In a paper published in Cluster Computing, Jasmine Kaur, Inderveer Chana and Anju Bala introduce an enhanced version of a serverless simulation framework that adds a virtual machine layer between physical servers and the containers that actually run user code. On top of that more realistic architecture, they bolt on a multi-agent reinforcement learning technique known as proximal policy optimization, or PPO, to decide where workloads should go. The results are striking: cold start latency drops by up to 29 percent, the number of active physical machines falls by as much as 66 percent, and energy consumption shrinks by up to 23 percent. For an industry under mounting pressure to curb the enormous electricity appetite of data centers, those numbers carry real weight.

To understand why the virtual machine layer matters, it helps to look at how serverless platforms are actually built. In principle, containers, lightweight packages of software that share the host operating system’s kernel, could run directly on physical machines. That is precisely how ServlessSimPro, the earlier simulation platform on which the new work builds, modeled the world. But in production deployments at major cloud providers and in popular open-source platforms such as Knative and OpenFaaS, containers are almost always nested inside virtual machines. The VM adds an extra boundary of isolation and security, but it also introduces an additional layer of resource management, with its own scheduling decisions, consolidation opportunities and performance overheads.

By ignoring that layer, earlier simulators produced results that diverged from what operators actually observe. A scheduling algorithm that looks brilliant in a flat, container-on-metal simulation may behave very differently when it must first choose a virtual machine, and only then a container within it. The enhanced framework bridges this architectural gap, giving researchers a testbed whose resource abstractions mirror the deployment models used in the real world. That matters because simulation is where most scheduling research lives: few academic groups can run experiments on hyperscale infrastructure, so the fidelity of the simulator effectively determines how transferable their ideas are.

The second half of the contribution is the scheduling brain itself. Rather than a single centralized controller making every placement decision, the researchers deploy multiple reinforcement learning agents, each powered by proximal policy optimization, working in parallel across the simulated cluster. PPO is a policy-gradient method that has become a favorite of the reinforcement learning community because it makes steady, clipped updates to its decision policy, avoiding the catastrophic performance collapses that can plague more aggressive learning algorithms. Each agent is responsible for a slice of the scheduling problem: deciding where containers should be placed, when virtual machines should be consolidated onto fewer physical hosts, and how incoming workloads should be distributed across the fleet.

Decentralization is not just an aesthetic choice. In a large data center, a single scheduler becomes a bottleneck and a single point of failure, and the state of the system grows so large that no one agent can see everything at once. Multi-agent architectures let decisions happen closer to the resources they affect, and the multi-agent formulation of PPO, often abbreviated MAPPO, allows the agents to learn coordinated behavior despite their partial views. The approach builds on the authors’ earlier work on multi-agent deep Q-learning for serverless job scheduling, but PPO’s more stable training dynamics make it better suited to the high-dimensional, continuous decisions involved in juggling containers, VMs and workloads simultaneously.

The headline metric, cold start latency, deserves particular attention because it is the Achilles’ heel of serverless computing. When a function is invoked and no warm instance exists, the platform must provision a container from scratch, an operation that can add hundreds of milliseconds or more to response time. For latency-sensitive applications, from interactive web services to real-time data pipelines, those milliseconds are the difference between a seamless experience and a visibly sluggish one. By learning placement and consolidation policies that keep likely-to-be-needed functions warm and co-located with available capacity, the MAPPO-driven scheduler cuts cold starts by nearly a third in the reported experiments.

The efficiency gains are equally significant. Consolidating workloads so that up to 66 percent fewer physical machines need to stay active is a dramatic reduction, and it cascades directly into the 23 percent cut in energy consumption. Data centers are among the fastest-growing consumers of electricity worldwide, and idle servers draw substantial power even when they process nothing. Any technique that lets operators switch off more of the fleet, without degrading the responsiveness users feel, translates into both lower costs and lower carbon emissions. The fact that the framework achieves these savings while simultaneously improving latency suggests the agents are finding genuinely smarter trade-offs, not merely shifting costs from one column to another.

The study also situates itself within a rapidly maturing ecosystem of simulation tools. Open-source options such as FaasSim, FaaS-Sim and SimFaaS have given researchers starting points for modeling function-as-a-service platforms, and commercial offerings from AWS, Microsoft Azure, Google Cloud and IBM have defined the de facto behavior the simulators try to capture. What has been scarce, the authors argue, is a comprehensive platform that combines realistic multi-layer resource modeling with the ability to train and evaluate learning-based schedulers. By pairing the VM-aware architecture with MAPPO, the framework becomes both a measurement instrument and a training ground, letting researchers prototype scheduling policies that can be benchmarked consistently before any real-world trial.

The broader significance of the work lies in the convergence of two trends: the industrial consolidation around serverless architectures and the rapid advance of multi-agent reinforcement learning as a practical tool for systems management. As edge computing, Kubernetes-based orchestration and energy-aware provisioning continue to reshape how cloud infrastructure is organized, the ability to simulate those systems faithfully, and to let learning agents discover scheduling policies that human engineers might never hand-design, could define the next generation of cloud efficiency. The Thapar Institute team’s framework offers the research community a way to explore that future at scale, on a laptop rather than a data center, and with an architectural honesty that earlier tools lacked. If the reported gains hold when such policies migrate from simulation to production clusters, the invisible machinery of the cloud may soon be run by algorithms that learned their craft in a simulator, one carefully modeled virtual machine at a time.

Subject of Research: VM-aware serverless simulation and multi-agent reinforcement learning for cloud job scheduling

Article Title: VM-aware serverless simulation framework with multi-agent proximal policy optimization for scalable and efficient job scheduling

Article References: Kaur, J., Chana, I., & Bala, A. (2026). VM-aware serverless simulation framework with multi-agent proximal policy optimization for scalable and efficient job scheduling. Cluster Computing, 29(13), Article 758. https://doi.org/10.1007/s10586-026-06585-w

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06585-w

Keywords: serverless computing, simulation framework, virtual machines, multi-agent reinforcement learning, proximal policy optimization, job scheduling, cold start latency, energy efficiency, cloud computing, resource utilization, container orchestration, data centers