101-Huawei OceanStor Dorado All-Flash Storage Systems Pre-Sales Training V1.5

Huawei OceanStor Dorado V6
All-Flash Storage Systems
Pre-sales Training
Security Level:
Intelligent IoE Era Drives Explosive Data Growth
Cloud 60 TB
4 billion 75 billion computing Daily training data volume
Interconnected IoT
users subscribers
Autonomous driving
5G Diverse Real-time vehicles
Mass data
applications processing
1 PB
30x 1,000,000x Daily production data
Traffic Graphics processing
capabilities
Data infrastructure
Digitally connected
8K, AR, VR factories
2 Huawei Confidential
Construct Data Infrastructure to Maximize Data Value
Industry
application
Governments Finance Carriers Public safety Large enterprises
Industrial
Data models Application enablement
enablement
Enablement
Processing
Data
infrastructure Storage
platforms Connection
Computing
Connections Enterprise wireless FTTx and WTTx WiFi Campus networks

IoE
Devices Sensors Mobile devices IoT Smart cameras Others
Redefine Data Infrastructure to Boost Digital Economy
Unleashed data potential and value

Data enablement
Training
AI
Database Big data Redefined data processing
Data processing
Block File Object HDFS

Redefined storage architecture
Data storage
Data access Diverse data connections
x86 Arm GPU NPU

Powerful computing capabilities
Computing
Transform Storage Infrastructure to Monetize Data
Larger-scale financial Smoother 5G One-stop

transactions communications government services
Peak-time transactions Billing processing time Government service efficiency
300% 67% 50%

Settlement for heavy workloads
150,000 transactions per second, Data does the legwork
within 4 hours at the beginning of
doubling profits
each month
Upgrade for Inclusive All-Flash Storage in Any Scenario
Huawei innovations make inclusive all-flash storage affordable
HDD SSD
Computing chip
capabilities
25%
Latency 2 ms 0.02 ms
Chip Batch processing
40%
Faster TCO 50%
latency of AI chips
SmartMatrix architecture Power Maintenance

5-year More CAPEX Footprint
capable of tolerating failures consumption cost
return rate 13.4% 0.8% Architecture
of up to 7 out of 8 controllers reliable 30% 82% 79% 50%
for always-on services
Data reduction for higher

Power space utilization More
consumption
10 W 3W Huawei AFA vs. HDD storage
Business Raw Effective efficient
capacity capacity
Lead the Market with Cutting-Edge OceanStor Dorado All-Flash Storage
All-flash storage market share

Gartner Magic Quadrant: Leader Global market growth rate: No.1
in China: No.1
Others
2019: Leader 10.8% 234%
2018: Challenger MacroSAN
4.7% Huawei
NetApp 33.4%
6.9%
Total:
USD 457.7M
IBM
10.6% 43%
2016: Further ahead 34%
27% 22% 17% 13%
Dell EMC
H3C
2014: Niche Player 14.6% Huawei Hitachi IBM NetApp Pure Dell HPE
18.9%
Storage EMC
Source: Gartner Magic Quadrant for

Primary Storage Source: IDC All-Flash Array Market Overview Source: Gartner Market Share, 2019 Q1
OceanStor Dorado All-Flash Storage Gains Industry Renown
Interop Grand Prize Verification by the ESG Lab Magnitude-9.0 earthquake

resistance certification
Proving that OceanStor

Recognizing Proving that OceanStor Dorado can prevent damage
OceanStor Dorado as Dorado can easily manage caused by vibration during
one of the most heavy-load applications for transportation, installation,
competitive products enterprises and operation
ESG — High Availability, Optimal Performance, Extremely Low Latency,
and 78% 5-Year TCO Savings with OceanStor Dorado V6
Enterprise Strategy Group is an IT analyst, research, One of the lowest response times while maintaining high performance
validation, and strategy firm that provides market intelligence
and actionable insight to the global IT community.
81 µs response time
Three out of four controllers offline in the test — I/Os run normal,
and response time is < 20 µ sec
Controller B Offline
Controller A Offline
220,000 IOPS
Controller C Offline
18 µ sec 20 µ sec
Up to 78% lower TCO over 5 years
Controller C Online
Controller A Online
Controller B Online
OceanStor Dorado All-Flash Storage Sets Benchmarks
First AI chip
First full NVMe series
World's fastest OceanStor Dorado 5000,

6000, and 18000 V3
First unified OceanStor Dorado V3
SAN & NAS
First all-flash storage
OceanStor V3
OceanStor Dorado
Industry's most stable Interop Grand Prize for No.1 performer in the
A-A solution the Best of Show Award SPC-1 benchmark
2009 to 2013 2014 to 2016 2017 2018 2019
OceanStor Dorado All-Flash Storage Highlights
Ever Fast Ever Solid AI-Powered
Industry's highest SmartMatrix fully- Intelligent full-lifecycle

performance and interconnected architecture management with AI chips
lowest latency for always-on applications and algorithms
20 million IOPS Tolerates failure of 7
0.1 ms latency out of 8 controllers AI-enabled O&M
Ever Fast
Chip-Powered Platforms for 20 Million IOPS
Chip-powered
Seamless storage upgrades for industry-leading

performance
Protocol-leading
Industry's
E2E NVMe for a high-performance expressway
highest
performance
Algorithm-accelerated
FlashLink® SSD-controller synergy with intelligent algorithms for

maximized all-flash performance
No.1 Performance with 5 Intelligent Chips
Network chip
1 Transmission 2x improved network latency
Kunpeng 920 processor

