More is Less in Kieker? : The Paradox of No Logging Being Slower Than Logging

Reichelt, David Georg and Jung, Reiner and Hoorn, André van and Hoorn, André van (2023) More is Less in Kieker? : The Paradox of No Logging Being Slower Than Logging. In: 14th Symposium on Software Performance :. UNSPECIFIED, DEU.

Text (ReicheltJungVanHoorn2023MoreIsLessInKieker-camera-ready) - Published Version
Available under License Unspecified.
Download (0B)

Text (ReicheltJungVanHoorn2023MoreIsLessInKieker-camera-ready)
ReicheltJungVanHoorn2023MoreIsLessInKieker-camera-ready.pdf - Published Version
Available under License Unspecified.
Download (240kB)

Abstract

Understanding the sources of monitoring overhead is crucial for understanding the performance of a monitored application. The MooBench bench mark measures the monitoring overhead and its sources. MooBench assumes that benchmarking overhead emerges from the instrumentation, the data collection, and the writing of data. These three parts are measured through individual factorial experiments. We made the counter-intuitive observation that MooBench consistently and reproducibly reported higher overhead for Kieker and other monitoring frameworks when not writing data. Intuitively, writing should consume resources and therefore slow down (or, since is parallelized, at least not speed up) the monitoring. In this paper, we present an investigation of this problem in Kieker. We find that lock contention at Kieker’s writing queue causes the problem. Therefore, we propose to add a new queue that dumps all elements. Thereby, a realistic measurement of data collection without writing can be provided.