• Have any questions?
  • info.zbook.org@gmail.com

Networking Requirements For Hyperconverged

14d ago
5 Views
0 Downloads
3.07 MB
37 Pages
Last View : 10d ago
Last Download : n/a
Upload by : Camille Dion
Share:
Transcription

Networking Requirements forHyperconverged Infrastructure (HCI)Christine McMonigal, IntelJ Metz, Cisco SystemsAlex McDonald, NetAppFebruary 5, 2019

Today’s PresentersChristine McMonigalIntelJ MetzCisco 2019 Storage Networking Industry Association. All Rights Reserved.Alex McDonaldNetApp2

SNIA-At-A-Glance 2019 Storage Networking Industry Association. All Rights Reserved.3

SNIA Legal NoticeThe material contained in this presentation is copyrighted by the SNIA unless otherwisenoted.Member companies and individual members may use this material in presentations andliterature under the following conditions:Any slide or slides used must be reproduced in their entirety without modificationThe SNIA must be acknowledged as the source of any material used in the body of any document containingmaterial from these presentations.This presentation is a project of the SNIA.Neither the author nor the presenter is an attorney and nothing in this presentation isintended to be, or should be construed as legal advice or an opinion of counsel. If youneed legal advice or a legal opinion please contact your attorney.The information presented herein represents the author's personal opinion and currentunderstanding of the relevant issues involved. The author, the presenter, and the SNIAdo not assume any responsibility or liability for damages arising out of any reliance on oruse of this information.NO WARRANTIES, EXPRESS OR IMPLIED. USE AT YOUR OWN RISK. 2019 Storage Networking Industry Association. All Rights Reserved.4

Some Ground Rules!This is a vendor-neutral presentationDifferent vendors have different solutions,there are no recommendations expressedor impliedThis is a technology-neutral presentationNo value-comparison with othertechnologies is expressed or impliedNothing in this presentation supersedesyour vendor best practices andrecommendations! 2019 Storage Networking Industry Association. All Rights Reserved.5

Why This PresentationSource: hyperconvergence-needs-networkingThe weak point in this new hyperconvergedworld comes at the interconnect level.Hyperconvergence vendors assume thatstorage and compute are their playgroundand more nodes will be sold to satisfyrequirements in the future. However, tointerconnect these nodes they must rely onexisting network infrastructure.- Tom Hollingsworth, @networkingnerd 2019 Storage Networking Industry Association. All Rights Reserved.Hyperconverged solutions can be software only,or hardware-and-softwareFocus is on compute and storageNetworking is often overlooked, as it is assumedto always “be there”Things that storage and compute people getwrong about the network:*The network is reliableLatency is zeroBandwidth is infiniteThe network is secureTopology doesn’t changeThere is one administratorTransport cost is zeroThe network is homogeneous* Source: Gosling, James. 1997. The 8 fallacies of distributed computing.”6

AgendaWhat is Hyperconvergence (HCI)?What is it? SDS and HCIWorkloads, size of deploymentsHCI Storage CharacteristicsUnderstanding HCI Reads and WritesOther key considerationsHCI Networking ConsiderationsNetwork topology options, speeds, classifying traffic flowsSizingNetwork settings and configurations“Stretching” your cluster 2019 Storage Networking Industry Association. All Rights Reserved.7

What Is Hyperconvergence? 2019 Storage Networking Industry Association. All Rights Reserved.8

Storage Trade-OffsThere is a “sweet spot” for storageDepends on the workload andapplication typeNo “one-size fits all”Understanding “where” the solution fitsis critical to understanding “how” to putit together 2019 Storage Networking Industry Association. All Rights Reserved.9

Storage Trade-OffsHyperconvergence fits hereNot just manageability - it’s uniformmanageabilityAll looks the same 2019 Storage Networking Industry Association. All Rights Reserved.10

Hyperconvergence - A SpecializedSoftware-Defined StorageNew infrastructure offering that is a type of Software DefinedStorage (SDS)Tight Integration of standard servers for compute and storage,networking and virtualization, in an all-in-one applianceIntegration of hypervisors and physical infrastructureIs it Block? File? Can vary depending upon vendorStorage is presented via a distributed filesystem orobject storeBlock: Almost always iSCSI; File: Almost always NFSHas an abstraction layer for control plane managementMagic Sauce: Each node can talk to each other node,centralized management, intuitive UI 2019 Storage Networking Industry Association. All Rights Reserved.11

