Renewable energy is an important step on the way to fight climate change. The energy produced by burning fossil resources is one of the main drivers of carbon emissions. But running a datacenter on renewable energy all the time is difficult. Usually - with only a few exceptions - your workloads do not run on renewable energy.
Martin Lippert spoke about using running workloads on renewable energy at OOP 2023 Digital.
Most sources of renewable energy are not available all the time and do not produce energy at a constant level all the time, Lippert said. The amount of energy that wind and solar produce fluctuates over time, and this usually doesn’t fit well with the moments in time when we consume the energy, he added. It is a huge challenge to align the production and consumption of renewable energy as closely as possible so that you can relinquish other energy sources:
The sentence that the datacenter uses, "100% renewable energy," usually only means that the operating company buys as much renewable energy as the datacenter consumes in total. It does not mean that the datacenter is fed all the time with enough renewables as it needs at that moment.
We don’t produce enough renewable energy for the overall energy needs, and the total energy consumption still goes up, Lippert said. Even if we dramatically increase the amount of renewable energy that is being produced, we might not get there fast enough.
Lippert suggested tackling the problem from multiple sides: more renewable energy production, less energy consumption, and more flexible workloads.
We need to think about workloads that are flexible with regards to when they run, where they run, and if they run at all, Lippert said. The more flexible the workloads are, the easier it is for a datacenter to save energy, the easier it is to align renewable energy availability with energy consumption, he mentioned:
Think about an option to automatically run a workload in the data center that has the most renewable energy available at a specific moment in time. Or to run batch workloads when renewable energy is available instead of at a fixed date and time. Or to not run workloads at all if they are not used.
A widely used technology stack for building business applications is Java and Spring Boot, Lippert said. While in the past, starting up those applications took many seconds, sometimes even minutes, this has changed dramatically, as he explained:
The GraalVM native image technology and the corresponding support for this technology in Spring Boot allows you to compile Spring Boot applications ahead-of-time into native executables that startup in milliseconds. This opens the door for probably millions of business applications around the globe to be more sustainable and more flexible with regards to scale-to-zero architectures and platforms.
The way you write and operate software has an impact on carbon emissions and climate change. Let’s not wait for others to solve the problem, let’s write greener software instead, Lippert concluded.
InfoQ interviewed Martin Lippert about handling zombie and idle workloads.
InfoQ: What problems do zombie workloads cause?
Martin Lippert: The term "zombie workloads" originates from the study zombie/comatose servers redux that analyzed a number of datacenters and their deployed workloads. In that study, they called workloads that were deployed, but never used, "zombies". Interestingly, the study revealed that zombie workloads are not a niche problem. They found between 1/4 to 1/3 of all deployed workloads to be zombies. That is huge.
Zombie workloads consume energy and hardware resources. Identifying those zombie workloads and removing them could save huge amounts of energy. The cloud management software from VMware, for example, has features built-in to automatically identify zombie workloads. And you will find similar features in other offerings as well. Automation is key here. You can’t do that manually - at least not at scale.
InfoQ: The study also identifies idle workloads, those that are used only sometimes. What can we do about them?
Lippert: We can’t delete those idle workloads and for some of those workloads, we can’t even predict exactly when they will be used and when not. One approach here could be to adopt so-called "scale-to-zero" architectures and technologies that allow you to run those workloads on demand only.
This is obviously not very straightforward for many workloads that we have today. Imagine a workload that takes many seconds, if not minutes, to startup. You can’t start up such a workload on demand quickly enough to serve the users needs. But there are architectures and technologies in place that allow you to refactor applications into smaller pieces and to turn them into applications that indeed startup in milliseconds - which seems to me good enough for many on-demand approaches.