for DISHTINY and a 12.5-fold speedup for graph coloring. Graph-coloring solution error was 73% lower under the best-effort condition. The comparison used the workloads and cluster configuration reported in the study.
The work develops a framework for measuring runtime behavior in best-effort code and applies it in two digital-evolution case studies. In the cluster case, DISHTINY used Conduit communication with configurable asynchronous send buffers. Messages could be dropped when those buffers reached capacity.
The cluster profiling used 3,600 DISHTINY cells per process and 2,048 graph-coloring nodes per process. Each SLURM allocation was run through 10 independent executions, providing repeated comparisons for the cluster benchmarks.
The average held up as the system grew
In a scaling test from 1 to 64 processes, DISHTINY retained 92% scaling efficiency under best-effort execution, compared with 47% under synchronous execution. Beyond 16 processes, best-effort performance remained stable, while synchronous execution degraded by 34%.
Quality of service, or QoS, measures showed a split picture at larger allocations. Between 64 and 256 processes, median QoS showed no detectable degradation, and straggling decreased by 4%. Maximum QoS measures, however, developed more extreme tails at the larger allocation. In practical terms, typical behavior stayed steady while the worst observed behavior became more volatile.
A localized hardware anomaly was also associated with unusual QoS behavior. Node lac-417 was associated with orders-of-magnitude more extreme QoS outliers for most measures, while median QoS remained stable. Including the node was associated with a 0.5% reduction in straggling.
The traffic analysis found incoming and outgoing traffic significantly anticorrelated in both configurations. Between-node imbalance coincided with a 12% execution-speed disparity, and that imbalance was not detectable after the speed disparity was accounted for.
A huge simulation observed through a small window
The second case study used a Cerebras Wafer-Scale Engine configuration with a 755-by-1,170 lattice, 883,350 processing elements, 44 gigabytes of aggregate on-device memory and 48 kilobytes per processing element. Instead of recording complete histories, the case used an approximate ancestry record built from sampled data.
The lineage method stored single-bit barcode markers in a fixed-capacity 64-bit region of the genome. When that region filled, new markers overwrote existing ones, maintaining an approximate ancestry record. In one reported run, the system evaluated 226.1 million agents for 5 million generations, sampled one genome per processing element about 15 times per second and yielded 9,000 snapshots over 10 minutes.
The harvested WSE genomes were thinned from nearly 8 billion to 1.1 billion, representing about 0.0001% of cumulative individuals. A localized transfer fault was associated with bit flips in around 5% of affected 32-bit words and corruption of 0.03% of extracted genomes, without compromising the on-device simulation. The finding concerns that localized fault and does not establish how other transfer faults would behave.
The record also exposed possible distortions
Phylogenetic inspection, meaning comparison of relationships among sampled lineages, found that parallel lineages disproportionately followed similar spatial paths. Many moved from the bottom to the top of the chip, suggesting a directional migration imbalance that could bias reconstructed histories. The broader frequency of the pattern and its impact after correction were not quantified.
Time-series recording showed mutator alleles reaching upwards of 99.9% prevalence in some cases even when they did not reliably fix.
The study is an arXiv preprint, version 1, dated 25 August 2026. It reports support from the Eric and Wendy Schmidt AI in Science Post-doctoral Fellowship, Schmidt Sciences, NSF programs and the U.S. Department of Energy. The findings remain computational case studies: the cluster results are tied to the reported workloads and configuration, and the wafer-scale fault was one localized event.
Paper data and sources
Original title: Trust, but Verify: Rigorously Profiling Best-Effort High-Performance Computing for Digital Evolution
Authors: Matthew Andres Moreno, Santiago Rodriguez Papa, Charles Ofria et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text