Download presentation
Presentation is loading. Please wait.
Published byVirgil Wells Modified over 8 years ago
1
FusionCube: Simple, Agile, Efficient —— Huawei One-stop Cloud Infrastructure Derek Liu, 163302 April 2013
2
1 Please delete this page before you present to customer Author / ID Liu Lei/163302 Department IT Elastic Computing Marketing Co-author / ID Wang Lingfei(199640), Feng Xingzhi(188809) Approver / ID Release Data April 2013 Please see notes for Key Messages
3
2 Enterprise IT Catching Up Business Drivers PUEApplication deployment days IT Business Drivers Competition Market IT department devotes 2/3 of time and resources to infrastructure operations IT Operations
4
3 Expectation of IT Resource Allocation Mission Critical Biz. Continuity DR Consolidate Virtualization Automate Optimize Modernize Mobile/Social Biz. Efficiency Innovate Host/Outsource Security IT Governance Compliance Biz. Analytics Cost Control Biz Alignment Risk Management
5
4 Converged Infrastructure : A re-design of IT infrastructure No external SAN No external switches No cables External storage Computing Unit External switches Storage Management Network Management Compute Storage Network Virtualization Platform Virtualization Management Server Management Converged component (compute, storage, network) One Managemeent Virtualization Platform
6
5 Converged Infrastructure Value Propositions Converged Infrastructure Simple No external switch No SAN One management interface Service: One throat to choke Agile Fast provisioning: pre-integrated, pre-validated Virtualized resource pool: easy to reposition Plug-and-play: Software defined components Performance Enables SW & HW optimization Distributed architecture enables high parallism Cost Effective High density: save space and power Higher utilization of hardware resources No cable
7
6 Virtualize P to V START 64% Automate P to V Automate Processes Phase I 32% P to V Automate Processes Converged Systems Converged Systems Converge Phase II 3% Source: IDC Estimates, March 2012 Industry Is Moving Toward Converged Infrastructure VMware Cisco EMC NetApp IBM HP HDS DELL
8
7 iNIC Card SSD 卡 GPU Computing Storage Network Virtualization Platform Unified Central Management Distributed Storage FusionCube – A True Converged Infrastructure
9
8 1.5 X IO Expansion Slots Mid-plane Throughput (Unidirectional 10hGb SerDes counted ) Local Storage Capacity 1.25 X Industry E9000 2.25 X 6 4 5.76Tbps 2.56Tpbs 15TB 12TB Convergence: Computing, Storage and Network integrated in one box I/O Acceleration: The only Blade server designed to support standard PCIe Cards inside the node Affordability: Lower memory cost up to 50% when max memory capacity modules populated *Cost comparison based on related vendors’ website quotation as of October 2012 + - $ Future Upgradability: Support next 4 generations of X86 CPUs and IO evolution to 100GbE & EDR Infiniband E9000: Designed for Converged Infrastructure E9000 Blade Server Front View Rear View
10
9 CH121CH240CH221CH222 Computing blade provides computing centric capability Performance blades provide balanced enhancement for computing and I/O capability, Balanced blade provides both matching storage and computing resources Aims at virtualization scenarios Aim at high performance scenarios, like Database, HPCAims at typical OA, VDI scenarios Half-width blade with 2 way Intel® Xeon® E5-2600 CPU (4/6/8 cores) 24 DIMM, 2*2.5’SSD/SAS/SATA, 1*PCIe 8x,2*Mezz 16x PCIe Full-width blade with 4 way Intel® Xeon® E5-2600 CPU (4/6/8 cores), 48 DIMM->1.5TB RAM, 8*2.5“ SSD/SAS/SATA , 2*Mezz 16x PCIe Full-width blade with 2 way Intel® Xeon® E5-2600 CPU (4/6/8 cores) 24 DIMM , 2*2.5“ SSD/SAS/SATA , 2*Mezz 16x PCIe , 2*16x PCIe standard card Full-width blade with 2 way Intel® Xeon® E5-2600 CPU (4/6/8 cores) 24 DIMM , 15*2.5“ SSD/SAS/SATA , 2*Mezz 16x PCIe , 1* 8x PCIe standard card Various Blades for Different Scenarios
11
10 CX610CX116 CX310 、 CX311 CX110CX317CX911 InfiniBand Switch GE Pass Through 10GE/FCOE SwitchGE Switch 10GE Pass Through 10GE/FC Switch 16*QDR/FDR Downstream 18*QDR/FDR Upstream 32*GE Downstream 32*GE Upstream FCoE/DCB/TRILL 32*10GE Downstream , 16*10GE+8*10GE/8*8 G FC Upstream 32*GE Downstream 12*GE+4*10GE Upstream 32*10GE Downstream 32*10GE Upstream 32*10GE/16*8G FC Downstream 16*10GE+8*8G FC Upstream Flexible Switching Modules
12
11 Optimization with Hardware Techniques Huawei own design Large size Cost effective Backplane and Switches Backplane switching 5.76-14.4 Tbps Switch delay =500 ns with PCIe Expansion Support Eth, FC, IB Bigger Memory Linear Cost 1.5x Height 2x RAM Size PCIe-SSDiNIC Card Offload vSwitch to hardware 3x-7x network I/O improvement
13
12 Start from small One chassis - Half configuration with 4 full-width blades Up to 8 chassis concatenation w/o external switches Up to 4096 vCores, 96TB RAM, and 1.9PB Storage Scale out with external switches Auto-discover, auto-configure Up to 20*3*512 = 30,720 vCores, 240TB RAM and 6PB Storage Scale On Demand Smoothly Single Chassis One Rack with 3 Chassis Multiple Racks Up to 20 Racks
