(v2026-spot-gpu): constant oversubscription of A10 GPUs

regarding [cluster-trace-v2026-spot-gpu](https://github.com/alibaba/clusterdata/tree/master/cluster-trace-v2026-spot-gpu):

I want to use this dataset for trace-driven simulations, however I am encountering issues due to large amounts of pending workloads when using this trace. When analysing, I noticed that the demand for certain GPU types is higher then the available capacity in the cluster. The worst case is by far A10 GPUs which are drastically oversubscribed throughout the entire trace.

I calculated the oversubscription using `gpu * instances` occupation in `[arrival, arrival + duration)`. This causes more and more workloads to pile up in the queue which would only drain at the very end of the trace. The corresponding paper lists an allocation rate of `84.59%` for A10 GPUs, my calculations yield `133%` on average.

Could you please clarify how the allocation rate in the paper was calculated? Maybe I am missing something or was there some kind of preemption or failures which would have caused the workloads to not run their entire duration?

Start with the cluster-trace-v2026-spot-gpu dataset and reproduce the reported A10 allocation rate using GPU, instances, arrival, and duration. Compare the calculation with the paper's methodology and determine whether preemption or failures explain the discrepancy; done means documenting a clear explanation for the 84.59% versus 133% results.

Written by the indexing model from the issue text.