Hyperconvergence and IT ChallengesTop IT Challenges Prompting the Purchase of HyperconvergedHyperconverged SolutionAdoption Rate amongSurveyed Enterprises27% 21%already adoptedplan to adoptin next year49%Infrastructure costsInfrastructure costs and OpEx46%Storage management challengesVM management challenges37%IT resources challengesStorage management challenges36%VM management challengesLack of IT resources32%Provisioningtakes too long 31%Specific workload/LOBneedsSpecific workload/LOB needsProvisioning takes too long30%Replacing outdated hardware2H’18: Hyperconverged Platforms Customer Research, Technology Business Research, Inc. 2019 Storage Networking Industry Association. All Rights Reserved.Replacing outdated hardware

Trade-Offs - StrengthsHyperconverged Cluster 2019 Storage Networking Industry Association. All Rights Reserved.Virtualizes compute and storage –pools the physical storage devicesand compute capabilitiesData distributed across cluster fordurabilityScales performance with capacitySupports multiple workloadsManage using familiar virtualizationtools and resources; reduces“management sprawl”Rapid (-ish) deployment timesEases refresh budgeting

Trade-Offs - ChallengesHyperconverged Cluster 2019 Storage Networking Industry Association. All Rights Reserved.One vendor only - cannot mix-andmatch software solutions from differentvendors; no interoperabilityHardware flexibility variesBalance between compute and storagecan be difficult to achieve, especially inchanging application environmentsWorkloads can compete for resourcesand cause performance problemsSome use cases may not be supported(e.g. Big Data, Hadoop, Spark, etc.)Growth of clusters places addedburden on network

Cluster Size Impacts NetworkClusters are 3 or more server nodesCan be as high as 128 nodes; maximumvariesMost deployments typically run between4-12 nodes, but this varies widelyDegree of scalability varies betweenHCI solutionsIn general, the more nodes morenetwork demandsTraffic between nodes places pressureon the network, from a control planeperspective 2019 Storage Networking Industry Association. All Rights Reserved.SOURCE: TBR HYPERCONVERGED PLATFORMS CUSTOMER RESEARCH, 1H’1815

Which workloads are run on HCI?WORKLOADS CURRENTLY RUNNING ONHYPERCONVERGED INFRASTRUCTUREHCI initially used for secondaryworkloadsNow, primary and businesscritical workloads are mostpopularNetworking bandwidth andlatency expectations differ forthese workloads60%60%Database management50%56%Data backup and recovery48%51%Business intelligence/analytics48%53%Business processing (ERP)Development operations(DevOps)48%50%Customer resource management(CRM)47%43%Productivity/collaboration (e.g.,email, file/print)43%Edge compute (e.g., on-site datastorage and analysis)51%41%N/AVDI (virtual %60%Percent of Respondents (n 201) 2019 Storage Networking Industry Association. All Rights Reserved.SOURCE: TBRTBR HYPERCONVERGED PLATFORMS CUSTOMER RESEARCH, 2H’182H18(n 201)1H18(n 203)16

Users run multiple workloadssimultaneously on HCI clustersData locality preferences varybetween HCI solutionsSome HCI solutions try to co-locatedata stored for specific workloads onthe same servers where theapplications are runningThe greater the number of workloadsrunning, the less likely data locality isachievablePercent of RespondentsMixed Workloads Prevalent50%SIMULTANEOUS WORKLOADS RUN ONHYPERCONVERGED40%39%30%20%10%0%20% 21%15%2% 3%1-2SOURCE: TBR36%3-45-62H18(n 201)20%7-88% 7%9-101H18(n 203)15%13%More than10Most HCI solutions allowprioritization of workloads/apps 2019 Storage Networking Industry Association. All Rights Reserved.SOURCE: TBR HYPERCONVERGED PLATFORMS CUSTOMER RESEARCH, 2H’1817

