Nikolas Herbst is now working as a post-doctoral researcher at the University of Würzburg at the Chair of Computer Science II. He served program committee co-chair for the ICPE in 2022. Moreover, he is active in the SPEC Research group as elected member of the steering committee and vice-chair of the Cloud working group, here co-organizing multiple editions of the HotCloudPerf ICPE workshop series.
Here is what Nikolas had to say about their work and the future of performance engineering:
Could you briefly summarize the core contribution of your paper and what you are most proud of regarding its impact?
The paper analyzed time series forecasting methods, available at that time in the R Forecasting package, concerning its strengths and weaknesses, preconditions and time-to-result expectations. The overarching goal was to achieve timely and accurate forecasts of request arrival rates on an ongoing basis. Such an online forecasting service was needed for proactive elastic scaling to match the fluctuating demand of dynamic real-world workloads with seasonal patterns, trends and bursts. The paper implemented a decision tree to select ensemble subsets of the available forecasting methods depending on the available length, resolution and further characteristics of the arrival rate time series.
After the conference, an extended version of the paper was published in a special issue of the Wiley Concurrency and Computation Practice and Experience journal. DOI: https://doi.org/10.1002/cpe.3224
The work was my very first paper submission as the result of my diploma thesis at the KIT (Descartes Research Group at the Software Quality Department) and in fruitful collaboration with IBM R&D labs in Böblingen. The MIP award honors ten years later had and still have very special impact on me personally giving motivation to keep pursuing a research career.
Looking back after ten years, did the impact or evolution of the work surprise you in any way? Is there anything you would approach differently if you were to do the work today?
Yes, the impact of the work in the community is still a great surprise for me, because the usability of the tool artifact was not optimal bridging between R and Java code. Also, parts of the concept felt a bit arbitrary, like selected threshold parameters. Yet, the work inspired the next building blocks of my dissertation, namely 1) LIMBO, an editor for dynamic workloads, 2) BUNGEE, metrics and a benchmark framework for elastic cloud systems, and 3) CHAMELEON, a proactive auto-scaler.
Already a few years later, we approached the same problem in a different way, by decomposing the time series before applying forecasts. This resulted in the TELESCOPE forecasting tool and corresponding publications between 2018 and 2020, also integrating machine learning techniques like tree boosting. And today in the AI era, the toolset for time series forecasting has again changed significantly.
If you could send one message back to yourself while writing this paper, what would it be?
The message would be short: Decompose the time series before applying forecasting techniques.
What exciting new challenges are you and your team focusing on at the moment?
In application domains like earth observation, resource-hungry workflows are executed to process satellite data to gain various important insights. Such scientific workflows are often platform dependent, monolithic scripts with low maintainability. We see huge potential to increase re-usability of workflow building blocks by helping domain researchers to adopt principles of component-based software engineering towards low- or no-code workflow engineering. Well-packaged workflow building blocks annotated with performance relevant information have the potential to boost workflow execution efficiency enabling a self-aware orchestration.
In your opinion, what emerging trends will dominate the field over the next few years?
Quantifying and taming the resource footprint of emerging workloads, especially AI services, towards matching the real value of running it will keep performance engineers busy for the next years.
Will we see you at ICPE 2027 and what is your top advice for junior researchers submitting to ICPE 2027?
Yes, I plan to be around at ICPE 2027 and looking forward to engage. My advice to junior researchers is: Do not fall the temptation to overly use generative AI beyond letting you help in small, controlled increments that can be fully understood and later defended.
What is one ICPE memory that has stayed with you?
I carry countless ICPE memories from great physical conference editions starting in Karlsruhe as a student helper, travelling to Prague, Dublin, Berlin, L’Aquila, Coimbra, London and Florence. To name one, it was definitely the MIP award talk and honors in Coimbra in 2023. It was not only special as the 1st physical conference for me after the pandemic. The venue was surrounded by beautiful gardens where walking, re-powering and discussing was possible simultaneously during the session breaks topped by Michelin-quality lunch-times.