Noisy-neighbour effects are the reason I stopped trusting any single benchmark run from shared capacity. Median of five or it did not happen.

30% is generous. ive seen worse on burst instances

这个问题在共享环境里基本无解，只能多跑几次取中位数。真要稳定就得独占，但成本又完全不一样了

yeah thats where i landed. median of five and stop pretending single runs mean anything

It’s frustrating because even increasing sample size only mitigates the symptom, not the underlying unpredictability of the shared environment. For latency-sensitive inference workloads, that variance can distort outcomes significantly, which is critical when you’re trying to compare models or optimize trading strategies. Have you tried any approaches to isolate or control for neighbor interference beyond just replication?

yeah but even median of five can hide weird outliers, feels like for real stability u gotta think about layering some kind of lightweight isolation or at least profiling artifacts from noisy neighbors 🤔