A Nose for a Node, PerfectScale Streamlines Kubernetes Optimization
The inherently distributed and interwoven complexity that typifies modern Kubernetes environments demands constant optimization. It’s a simple fact of cloud-native life. There’s always scope to manage resources more efficiently, more cost-effectively and securely within the realm of whatever governance mandate oversees any given deployment.
Aiming to provide the requisite levels of constant Kubernetes optimization control needed in a continuous and autonomous way is PerfectScale. The company’s Kubernetes optimization and governance platform is aligned to serve platform engineering teams across DevOps environments with tools to improve the stability of what it calls a “complete K8s environment” in live production. But what is Kuberenetes completeness in this context?
How Low Can You Node?
In this case, we’re talking about Kubernetes optimization which has now reached a level of advanced node-level optimization. This granular lower-level control alongside upcoming GPU support is hoped to streamline Kubernetes performance so that teams can optimize resources intelligently across their entire stack.
Optimization at this level requires management tool technologies capable of efficiently allocating resources (CPU power – and now GPU power too, plus memory and storage) to individual Kubernetes “worker nodes” so that workloads are more evenly distributed, appropriately sized and allocated to nodes that may be optimized in specific ways (with specific hardware or software configurations) all with a view to cost-performance efficiency. Also including system controls to manage container “rightsizing” as well as prioritization and scheduling, this practice also straddles “node affinity” i.e. the implementation of rules that control pod placement to specific Kubernetes nodes for the same optimization objectives. That’s a lot of what PerfectScale will be doing here.
The company points to studies that talk about Kubernetes resource waste, often occurring as a result of over-allocated node space. Whether one study finds 30% wastage and another 51% waste isn’t important, what matters is the fact that an inherent allocation inefficiency is not uncommon.
Pinpointing Idle Capacity
PerfectScale offers its branded Infrafit recommendations service to provide data-driven insights into node allocation and utilization. The concept here hinges on helping organizations to minimize waste, reduce cloud costs and decrease their environmental impact. Now, with the latest release of Infrafit announced during KubeCon + CloudNativeCon America 2024 in Salt Lake City this November, organizations gain tailored recommendations on optimal node types and resource configurations. This means they can pinpoint idle capacity and plan budgets effectively.
“PerfectScale’s flagship solution, PodFit, is designed to autonomously right-size and scale Kubernetes workloads, empowering our customers to reduce costs and enhance system resilience,” said Amir Banet, CEO and co-founder of PerfectScale. “But our mission goes further – to simplify the complexities of ongoing day-2 Kubernetes operations. With the addition of InfraFit recommendations, we now offer a solution for full-stack Kubernetes cost and performance optimization.”
PerfectScale’s InfraFit and PodFit solutions integrate with ClusterAutoscaler (a component that automatically adjusts the size of a Kubernetes Cluster so that all pods have a place to run and there are no unneeded node) and Karpenter (a well–known technology service that works to enhance node autoscaling efficiency) for additional cost savings.
Advanced GPU Utilization
As organizations increasingly adopt AI and LLMs to power applications, GPU usage has surged, driving up cloud costs and creating new complexities in resource management.
“GPU represents a significant investment for organizations, accounting for up to 75% of hourly cloud costs,” said Eli Birger, CTO and co-founder of PerfectScale. “However, managing GPU resources, including their allocation and optimization, presents a complex challenge arising from the need to balance resource distribution while ensuring workload performance. Misconfiguration can lead to system instability, making it essential for organizations to effectively monitor and adjust GPU usage without compromising the efficiency and reliability of workloads.”
PerfectScale’s upcoming GPU support announced at KubeCon + CloudNativeCon America 2024 is designed to provide real-time visibility into GPU allocation and utilization for each workload on a node, identifying idle resources and highlighting significant cost-saving opportunities. The company says that this new capability sets the stage for enhanced data-driven GPU allocation recommendations and automated optimization, ensuring that organizations can effectively balance performance and efficiency in their workloads.
PerfectScale invites early adopters to join the waitlist for exclusive early access to these GPU optimization features.
