The CERN OpenStack Cloud

Transcription

1

2 The CERN OpenStack Cloud Jan van Eldik for the CERN Cloud Team 2

3 Agenda Objective: why a private Cloud at CERN? Bulk compute, server consolidation Overview of OpenStack Cloud service Components & numbers Organization humans in the loop Operations Upgrades, support, performance, Outlook 3

4 Why a private Cloud? Compute needs of the organization growing and changing Bulk compute Compute-intensive Low SLA the more, the better Server consolidation Web servers, dev boxes, build nodes, Linux, Windows Cattle *and* pets 4

5 Virtualization helps a lot Stop handing out physical boxes Customers {do not care,do not want to care,should not care} about managing hardware Hardware over-dimensioned for most usecases and we only buy one type Good experience with earlier Virtualization Kiosk 3000 Linux and Windows VMs Hyper-V, managed by Microsoft VMM 5

6 Enter the Cloud Cloud promise attractive for CERN Provide API to customers Scale out resources to CERN-size Lots of potential beyond our earlier offerings Bursting, federating, block storage, Single pane of glass 2012: decide to consolidate resource provisioning under a single service OpenStack based Pets *and* Cattle KVM *and* HyperV 6

7 VMs! IT/CM Group Meeting - Jan 12,

8 CERN OpenStack Project July 2013 CERN OpenStack Production February 2014 CERN OpenStack Havana October 2014 CERN OpenStack Icehouse March2015 CERN OpenStack Juno September 2015 CERN OpenStack Kilo September 2016 CERN OpenStack Liberty 5 April September April October April October April October 2015 ESSEX FOLSOM GRIZZLY HAVANA ICEHOUSE JUNO KILO LIBERTY MITAKA Nova (*) Swift Glance (*) Horizon (*) Keystone (*) Nova (*) Swift Glance (*) Horizon (*) Keystone (*) Quantum Cinder Nova Swift Glance Horizon Keystone Quantum Cinder Ceilometer (*) Nova Swift Glance Horizon Keystone Neutron Cinder Ceilometer (*) Heat Nova Swift Glance Horizon Keystone Neutron Cinder Ceilometer Heat Ironic Trove Nova Swift Glance Horizon Keystone Neutron Cinder Ceilometer Heat (*) Rally (*) Nova Swift Glance Horizon Keystone Neutron Cinder Ceilometer Heat Rally Manila Nova Swift Glance Horizon Keystone Neutron (*) Cinder Ceilometer Heat Rally Magnum (*) Barbican (*) Ironic Nova Swift Glance Horizon Keystone Neutron Cinder Ceilometer Heat Rally Magnum Barbican Ironic Mistral Manila (*) Pilot Trial 8

9 Nova Cells Top level cell Runs API service Top cell scheduler Child cells run Compute nodes Scheduler Conductor 40+ cells Version 2 coming Default for all 9

10 Block storage Predominantly Ceph, Geneva * 2, Budapest Some NetApp From ~200TB total to ~45Cinde0 TB of RBD + 50 TB RGW Example: ~25 puppet masters reading node configurations at up to 40kHz Ceph is backend for Glance images CephFS with Manila is under investigation for cluster filesystems OpenStack Glance + Cinder iops 11

11 Operations: Containers Magnum: OpenStack project to treat Container Orchestration Engines (COEs) as 1 st class resources Production service since Q Support for Docker Swarm, Kubernetes, Mesos - Storage drivers for (CERN-specific) EOS, CVMFS Many users interested, usage ramping up - GitLab CI, Jupyter/Swan, FTS, 12

12 Some selected numbers 7000 compute nodes 190K cores, ~70% in compute-intensive VMs 40+ cells Virtual Machines 800 shared projects, 2478 user projects 100s of VMs created/deleted each hour 3000 Volumes in Ceph 13

13 Status In production: >190K cores >7000 hypervisors ~100,000 additional cores being installed in next 6 months 90% of CERN s compute resources are now delivered on top of OpenStack 14

14 Humans in the loop Multiple CERN teams Openstack team: service managers and service developers Procurement, hardware mgmt. teams Ceph, DBoD teams Upstream OpenStack superusers Development, large deployment, board representative RDO Packaging, testing 15

15 Operations User support Tickets, documentation, etc. Deployment RDO packages, with some customizations/selected patches Puppetized configuration Upgrades Component by component, twice a year Operating systems CentOS 7.2 -> 7.3 ongoing SLC6 ( CentOS6 ) phase-out finished HyperV -> KVM migrations have started 16

16 VM migrations 5000 VMs migrated to newer hardware Live when possible Cold when necessary (CPU architecture, attached volumes) To retire older hosts Upgraded SLC6 hypervisors to CentOS7 Pre-req for upgrade to Nova Liberty 18