2 Compute 2x higher performance
AI chip
3 Intelligence 50% better read hit ratio
4 Storage SSD controller chip

2x improved write latency
BMC chip
5 Management
30% improved fault location
Intelligent Interface Chip Doubles Front-End Access Speed with
Multi-Protocol Parsing
• Industry-leading interface chip for on-

demand use with:
− 8 Gbit/s, 16 Gbit/s, and 32 Gbit/s Fibre
Channel
− 10GE, 25GE, 40GE, and 100GE
• Offloading protocol parsing from general-

purpose CPUs, reducing access latency by
50% (160 μs to 80 μs)
Kunpeng 920 Processors Power No. 1 Performance
25% better performance High performance Kunpeng 920-6430 (64-core, 2.6 GHz) 930+
than mainstream CPUs SPECint®_rate_base2006 Competitors
750
benchmark result
* Huawei lab test data. The benchmark test is conducted using
the SPEC test program compiled by the GCC compiler for Linux.
48 cores
High density
4 Kunpeng 920 processors
100 Gbit/s RoCE
Gbit/s x 16 lanes
Acceleration
SAS 3.0, 12
on a single controller
8-channel
engine
DDR4
Acceleration Offloads compute-intensive storage

algorithms to free computing power:
engine compression,
decompression, RAID, EC,
Acceleration engine encryption, decryption
AI Chip — 50% Higher Read Cache Hit Ratio
Intelligent cache
• Analyzes and learns I/O rules of various application

models with machine learning.
• Improves the read hit ratio and increases system

performance by up to 20%+.
Ascend 310
Semantic-related I/Os
FP16: 8 TeraFLOPS
with machine learning
Read cache hit ratio
10000001
1010101010
50%
Intelligent Controller Chip for SSDs Halves Write Latency
• Faster data access on SSDs with core FTL algorithm
• Write latency halved from 40 μs to 20 μs on NVMe

SSDs with light loads
BMC Chip Shortens Fault Recovery
• Unified and intelligent fault management
• 93% fault locating accuracy

• Fault recovery shortened from 2 hours
to 10 minutes
E2E NVMe for Full Series — Every Product Model Is Ever Fast
HBA
Server
Switch
0.1 ms
Test model: 7:3 read/write, 130 μs read
16 Gbit/s
Fibre Channel latency, 50 μs write latency
or 10GE 32 Gbit/s FC-NVMe
NVMe over 100 Gbit/s RDMA Time = 130 x 0.7 + 50 x 0.3 = 106 μs
SCSI
INTERPOSER
SAS Fewer
Shorter path Wider channel
NVMe interactions
562 μs 106 μs
E2E Every model

Full series
ever fast
NVMe AFA
*Front-end NVMe over 100 Gbit/s RDMA will be available

in the next version.
SAS AFA *Deployment of FC NVMe or NVMe over ROCE due to
the maturity of ecosystem.
NVMe Reduces Protocol Processing Latency and Accelerates
Data Transmission
App
Lower latency: communication
interactions drop from 4 to 2
Block layer
Controller SSD
Controller 1. Transfer command
SCSI Initiator NVMe
2. Ready to transfer
SAS
3. Transfer data
SAS
4. Response
SAS Target NVMe 1. NVMe write command

NVMe
2. NVMe write finished
SCSI
SAS protocol stack NVMe protocol stack NVMe averages far shorter storage latency than SAS 3.0
E2E Hardware and Software Acceleration for I/O Paths
Server
NVMe for various networks reduces overhead by
40%
RoCE or FC
Protocol parsing offloaded for 50% lower latency
Controller enclosure
Front-end Front-end Front-end Front-end • TOE interface card chip
interface interface interface interface
module module module module • ASIC I/O balance/allocation
30 μs
50 μs
Controller Controller Controller Controller
Lock-free processing of concurrent I/Os with
NVMe multi-queue polling
Back-end Back-end Back-end Back-end
interface interface interface interface
module module module module
100 μs
30 μs
SSD enclosure Prioritized processing for read requests
Uncompromising Performance with Innovative FlashLink® Algorithms
Controller with 5 chips

