Infrastructure at CERN Scale - DevOps Porto

Transcript

1. Infrastructure at CERN Scale Ricardo Rocha - CERN Cloud Team @ahcorporto ricardo.rocha@cern.ch
2. Founded in 1954 What is 96% of the universe made of? Fundamental Science Why isn’t there anti-matter in the universe? What was the state of matter just after the Big Bang?
12. ~70 PB/year 700 000 Cores ~400 000 Jobs ~30 GiB/s 200+ Sites
13. Computing at CERN Increased numbers, increased automation 1970s 2007
14. Computing at CERN Increased numbers, increased automation 1970 2007
15. Computing at CERN Increased numbers, increased automation 1970 2007
16. Computing at CERN Increased numbers, increased automation 1970 2007
17. CERN IT Today 200+ people Storage, Computing and Monitoring, Databases, Network, DC, … Batch Systems and Core Physics Services But also campus services: hotel, bike service, wifi, … Common for teams to work on 2 weeks sprints, even for operations Rota system per team ServiceNow for end user support tickets
18. Provisioning Deployment Update Physical Infrastructure Days or Weeks Minutes or Hours Minutes or Hours Utilization Poor Maintenance Highly Intrusive
19. Provisioning Deployment Update Physical Infrastructure Days or Weeks Minutes or Hours Minutes or Hours Utilization Poor Maintenance Highly Intrusive Cloud API Virtualization Minutes Minutes or Hours Minutes or Hours Good Potentially Less Intrusive
20. OpenStack Private Cloud 3 Separate Regions (Main, Batch, Point 8) Scalability, Rolling Upgrades Regions split in multiple Cells Often matching hardware deliveries Different configurations and capabilities Single hypervisor type (KVM, used to have HyperV as well) CELL 1 MAIN CELL 2 CELL N Compute GPU Compute Nova Network Neutron Neutron
21. OpenStack Private Cloud 3 Separate Regions (Main, Batch, Point 8) Scalability, Rolling Upgrades, Regions split in multiple Cells Often matching hardware deliveries Different configurations and capabilities Single hypervisor type (KVM, used to have HyperV as well) CELL 1 MAIN CELL 2 CELL N Compute GPU Compute Nova Network Neutron Neutron
22. OpenStack Private Cloud 3 Separate Regions (Main, Batch, Point 8) Scalability, Rolling Upgrades, Regions split in multiple Cells Often matching hardware deliveries Different configurations and capabilities Single hypervisor type (KVM, used to have HyperV as well) CELL 1 MAIN CELL 2 CELL N Compute GPU Compute Nova Network Neutron Neutron
23. Networking Flat, segmented network (broadcast domains) Hierarchy of Primary (hypervisors) and Secondary (VMs) services CELL NODE 2 NODE 1 VN V2 V1 V3 V2 V1 V3 S513-V-IP123 137.1XX.43.0/24 ( Primary Service ) S513-V-VM908 188.1XX.191.0/24 ( Secondary Service ) Hypervisors Virtual Machines
24. OpenStack Private Cloud Automate everything! Puppet based deployment of all components Including control plane running on VMs Same is true for most CERN services Workflows for all sorts of tasks Onboarding new users, project creation, quota updates, special capabilities Overcommit, Pre-emptible instances, Backfilling workloads
25. Provisioning Deployment Update Physical Infrastructure Days or Weeks Minutes or Hours Minutes or Hours Utilization Poor Maintenance Highly Intrusive Cloud API Virtualization Minutes Minutes or Hours Minutes or Hours Good Potentially Less Intrusive Containers Seconds Seconds Seconds Very Good Less Intrusive
26. Lingua franca of the cloud Managed services offered by all major public clouds Multiple options for on-premise or self-managed deployments Common declarative API for basic infrastructure : compute, storage, networking Healthy ecosystem of tools offering extended functionality Kubernetes
27. Lingua franca of the cloud Managed services offered by all major public clouds Multiple options for on-premise or self-managed deployments Common declarative API for basic infrastructure : compute, storage, networking Healthy ecosystem of tools offering extended functionality Kubernetes
28. GitOps for Automation We were already doing similar things with Puppet Git as the source of truth for configuration data Allowing multiple choices of deployment models 1 ⇢ 1: Currently the most popular: one application, one cluster 1 ⇢ *: One application, multiple clusters (HA, Blast Radius, Rolling Upgrades) * ⇢ *: Separation of roles, improve resource usage Meta Chart git push FluxCD git pull Helm Release CRD Helm Operator
29. Cluster Creation Image Pre-Pull Data Stage-In Process 5 min 4 min 4 min 90 sec Kubernetes More than just infrastructure management Potential to ease scaling out data analysis on-demand Challenge: Re-processing the Higgs analysis in under 10min Processing a dataset of ~70TB of data split in ~25000 files
30. OpenStack Magnum 70 TB Dataset Job Results Interactive Visualization Aggregation 25000 Kubernetes Jobs
31. Cluster on GKE Max 25000 Cores Single Region, 3 Zones 70 TB Dataset Job Results Interactive Visualization Aggregation 25000 Kubernetes Jobs
33. Monitoring From ~40k machines More than 3TB/day compressed Modular architecture Decoupled producers / consumers Built-in stream processing Multiple backends with different SLAs Credit: Diogo Lima Nicolau
34. Credit: Diogo Lima Nicolau
35. Service Availability Historical View • Availability per service • Outages integration Credit: Diogo Lima Nicolau
36. Alarming Local (on the machine) • Simple Threshold / Actuators On dashboards • Grafana alert engine External • Alarm source Integrated with ticketing system • Service now Credit: Diogo Lima Nicolau
37. Challenges Do more with similar resources High Luminosity Large Hadron Collider x7 collisions per second, x10 Data and Computing Machine Learning Considered for fast simulation, detector triggers, anomaly detection, … Accommodate accelerators and scale this new type of workload GPUs, TPUs, IPUs, FPGAs, ...
38. Questions ? @ahcorporto ricardo.rocha@cern.ch http://visits.cern/
39. User Notebook Distributed Compute Build, Validate Model Train at Scale Persistent Storage for Feedback 1. 2. 3. 4. Serving