17 Automate provisioning with Automate routine procedures - Common place for workflows - Clean web interface - Scheduled jobs, cron-style - Traceability and auditing - Fine-grained access control - Disable compute node Disable nova-service Switch Alarms OFF Update Service-Now ticket Procedures for - OpenStack project creation - OpenStack quota changes - Notifications of VM owners - Usage and health reports - Notifications Other tasks Send to VM owners Save intervention details Send calendar invitation Add remote AT job Post new message broker 19

18 Optimizing CPU performance 70% of cores used for compute-intensive workloads Low-SLA CPU passthrough NUMA-aware flavours CPU overhead ~3% for 24-core VMS (details in backup slides) 20

19 Outlook Operate & grow the service Add new hardware, upgrade components, Get experience with Magnum Nova-network -> Neutron On a per-cell basis Phase-out HyperV, NetApp Investigate possible service additions Ironic, Manila, Neutron LBaaS, 21

20 Summary OpenStack at CERN in production since 3.5 years - We re working closely with the various communities - OpenStack, Ceph, RDO, Puppet, Cloud service continues to grow and mature - While experimental, good experience with Nova cells for scaling - Experience gained helps with general resource provisioning - New features added (containers, identity federation) - Expansion planned (bare metal provisioning) Confronting some major operational challenges - Transparent retirement of service hosts - Replacement of network layer (read about our recent 7M req/s Magnum & Kubernetes!) 22

21

22 10 9 HS06 sec / month PB/year LHC: Data & Compute Growth CMS ATLAS ALICE LHCb Collisions produce ~1 PB/s Run 1 Run 2 Run 3 Run GRID ATLAS CMS LHCb ALICE What we can afford 20 0 Run 1 Run 2 Run 3 Run 4 24

23 OpenStack at CERN In production: 4 clouds >200K cores >8,000 hypervisors 90% of CERN s compute resources are now delivered on top of OpenStack 25

24 CERN Cloud Architecture (1) Deployment spans our two data centers - 1 region (to have 1 API), ~40 cells - Cells map use cases hardware, hypervisor type, location, users, Top cell on physical and virtual nodes in HA - Clustered RabbitMQ with mirrored queues - API servers are VMs in various child cells Child cell controllers are OpenStack VMs - One controller per cell - Tradeoff between complexity and failure impact 26

25 CERN Cloud in Numbers (2) Rally down Every 10s a VM gets created or deleted in our cloud! images/snapshots - Glance on Ceph volumes - Cinder on Ceph (& NetApp) in GVA & Wigner Only issue during 2 years in prod: Ceph Issue

26 dev environment Simulate the full CERN cloud environment on your laptop, even offline Docker containers in a Kubernetes cluster - Clone of central Puppet configuration - Mock-up of - our two Ceph clusters - our secret store - our network DB & DHCP - Central DB & Rabbit instances - One POD per service Change & test in seconds Full upgrade testing 28

27 Rich Usage Spectrum Batch service - Physics data analysis IT Services - Sometimes built on top of other virtualised services Experiment services - E.g. build machines Engineering services - E.g. micro-electronics/chip design Infrastructure services - E.g. hostel booking, car rental, Personal VMs - Development rich requirement spectrum! 29

28 Usecases (1) Server consolidation: Service nodes, dev boxes, Personal VMs, Performance less important than durability Live-migration is desirable Persistent block storage is required Linux KVM, Windows HyperV Starting to run Win VMs under KVM Pets usecase : 32K cores, 7500 VMs 30

29 Usecases (2) Compute workloads Optimize for compute efficiency CPU passthrough, NUMA aware flavours Still, very different workloads IT Batch: LSF and HTCondor, longlived VMs, 8 and 16-core VMs, full-node flavors CMS Tier-0: medium-long, 8-core VMs LHCb Vcycle: short-lived, single core VMs Low-SLA, cattle usecase 150K cores, compute nodes 31

30 Wide Hardware Spectrum The ~6700 hypervisors differ in - Processor architectures: AMD vs. Intel (av. features, NUMA, ) - Core-to-RAM ratio (1:2, 1:4, 1:1.5,...) - Core-to-disk ratio (going down with SSDs!) - Disk layout (2 HDDs, 3 HDDs, 2 HDDs + 1 SSD, 2 SSDs, ) - Network (1GbE vs 10 GbE) - Critical vs physics power - Physical location (Geneva vs. Budapest) - Network domain (LCG vs. GPN vs. TN) - CERN CentOS 7, RHEL7, SLC6, Windows - Variety reflected/accessible via instance types, cells, projects variety not necessarily visible to users! - We try to keep things simple and hide some of the complexity - We can react to (most of the) requests with special needs 32

