Slide 1

Slide 1 text

Hyper-Efficient Serverless Platforms on Azure Kubernetes Service with Fermyon Platform for Kubernetes

Slide 2

Slide 2 text

Kubernetes Efficiency Pain Points Actual resource utilization vs requested resources Requested resources vs allocatable resources low demand should drive down scale Workload Rightsizing 01 Demand Based Downscaling 02 Cluster bin packing 03 R E S O U R C E S A P P L I C A T I O N D E V E L O P E R P L A T F O R M A D M I N From Google Cloud's ‘State of Kubernetes Cost Optimization’ P A I N P O I N T

Slide 3

Slide 3 text

P A I N P O I N T Containers & Serverless have a cold start delay problem. Time to first “serverless byte”

Slide 4

Slide 4 text

Containers are often complex. P A I N P O I N T Many layers of operational dependencies – Kernel, Drivers, OS + Utilities than need to be deployed to along with your code.

Slide 5

Slide 5 text

Containers can be very expensive. P A I N P O I N T Common practices to optimize containers lead to over-provisioning and over-consuming resources.

Slide 6

Slide 6 text

Small package sizes Startup time Portable Sandboxed

Slide 7

Slide 7 text

Developer tool for building and running event-driven serverless WebAssembly applications Spin is an open-source project, built with open standards like WASI and the WebAssembly Component Model. spin new spin build spin up 4.8K GITHUB ★ 75+ CONTRIBUTORS github.com/fermyon/spin

Slide 8

Slide 8 text

Spin Apps Run Everywhere

Slide 9

Slide 9 text

Lab Configuration Single Node AKS Azure Region: Frankfurt Standard_D4ds_v5 4 CPUs 16 Gigs Memory AMD64 Max Pods Per Node 250 $271.56 monthly Fermyon Platform for Kubernetes AMD64 Single Node AKS Azure Region: Frankfurt Standard_D4ds_v5 4 CPUs 16 Gigs Memory AMD64 Max Pods Per Node 250 $271.56 monthly Azure Functions on Kubernetes AMD64 Single Node AKS Azure Region: Frankfurt Standard_D4ds_v5 4 CPUs 16 Gigs Memory AMD64 Max Pods Per Node 250 $271.56 monthly SpinKube AMD64

Slide 10

Slide 10 text

Benchmark Constraints • Deploy as many Workloads as possible • Cluster must handle 120.000 requests per hour • 20k requests every 10 minutes • Apps are requested randomly • Cluster may not run into problems like • PID or IP exhaustion • Out-Of-Memory exceptions

Slide 11

Slide 11 text

Demo Let’s explore the clusters

Slide 12

Slide 12 text

No content

Slide 13

Slide 13 text

Conclusion Cluster Workloads Avg. Response Time Costs per Workload Azure Functions on AKS 35 3.353ms $7.76 SpinKube on AKS 135 2.383ms $2.01 Fermyon Platform for Kubernetes on AKS 3,500 2.529ms $0.06 Hyper-Efficiently packed and performant WebAssembly with Spin and Fermyon Platform for Kubernetes leads to lower cost

Slide 14

Slide 14 text

Hyper-efficient serverless on Kubernetes, powered by WebAssembly

Slide 15

Slide 15 text

Spin https://fermyon.com/spin SpinKube https://spinkube.dev Fermyon Platform for Kubernetes https://www.fermyon.com/platform Resources

Slide 16

Slide 16 text

Spin Operator A Kubernetes operator to deploy Spin applications. Containerd-shim-spin A ContainerD shim for running Spin Applications. Runtime-class manager A simple way to install Wasm runtimes. Spin Kube plugin A Spin plugin for interacting with Kubernetes.

Slide 17

Slide 17 text

Hyper-efficient Application Density 5.000 apps/node with instant response-time Simplified Operations No need for complicated scaling rules or resource reservations Developer Portal Experience UI to simplify app management Cloud Vendor Integrations KV Store, SQL, Blob, and LLM integrations