AgendaWhat is Hyperconvergence (HCI)?What is it? SDS and HCIWorkloads, size of deploymentsHCI Storage CharacteristicsUnderstanding HCI Reads and WritesOther key considerationsHCI Networking ConsiderationsNetwork topology options, speeds, classifying traffic flowsSizingNetwork settings and configurations“Stretching” your cluster 2019 Storage Networking Industry Association. All Rights Reserved.18

HCI Storage Characteristics 2019 Storage Networking Industry Association. All Rights Reserved.19

Understanding Networking DifferencesTraditional Storage Environments“North-South” traffic from host to storageStorage is centralizedStorage services are maintained inside of storage arrayStorage Network is tuned for storage trafficManagement confined to each “layer” (compute,network, storage)High Availability managed at each levelActive/Active, Active/Passive multipathingNSPOF component architectures for devicesStorage resilience maintained in Arrays/FilersNetwork resilience maintained at L2-L4 or Fibre ChannelVM resilience maintained by hypervisor in hostsetc. 2019 Storage Networking Industry Association. All Rights Reserved.20

Hyperconverged Storage Traffic/Storage Abstraction M 2019 Storage Networking Industry Association. All Rights Reserved.Heavily increased “East-West Traffic”Demand for higher bandwidth andlower over-subscription ratios arecommon, especially for server-toserver communicationData is distributed to multiple nodesand load-balanced, creating additionaltraffic on the networkEach additional node increasesoverall bandwidth requirements,especially for load-balancingalgorithms21

HCI Reads and WritesGoal is to make I/O transparent to the applicationApplication just fires I/O at virtualized disksReads and writes require co-ordination between nodesManagement software (sometimes an accelerator) on each nodecollects the I/O and decides what to do with itReads: looks up which node(s) holds and asks for itDelay here isn’t helpfulWrites: can be “coalesced”, cached and subsequently passed to node(s)for writingOptimizations can be performed (compression, deduplication)Delay here can (within limits) be helpfulNetwork traffic includes east-west node management “chatter”Not all I/O need end up on the networkCan require RAID on the nodeErasure Coding sharding requires I/O on the network 2019 Storage Networking Industry Association. All Rights Reserved.22

All Flash & NVMe SolutionsHCI abstracts away the details of storagemanagement but, you can only virtualize the actual HWunderneathEnsure HW selection will meet workloadperformance requirementsAll-flash and NVMe solutions deliver higherbandwidth, lower latency*Beware the shifting bottleneck! but, the network will become a bottleneck unlessusers plan for itNewer technologies such as RDMA, NVMe-overFabrics (NVMe-oF) and/or higher network speedscan help avoid bottlenecksWatch for stressing the network with thesenew components*See SNIA NSF “Under the Hood with 75515 2019 Storage Networking Industry Association. All Rights Reserved.23

What is a “Noisy Neighbor?”Resource contention between virtual machinesDifferent I/O requirements negatively impacting resources(compute/network/storage)Trade-offs: Larger Clusters versus Smaller, focused ClustersMitigation:Things like RDMA helpRDMA gets rid of the host bufferingMakes the network more responsibleRelieves the pressure on caching on the host memory andcompute cycles, but pushes the burden into the networkHowever You’re throwing everything onto the network, so you have tomake sure the network can cope with it 2019 Storage Networking Industry Association. All Rights Reserved.Source: Stephen Foskett. ”The I/O blender-part-2-virtualization/

Block or File?At this point, with modern kit, block and file don’t matter at this scaleRecall: these workloads are well-understoodFor vendors that offer either option, it becomes dealer’s choicePerformance is roughly the same, and generally not an issueThis is not where the performance issues reside 2019 Storage Networking Industry Association. All Rights Reserved.25

AgendaWhat is Hyperconvergence (HCI)?What is it? SDS and HCIWorkloads, size of deploymentsHCI Storage CharacteristicsUnderstanding HCI Reads and WritesOther key considerationsHCI Networking ConsiderationsNetwork topology options, speeds, classifying traffic flowsSizingNetwork settings and configurations“Stretching” your cluster 2019 Storage Networking Industry Association. All Rights Reserved.26