31 Basic Building Blocks: Instance Types Name vcpus RAM [GB] Ephemeral [GB] m2.small m2.medium m2.large m2.xlarge m2.2xlarge m2.3xlarge ~80 flavors in total (swap, additional/large ephemerals, core-to- RAM, ) 33

32 Basic Building Blocks: Volume Types Name IOPS b/w [MB/s] feature Location Backend standard GVA io fast GVA cp critical GVA cpio critical/fast GVA cp2 n.a. 120 critical/windows GVA wig-cp critical WIG wig-cpio critical/fast WIG m2.* flavor family plus volumes as basic building blocks for services 34

33 Operations: Network Migration We ll need to replace nova-network - It s going to be deprecated (really really really this time) New potential features (Tenant networks, LBaaS, Floating IPs, ) We have a number of patches to adapt to the CERN network constraints - We patched nova for each release - neutron allows for out-of-tree plugins! 35

34 Operations: Federated Clouds Public Cloud such as Rackspace or IBM Many Others on Their Way NecTAR Australia Brookhaven National Labs ALICE Trigger 9K cores CERN Private Cloud 160K cores INFN Italy ATLAS Trigger 28K cores CMS Trigger 13K cores Access to EduGAIN users via Horizon - Allow (limited) access given appropriate membership 36

35 IO Performance Issues Symptoms - High IOwait - Erratic response times User expectation - Ephemeral HDD Mirrored spinning drives on the host IOPS shared among guests 37

36 IO Performance: Scheduling deadline default (on RH s virtual-guest profile) - prefers reads (heavy load) - can block writes up to 5 secs! - sssd DB update at login time - CFQ better suited IO benchmark starts Switch from deadline to CFQ IO benchmarks continues 38

37 IO: Cinder Volumes at Work IOPS QoS changed from 100 to 500 IOPS ATLAS TDAQ monitoring application Y- Axis: CPU % spent in IOwait Blue: CVI VM (h/w RAID-10 with cache) Yellow: OpenStack VM Going to 1000 IOPS doesn t help. If the application is not designed to keep enough requests in-flight, more IOPS do not reduce the time spent in IOwait. We need to reduce the latency as well 39

38 IO: Block-level Caching Single SSD to cache in front of HDDs - flashcache, bcache, dm-cache, We chose bcache: - Easy to setup - Strong error handling On a 4 VM hypervisor: ~25-50 IOPS/VM ~1k-5k IOPS/VM 40

39 IO: bcache and Latency Use a VM on a bcache hypervisor Blue: CVI VM (h/w RAID-10 w/ cache) Yellow: OpenStack VM Red: OpenStack on bcache HV SSD block level caching sufficient for IOPS and latency demands. 41

40 IO: ZFS ZIL/l2arc Meanwhile: SSD-only hypervisors - Fast, but (too) small for some use cases Re-use cache idea leveraging ZFS features - Have partitions on the VM (on the SSDs!) which ZFS can use for the l2arc cache and the ZFS Intent Log - Use a fast volumes as the backing store We re testing this setup at the moment 42

41 CPU Performance Issues The benchmarks on full-node VMs was about 20% lower than the one of the underlying host - Smaller VMs much better Investigated various tuning options - KSM*, EPT**, PAE, Pinning, +hardware type dependencies - Discrepancy down to ~10% between virtual and physical Comparison with Hyper-V: no general issue - Loss w/o tuning ~3% (full-node), <1% for small VMs - NUMA-awareness! *KSM on/off: beware of memory reclaim! **EPT on/off: beware of expensive page table walks! 43

42 CPU: NUMA NUMA-awareness identified as most efficient setting - Full node VMs have ~3% overhead in HS06 EPT-off side-effect - Small number of hosts, but very visible there Use 2MB Huge Pages - Keep the EPT off performance gain with EPT on More details in this talk 44

43 Operations: NUMA/THP Roll-out Rolled out on ~2 000 batch hypervisors (~6 000 VMs) - HP allocation as boot parameter reboot - VM NUMA awareness as flavor metadata delete/recreate Cell-by-cell (~200 hosts): - Queue-reshuffle to minimize resource impact - Draining & deletion of batch VMs - Hypervisor reconfiguration (Puppet) & reboot - Recreation of batch VMs Whole update took about 8 weeks - Organized between batch and cloud teams - No performance issue observed since 45

44 Future Plans Investigate Ironic (Bare metal provisioning) - OpenStack as one interface for compute resource provisioning - Allow for complete accounting - Use physical machines for containers Replace Hyper-V by qemu/kvm - Windows expertise is a scarce resource in our team - Reduce complexity in service setup 46