Running I/O intensive workloads on Kubernetes, by Nati Shalom
1 like526 views
Kubernetes was originally targeted for running large scale web applications. I/O intensive workload represents a class of high-end applications such as network services, trading applications, database services that require high-speed access to hardware resources and often users specific hardware or CPU features to maximize their performance.
Ad
More Related Content
What's hot (20)
Similar to Running I/O intensive workloads on Kubernetes, by Nati Shalom (20)
Ad
More from Cloud Native Day Tel Aviv (20)
Ad
Recently uploaded (20)
Running I/O intensive workloads on Kubernetes, by Nati Shalom
- 1. 1 All Clouds Any Service Unlimited Locations Running I/O Intensive Workload On Kubernetes November 2019
- 3. 33 I/O Intensive Workload Defined Networking Low latency ML Application HPC Application
- 4. 44 Optimizing Kubernetes I/O Performance Optimize the container to host performance ratio The focus of this presentation
- 5. 55 Normalising the workload management I/O Intensive Workload Standard Application Highly Optimized Infrastructure Environment Standard Public Cloud Highly Optimized Public Cloud Resources Common automation through CI/CD pipeline EaaS The focus of this presentation QA vs Production Environment Best Cost/Performance
- 6. 6 Optimizing Kubernetes Performance through H/W acceleration
- 7. 77 Kubernetes H/W Acceleration Features HW Platform Intel® Ethernet Network Adapter Intel® QuickAssist Technology Intel® DC SSD Intel® Programmable Acceleration Card ▪ Multiple network interfaces for VNFs ▪ High performance Data Plane (E-W, N-S) ▪ Node Feature Discovery - platform capabilities ▪ Device Plugin (SR-IOV, Intel® QAT, FPGA…) ▪ CPU Core-Pinning and isolation for K8s pods ▪ Huge Page allocation ▪ Topology Manager - Guarantee NUMA alignment ▪ Platform Telemetry ▪ Installation playbook
- 9. 99 Benchmark Use CaseAppsSWPlatform Orchestration HW Platform Policy Driven for Multi-Cloud Services HTTPS Bench Intel® Ethernet Network Adapter Intel® QuickAssist Technology Intel® DC SSD Intel® Programmable Acceleration Card GitHub ▪ Multiple network interfaces for VNFs ▪ High performance Data Plane (E-W, N-S) ▪ Node Feature Discovery - platform capabilities ▪ Device Plugin (SR-IOV, Intel® QAT, FPGA…) ▪ CPU Core-Pinning and isolation for K8s pods ▪ Huge Page allocation ▪ Topology Manager - Guarantee NUMA alignment ▪ Platform Telemetry ▪ Installation playbook TOSCA Blueprint Crypto Crypto Open sourced, up-streamed Reference Implementation for Kubernetes* Scaled through Intel® Network Builders and Intel® Select Solutions
- 10. 10 I/O Optimised K8S Cluster Hardware Architecture QAT Enabled NodeStandard Node
- 11. 11 Phase 1 – Setting up the I/O Optimised K8S Environment 11 Running I/O Intensive cloud native workload on private cloud while being connected to data services on public cloud The Environment Includes: ○ Multi Cloud: AWS / Azure , OpenStack ○ Multi Cloud Networking. ○ Kubernetes Cluster Optimised for I/O Intensive workload ○ DBaaS - on VM’s EaaS - Environment As a Service I/O Intensive Cloud Native Stack AWS VPC, Load Balancer Azure US East, West Dev, Test, Prod OpenStack Kubernetes DBaaSAnsible VPN Private
- 12. 1212 https://youtu.be/PxRxbhA6gCI Creating the Environment High Level Overview
- 13. 1313 Leveraging Kubernetes Node Feature Discovery ○ Each node is labeled with a list of hardware capabilities ○ APP A —> SRIOV & Turbo Boost is supported only by node 2. Node Feature Discovery in Kubernetes Cluster Master Node 1 Node 2 SR-IOV Application A Turbo Boost node selector Node 2 NFD DISCOVERY POD SR-IOV Turbo Boost NUMA QAT IOMMU Application B Node 1 NFD DISCOVERY POD QAT Turbo Boost IOMMU QAT Application B IOMMU node selector Application AKubernetes Node labels
- 14. 14 Expanding to Public Cloud Workloads placement utilizing Enhanced Platform Awareness(Intel EPA)
- 15. | ©2019 F5 NETWORKS15 Public Cloud KUBERNETES CLUSTER QAT Node Node BIG-IP Virtual Edition(s) F5 CONTAINER INGRESS SERVICES Inspecting Traffic K8s Node Labels COTS/COTS+ matchmaking Ingress examples Security (PAC, DDOS) Encrypt/Decrypt (QAT) Traffic steering (at F5 booth) Traffic re-encrypted (using Intel® QAT offload) and sent to another destination in public cloud Encrypted Unencrypted QuickAssist Technology <none> EPA, DPDK , AWS ENA Demo setup Cloudify Leverages Node Feature Discovery to Optimize NGINX Service Placement Node feature discovery (NFD) Apache Benchmark Apache Benchmark Server
- 16. 16 Node labels & POD placement using “nodeSelector” 16 Node Capabilities Application (POD) Requirements QAT Enabled Node Capabilities QAT Enabled POD
- 18. 18 Phase 2 – Placement Policy 18 NGINX deployed on QAT K8S node NGINX deployed on any K8S node
- 19. 1919 https://youtu.be/3WlLg6u6ef4 QAT Enabled Environment Detailed Overview
- 20. 20 Hardware Architecture Performance Results 2X Performance Gain
- 21. 2121 Normalising the workload management I/O Intensive workload Standard Application Highly Optimized Infrastructure Environment Standard Public Cloud Highly Optimized Public Cloud Resources Common Automation through CI/CD pipeline EaaS The focus of this presentation QA vs Production Environment Best Cost/Performance
- 22. 2222 EaaS - Environment As a Service 22 EaaS - Environment As a Service Web App Environment HPC Environment Analytics Environment Dev/ QA/ Prod F5, VPC, Express Route CI / CD US East, West Dev, Test, Prod Palo Alto, V-WAN US East, West, Asia Prod Dedicated Interconnect US East, West, Europe, Asia Prod One Pipeline, Multiple Environments Decoupling the CI/CD pipeline from the infrastructure environment. Package each environment into high level reusable services Each environment serves many applications We call it Environment As a Service - EaaS Intelligent automation - reduced deployments by 98%, from 5 days to 2.5 hours. Achieved 60% More Velocity & Accelerated Time to Value with Cloudify. (FiCO) The Results:
- 23. 2323 Why Cloudify ? 23 Cloudify Is leading at Orchesting Heterogeneous and Distributed Services Multi Cloud Multi Site Clusters Deployments Edge DistributedCentralised Heterogeneous (Multi site/Cloud) Homogeneous
- 24. 242424 Key Customers and Strategic Alliances 24
- 25. 2525 Thank You.