Request Location.

It is well known that the latency of a request can vary from a fraction of a millisecond to tens of milliseconds, based on its location compared to previous requests, as well as caching policies at the array. Variability in seek and rotational delays can cause an order of magnitude difference in service times. This makes it difficult to estimate the baseline IO latency corresponding to the latency with no queuing delay. Thus a sudden change in average latency or in the ratio of current values to the previous average may or may not be a signal for overload. Instead, we look at average latency values in comparison to a latency threshold $\cal {L}$ to predict congestion. The assumption is that latencies observed during congestion will have a large queuing delay component, outweighing increases due to workload changes (e.g., sequential to random).