HCI Networking Considerations 2019 Storage Networking Industry Association. All Rights Reserved.27

Network QualityWorth RepeatingWorth RepeatingThe quality of service, reliability, availability, scalability,and overall performance of Ethernet fabric ultimatelydefines the capabilities of the HCI systemThings that storage and compute people get wrongabout the network:*Worth RepeatingWorth RepeatingWorth RepeatingWorth RepeatingWorth Repeating 2019 Storage Networking Industry Association. All Rights Reserved.The network is reliableLatency is zeroBandwidth is infiniteThe network is secureTopology doesn’t changeThere is one administratorTransport cost is zeroThe network is homogeneousSource: *Gosling, James. 1997. The 8 fallacies ofdistributed computing.”28

How Many Networks?Can be multiple:Management NetworkVM VLAN NetworkStorage NetworkMetadata Network“Motioning” NetworkDifferent networks have different bandwidthand latency requirementsRemember: network bandwidth is measured in bits,not bytesEasy math: Divide by 10 to get the bytes; 10Gb 1GB throughputMinimum: 10GbE (some have maximum of 25GbE)- read vendor documentationStorage interfaces are generally much fasterthan network interfacesConsider NVMe impact on network 2019 Storage Networking Industry Association. All Rights Reserved.29

Small, Purpose-Built ClusterReduce Noisy Neighbor ScenariosFine-tune specific services for limited selection ofapplicationsSmaller fault domainsManagement interface is identicalNot many networking implications (comparativelyspeaking) for small, purpose-built clusters 2019 Storage Networking Industry Association. All Rights Reserved.30

Large Mix-Use ClusterMore protection and space efficiencyMore options for data protection(e.g., Erasure Coding or RAID levels)Better sharing efficiencyReduction of impact on individualresourcesIncrease in networking knowledge,management and QoS requirements 2019 Storage Networking Industry Association. All Rights Reserved.31

Vendors Want You To ConsiderNetwork SettingsSpanning TreeLoad BalancingJumbo FramesLink Aggregation/BondingThese things you will want to take your vendor’s adviceYou want the simplicity - you should follow vendors’ guidelines to the letter 2019 Storage Networking Industry Association. All Rights Reserved.32

“Stretching” A ClusterSome clusters may need multicast supportIn most cases, latency requirements are strictOften 2-5ms maximum Round Trip Time (RTT)Writes are replicated to the other site, which means thateach write could take 5ms if that’s your RTTMost solutions have SSD/NVMe caching, so yournetwork will be the biggest problemFor every ms of latency, that’s a ms worth of cacheExpensiveEvery ms is a huge amount of I/O that you can’t doBecause Physics! 2019 Storage Networking Industry Association. All Rights Reserved.33

SummaryHCI is a simple solution with complexity;complexity is under the coversListen to the vendorThey’ve done the blood, sweat and tearsUnderstands what is required from thesolutionUnderstanding the principles, and whatis going on, makes you a better end user 2019 Storage Networking Industry Association. All Rights Reserved.Very Special Thanks:Duncan EppingCormac HoganIvan PepelnjakChris TwiggChris DunkTom HollingsworthStephen FoskettJase McCartyPhil White34

Our Next NSF WebcastThe Scale-Out File System Architecture OverviewFebruary 28, 2019https://www.brighttalk.com/webcast/663/346111 2019 Storage Networking Industry Association. All Rights Reserved.35

After This WebcastPlease rate this webcast and provide us with feedbackThis webcast and a PDF of the slides will be posted to the SNIANetworking Storage Forum (NSF) website and available on-demandat www.snia.org/libraryA full Q&A from this webcast, including answers to questions wecouldn't get to today, will be posted to the SNIA-NSF blog:sniansfblog.orgFollow us on Twitter @SNIANSF 2019 Storage Networking Industry Association. All Rights Reserved.36

Thank You

48%Now, primary and business critical workloads are most popular Networking bandwidth and latency expectations differ for these workloads 60% 50% 48% 48% 47% 43% 41% 36% 27% 60% 56% 51% 53% 50% 43% 51% N/A 40% 32% 0% 20% 40% 60% Database management Data backup and recovery Business intelligence/analytics