Computing Response Time Distributions Using Iterative Probabilistic Model Checking

Abstract

System designers need to have insight in the response times of service systems to see if they meet performance requirements. We present a high-level evaluation technique to obtain the distribution of services completion times. It is based on a high-level domain-specific language that hides the underlying technicalities from the system designer. Under the hood, probabilistic real-time model checking technology is used iteratively to obtain precise bounds and probabilities. This allows reasoning about nondeterministic, probabilistic and real-time aspects in a single evaluation. To reduce the state spaces for analysis, we use two sampling methods (for measurements) that simplify the system model: (i) applying an abstraction on time by increasing the length of a (discrete) model time unit, and (ii) computing only absolute bounds by replacing probabilistic choices with non-deterministic ones. We use an industrial case on image processing of an interventional X-ray system to illustrate our approach.

Efficiently Computing Latency Distributions by Combined Performance Evaluation Techniques

“…We use advanced model checking, which is inherently hard (as shown in [18]), mainly because of two reasons. First, the model is complex because it has both nondeterministic and probabilistic traits: nondeterminism occurs when two atomic processes try to access a resource at the same time, and when to decide whether atomic process "Refine" is executed either once or twice.…”

“…The minimum simulated result is an upper bound for the lower bound, and vice versa. To this end, we apply the MODES simulator [9] to a Modest model, derived from an iDSL model (as in [19,18]). …”

“…Exhaustive methods, like (probabilistic) model checking [18], do potentially provide the required accuracy and flexibility in model choice. However, they do not scale well and typically require many computational resources and time.…”

“…In previous work, we developed a method to generate latency distributions via iterative probabilistic model checking [18] for iDSL [20,19], a high-level language and tool chain for performance evaluation of service systems. It delegates low-level performance queries to the Modest toolset [9].…”

“…First, in the method [18] the user had to determine the right model abstraction manually. We have automated this by evaluating the execution time of iDSL for various model abstractions and then selecting the best model.…”