14
13 No SAN Design: DAS Based Distributed Storage High Performance Parallel I/O 10X total IO throughput 3-5X IOPS improvement High Reliability Replications cross nodes Quick data rebuild (30min vs. 12hrs for 1 TB) High Scalability Up to 2000 nodes Linearly scalable in both capacity and performance Storage Blades Server Hypervisor Server volume1 Server volume2 Server volume3 FusionStorage VM Client VM I/O FusionSphere Hypervisor FC/IP network HDD SAN/NAS
15
14 Unified Management Interface for All Components Management Portal Resource Pool Unified Monitoring
16
15 n FusionSphere: Huawei’s Virtualization Platform
17
16 Start Purchase Shipping System Integration App Deployment Integrated Testing Platform Setup Business Online Equipment Testing 1-3W 4-8W 1W 10-18W FusionCube: Platform Ready in 2 hours Solution Design 1-3W Plug and Play Quick Deployment Avoid Human Mistake With Best Practice Proven Solution Optimization Simple Purchasing Faster Shipping Better Experience All In One Box Solution
18
17 Re-cap: What is FusionCube What is FusionCube ? A true converged infrastructure that relieve IT department from tedious maintenance works. E9000 + FusionSphere + FusionStorage Pre-integrated, Pre-validated, Unified Management + SW & HW Optimizations
19
18 Data Warehouse Private Cloud Typical Scenarios for FusionCube Resources utilization Business agility Management complexity Performance improvement Cost effective High scalability
20
19 Data Warehouse Performance is about IO throughput Join Group Avg/Sum Query processing Result visualization Table scan, disk IO throughput is bottleneck RAM size limit, write group temporary data to disk. Disk IO latency and throughput is bottleneck. Read temporary data back, disk IO latency is bottleneck Aggregated IO throughput of storage is the main bottleneck!
21
20 128 GBps IO throughput per chassis, bottleneck is significantly reduced Cost effective based on standard components. Storage Blade PCI-e SSD FusionCube reduces IO throughput bottleneck Storage Blade PCI-e SSD Storage Blade PCI-e SSD Storage Blade PCI-e SSD Compute Blade PCI-e SSD CPU Compute Blade PCI-e SSD CPU InfiniBand PCI-e SSD as main storage removes bottleneck to storage media Fastest interconnecting network removes bottleneck between storage and computing nodes. Large RAM size means more cache and less IO Distributed storage removes central controller bottleneck
22
21 Cost effective alternative DW Infrastructure Traditional SMP based DW DB + Miniframe/Server + Central Storage Hit performance bottleneck Can NOT scale cost effectively Switch FC Switch SAN Exadata TeraData FusionCube MPP Parallel Database DW Reference Architecture Performance gain from HW assisted SW optimization Existing DB application requires no change Expensive Scale-out with high performance Expensive Existing DW application may need modification Cost effective alternative solution Based on innovative scale-out architecture High performance: Distributed storage + HW optimization Existing DB and application run without modification (Oracle/HANA/SybaseIQ/..) Database Appliance
23
22 Compare FusionCube and Exadata for OLAP 1.Normally a OLAP query goes through several phases JOIN/GROUPING/SUM 2.Because the amount of data is normally large and can’t be fit into cache (RAM or SSD), data needs to be read from disk. At JOIN, a full table scan is normally taking place. At GROUPING, normally 2X to 3X more IO will take place. Also, at GROUPING, temporary data will need to be kept for later stage, so disk write is needed here (normally OLAP is dominated with disk read). 3.So in general, OLAP is disk IO intensive. 4.When there are concurrent queries going on, the total IO throughput is normally the bottleneck (in other words, the storage network is bottleneck). This explains Exadata, MPP database or FusionCube advantage over SAN (they all have 10X more IO throughput than their SAN counterpart at comparable price). Now compare FusionCube and Exadata for OLAP 1.Different approach: Exadata mainly employs software optimization, for example, Smart Scan, to overcome the IO throughput bottleneck and improve OLAP performance. FusionCube mainly employs hardware optimization, for example, using SSD as main storage. 2.FusionCube beat Exadata X2 mainly because FusionCube use PCIe-SSD while Exadata X2 uses SAS disks as main storage. In particular, at GROUPING stage, FusionCube has much faster rate for writing temporary data. This process can’t be optimized by Exadata's Smart Scan which works for read operation only. This explains our testing result from Huawei IT project. 3.FusionCube will be at least comparable with Exadata X3 because X3 still uses SSD as cache. Even it increased cache size, it still can’t help much with write operations. However, with Smart Scan, this cache size increase will help read operation so overall performance will be increased. According to Oracle’s measure, X3 will double X2’s performance which would be comparable with FusionCube’s performance which is about double of X2’s.
