Using Site-Level Modeling to Evaluate the Performance of Parallel System Schedulers

Abstract

Abstract

Workload Modeling for Computer Systems Performance Evaluation

“…For example, a model can include an explicit feedback loop in which system performance affects subsequent load generation [613]. A log only contains a list of load items, and we don't know if and how they may have depended on whatever happened before.…”

“…This is an extension of the well-known distinction between open and closed system models: in open models new work is completely independent of system conditions, whereas in closed models new work arrives only as a direct consequence of a previous termination. But in real life the situation is more complicated, and feedback from the system's performance to the users can affect the generation of additional work in various ways [613,614].…”

“…A session is then a sequence of job submittals, and a break in the sequence is interpreted as a break between sessions [34,617,74,613,600,384]. However, jobs always have some gaps between them, even if they are part of the same session.…”

“…In such a scenario the user may submit several jobs concurrently. One possible interpretation in this case is to consider interarrival times between the jobs that make up a single batch, and think times between the batches [613,747]: However, the main problem is handling long jobs. Consider a case where a user's activity includes a very long job, say one that runs overnight.…”

“…This is a self-regulating effect that reduces the risk of system saturation, and may be regarded as part of online behavior [241]. Thus models with feedback may lead to dramatically different results regarding system behavior under load; specifically, oblivious models will tend to generate pessimistic predictions, whereas models with feedback will be more optimistic regarding system robustness [613,600,687,551,750].…”