The end of October 2023 was an exciting time, since the 29th ACM Symposium on Operating Systems Principles (SOSP ’23) took place in Koblenz, Germany and the 14th ACM Symposium on Cloud Computing (SoCC ’23) in Santa Cruz, CA, USA. Both conferences featured exciting keynote talks, paper presentations and poster sessions with attendees from all over the world, academia and industry.
At SOSP ’23, Margo Seltzer delivered the award talk for receiving the prestigious ACM Athena Lecturer award, that recognizes and celebrates women researchers who have made fundamental contributions to Computer Science. The topic of her talk, as described in her own words, was:
“It’s 2023, and the answer to every system performance or optimization problem is “machine learning.” But what kinds of models are appropriate for these applications? I’m going to try to convince you that, as in good system design, “simpler is better.” And, in this case, simpler has many benefits: simpler models are typically more efficient in both space and time, they are frequently transparently interpretable, and they produce accuracy and generalization equivalent to the fanciest deep learning model you can build.”
This school of thought was echoed in various talks at SoCC ’23, as well. Christos Kozyrakis made a point on how training One Model to Rule Them All can be very ambitious and challenging with potentially prohibitive costs in terms of money, time and resources for deployment. In contrast, researchers from Huawei Research in Edinburgh, showed potential for high prediction accuracy when training one global model in the context of serverless workloads (Joosen et al.).
Apart from the battle between simple vs. complex and many vs. one machine learning models, there were prominent discussions in both conferences on aiming for simplicity in the system design and avoiding unnecessary complexity, even avoiding the use of machine learning. For example, researchers from UBC, Canada, made a case on how using parametric regression can significantly simplify function rightsizing compared to black-box optimization techniques, when configuring memory to execute a serverless function (Moghimi et al.). In a different session, researchers from IMDEA Software, Spain, questioned whether it is even necessary to use machine learning to forecast cloud resource usage. Their answer is no, at least for the most part, since cloud resource usage exhibits high data persistence over time, meaning that the values do not change significantly in a short time window, e.g., every 5 minutes (Christofidi et al.).
All these insights and conversations reminded me of the K.I.S.S. principle, first noted around 1938, according to which systems work best when they are designed to be simple, avoiding unnecessary complexities. The K.I.S.S. acronym comes across in several variations, originally stated as “Keep it Simple, Stupid!”. In the context of machine learning in computer systems (MLSys), I think it’s more suitable to K.I.S.S.: Keep it Simple, Smart!
Why smart? Because we need to use our human intelligence to identify clever and insightful ways that will enable simple approaches to be effective amidst the increased complexity of systems, user behaviors and observed patterns.
However, what is considered as simple can be subjective. According to the presentations and conversations at SOSP ’23 and SoCC ’23, what is perceived as a simple approach, seems to range across heuristics, data-driven solutions, statistical methods like linear regression models and machine learning techniques such as decision trees.
In conclusion, keeping it simple doesn’t necessarily make things easier. It creates a need for extensive analysis and exploration to extract insights that will enable a simple, yet smart, system design to be effective.
About the author: Thaleia Dimitra Doudali is an Assistant Professor at IMDEA Software Institute in Madrid, Spain, working at the intersection of machine learning and computer systems. She received her Ph.D. from Georgia Tech, USA, advised by Ada Gavrilovska.
Acknowledgement: Kudos to Ada Gavrilovska for coming up with the rephrase of KISS: Keep it Simple, Smart!
Disclaimer: The author only attended SoCC ‘23 in person, thus most paper references are from presentations and conversations at that venue.
Picture Credit: Thaleia Dimitra Doudali
Editor: Dong Du (Shanghai Jiao Tong University), Tianyin Xu (University of Illinois at Urbana-Champaign)