24
23 FusionCube vs. Exadata X2 and X3 Single RackExadata X2FusionCube2.0Exadata X3FusionCube2.1 Effective storage capacity50.4TB57.6TB50.4TB115.2TB Max RAM1.8TB6.9TB2.9TB6.9TB SSD Capacity5.3TB(Cache)57.6TB22.4TB(Cache)115.2TB HDD Capacity50.4TBNA50.4TBNA InfiniBand40Gbps56Gbps40Gbps56Gbps Max IO throughput75GBps96GBps100GBps128GBps Actual IO throughput (tested)25GB78GBNA100GB Concurrent queries6001800 900 ( Estimate ) 2300 用户( Estimate ) 8*1U DB Servers : 2*X5675, 96G RAM 2*40Gb IB; 2*10Ge; 14*2U Computing Servers 2*X5640, 24G RAM 2*40Gb IB; 4*100G SLC Flash 12* 15K SAS 3*12U E9000 Blazes 2*E2680, 384G RAM 2*56Gb IB; 2*10Ge; 2*2.4T MLC Flash FusionCube beat Exadata X2: SSD vs. SAS; Estimated to be comparable with X3. Both have high throughput, but FusionCube uses SSD as main storage while Exadata uses SSD as cache.
25
24 Background and problems : Huawei financial reporting and analytics system 40TB, increase 3TB per month, running Exadata X2-2 Response time drops as amount of data and users increase Support 300 users now, expect to support 1000 users Expansion of Exadata is expensive and hard to justify ROI Solution : Employ FusionCube+Oracle OBIEE RAC without application change Result : Response time dropped 27% , performance increased 3X-9X Support >1000 concurrent users Data Warehouse Case Study 1/3 Rack 1.2 Rack VS FusionCubeExadata X2-2 680.5 Seconds 163.8 sec. FusionCubeExadata Avg. Time (Top3 SQL query time) Rack Computing Cores Storage Cores RAMSSDSAS FusionCube1/3 128 ( converge ) 3T38.4T0 Exadata v21.2722041T6T120T FusionCube is deeply optimized for data warehouse, 4x cost effective than Exadata
26
25 Traditional IT Inefficiency Calls Private Cloud Computer Rooms App PC App PC App PC App Roaming Users Server Farms App IT Staff Teachers Students Industry Collaborators IT system of a logistics research lab in a large university. Applications are tightly coupled with hardware Utilization is low, especially for some precious software Resource provisioning is slow Affecting research and teaching activities Limiting collaboration with industry colleagues Various hardware resources are hard to manage Maintenance is nightmare Waste of hardware and software resources ….
27
26 Research, Teaching Internal Tools VDIVM Industry Collaboration 3D Design, VDI DB Storage IT Administrator TeachersStudentsIndustry Collaborators Efficiency One platform support all IT needs All resources under one management Utilization increased due to easy access Flexibility Dynamic resource (re)assignment according to demand Fast response to research and teaching requirement Accessibility Applications and resources can be accessed at anytime and anywhere by anybody Value Proposition Internet Intranet FusionCube as Cloud Infrastructure
28
27 Benefits of Private Cloud 99.95% <30% >60% Cost Savings 节省 20% 1W~1M <1Min Efficiency 2-4 Hours 3Min HA Reliability User Experience Buy hardware Install software Staff User involved Flexible Focus on research Resource assignment Resource utilization Troubleshooting No SAN 24 hours Standalone Hardware 99% Precious software
29
28 Case Study ROI Analysis BeforeAfterComparison CPU utilization10-30%40-60%4X higher CAPEX 36 Servers, separate storage16 Servers (including storage)56% saving Operation efficiency Independently for computing, storage and networking Unified managementImproved efficiency Power consumption 42KW10KW 76 % saving Resource acquiringWeek(s)30 VMs in one minute50X faster Application software management Runs on dedicated hardwareDedicated VM, shared hardware More people can share the software Configuration managementWeek(s)1 Hr50X faster Reliability99%HA, 99.5% Downtime reduced by 82 hours per year Event managementHoursMinutes10X faster Unified hardwareNOYESEasy maintenance Self serviceLackDesktopReduces number of tickets
30
29 FusionCube – One-stop Cloud Infrastructure FusionStorage d All in one box Unified Management Flexible configuration Resource pool management High density design HW & SW optimizations High performance converged storage Simple Agile Efficient
31
Copyright©2012 Huawei Technologies Co., Ltd. All Rights Reserved. The information in this document may contain predictive statements including, without limitation, statements regarding the future financial and operating results, future product portfolio, new technology, etc. There are a number of factors that could cause actual results and developments to differ materially from those expressed or implied in the predictive statements. Therefore, such information is provided for reference purpose only and constitutes neither an offer nor an acceptance. Huawei may change the information at any time without notice. HUAWEI ENTERPRISE ICT SOLUTIONS A BETTER WAY
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.