Detecting Bottleneck in n-Tier IT Applications Through Analysis

Abstract

Abstract.As the complexity of large-scale enterprise applications increases, providing performance verification through staging becomes an important part of reducing business risks associated with violating sophisticated service-level agreement (SLA). Currently, performance verification during the staging process is accomplished through either an expensive, cumbersome manual approach or ad hoc automation. This paper describes an automation approach as part of the Elba project supporting monitoring and performance analysis of distributed multi-tiered applications that helps in bottleneck detection. We use machinelearning to determine service-level objectives (SLOs) satisfaction and locate bottlenecks in candidate deployment scenarios. We evaluate our tools with TPC-W, an on-line bookstore, and RUBiS, an on-line auction site.

Bottleneck Detection Using Statistical Intervention Analysis

“…For any workload w ∈ WS an impact assessment model for the first difference of any metric value Y w can be formulated in terms of Equation 3. Note that we use the first difference as approximation of the first derivative consistently with our findings in [4].…”

“…If a SLA was not met (SLO-satisfaction drops significantly) in a certain scenario, a three-step detection process began: staging the system with varying workloads and collecting performance data from system-level and applicationspecific metrics, training a machine learning classifier with the data, and finally querying the trained machine learning classifier to identify potential bottlenecks. Please refer to [4] for more details. While our three-step methodology proved to be successful, it mainly relies on machine learning algorithms to execute the final performance modeling and classification.…”

“…Secondly machine learning classifiers produce a model that is not necessarily interpretable in a trivial manner. We discussed suitable interpretations in [4]. Nevertheless, this led to a residual degree of uncertainty in the interpretation of the analysis results.…”

“…Individual satisfaction levels are determined for each of the SLO components. Interested readers can refer to [4] and [11] for more details. For simplicity reasons we solemnly use the response time of the system from the SLO-evaluator in the presented analysis.…”

“…This approach is distinct from our previous work that used machine learning [4]. We introduce a deterministic algorithm, which has proven to be very effective in the Elba environment.…”