Core Reconstruction 81.79%
53.13%
0 0
01 Many-core Core
0 00 1 Core
1 scheduling 300 min/TB 15 min/TB Ordered 1 2 3 4 5
1
1 0 0
0 by reboot
0 1 0
1 1
0
1 10 Core sequence
10 000 0
0 10 1 1
11 1
1
00 0 0
Smart SSD
10 enclosure Kunpeng + Many-core algorithm Kunpeng + Service grouping AI chip + Cache algorithm
0 11
1101 1
0
0 10 0
10
0 1
0
1 01 1 2x computing power 20x reconstruction speed 50% cache hit ratio
0110 0
01
01001 1
10
1 0
01 0 0
01
11011 1
10
0 100 0
1 1 Data read
01
11
011 1
001 0
Metadata
Data write
1 0
11 1
New data
1
SSD111 1 Garbage Advanced feature
0 1 collection data Reconstruction
1 Garbage collection
1
Full-stripe writes Multi-streaming data separation Global I/O priority adjustment
Less write amplification Less garbage collection Lower latency
FlashLink®
— Intelligent Many-Core Algorithm Streamlines CPUs with Kunpeng Power
Reads and Data exchange Protocol Data

writes channel parsing flushing Read I/O 1 Read I/O 2 Write I/O 1 Write I/O 2
(LUN, LBA), Data
64 MB
… vNode vNode vNode Core Core Core Core Core Core Core Core
N1
Core Core Core Core
N7 N2
DHT CPU
N6 ring N3
CPU CPU Core Core Core Core Core Core Core Core
N5 N4
Dedicated Dedicated Shared Read/write I/O group
CPU partitioning in controllers Core grouping in CPUs Core-based I/O binding
2x CPU processing efficiency, 20% lower latency
FlashLink®
— Intelligent Many-Core Algorithm Dynamically Schedules CPU Resources
Controller CPU Core group

LUN enclosure LUN slicing
Controller
Core Core Core
…
Data
Read I/O 1
Slice 10 exchange
Core Core CPU core grouping
Slice 9
Controller
Core Core Core
Slice 8
Read/Write Read I/O 2
Slice 7
Core Core
Dynamic prioritized scheduling
Slice 6
Controller
Slice 5 Core Core Core
……4
Slice Data update Write I/O 1
Core Core
Slice 3
Slice 2
Service load isolation
Controller Core Core Core
Data
Slice 1 Write I/O 2
reduction
Core Core
Slice 0
FlashLink®
— Enclosure-Controller Synergy for Service Grouping and Computing Power Unleashing
Controller enclosure Smart SSD enclosure with intelligent chips

Front-end
interface module
Front-end
interface module
Front-end
interface module
Front-end
interface module 30%
Computing power
sharing
Data Data
reconstructionController
Controller reconstruction
Controller Controller
Back-end Back-end Back-end Back-end Impact on controller

interface module interface module interface module
Data reconstruction time
interface module performance (maximum IOPS)
30 min/TB 15 min/TB 15% 5%
Smart SSD enclosure + Kunpeng Smart SSD enclosure + Kunpeng

processor processor Data reconstruction bandwidth Controller CPU utilization
80 MB/s 200 MB/s 70%
FlashLink®
— AI Chip Algorithm for Real-Time Learning and Self-Optimization
Private
Real-time collection Online learning Feedback optimization

100+ built-in probes Tera-level FLOPS boosted by Real-time algorithm
Real time collection of high-speed online learning of AI adjustment optimizes models
business workload index chips in a 10 M+ IOPS scenario for customized cache
prefetching
Private Cloud VDI scenario

Cache hit ratio: 30% Latency: 67%
100 1 151
81.79
80
2 104
Ascend 310 60
40
56
3 103
FP16: 8 TeraFLOPS 20
4 95
0
Ordered 1 2 3 4 5
5 90
by reboot
sequence Ordered by reboot
sequence 0
F1500 OceanStor Dorado 6000 V6
μs 0 50 100 150 200
Scenario: 10 linked clone VMs in VM reboot
FlashLink®
— Sequential Writes of Large Blocks Reduce Write Overhead and Read/Write Pressure on SSDs
Technical highlights
 Controllers detect SSD data layouts.
 Discrete data blocks are aggregated to a continuous large block.
OceanStor Dorado
 Large blocks are written into SSDs sequentially.
Benefits
 Capitalizes on SAS bandwidth
 Less garbage collection
FlashLink®
— Multi-Streaming Data Separation Reduces SSD Garbage Collection
Combined storage of metadata Separate storage of metadata

and other data and other data
Metadata separation
Metadata Other data
FlashLink®
— I/O Priority Adjustment for Stable and Low Latency
Read/Write
数据读写 1st Read/Write
数据读写 1st
高级特性
Advanced features 1st 高级特性
Advanced features 2nd
Cache批量写
Cache flushing 1st I/O priority adjustment cache批量写
Cache flushing 3rd
硬盘重构
Reconstruction 1st 硬盘重构
Reconstruction 4th
垃圾回收 1st 垃圾回收 5th

Garbage collection Garbage collection
Read and write I/Os are processed first to

All I/Os are handled chronologically. minimize read/write latency.
No.1 performance in SPC-1 benchmark
20,000,000 IOPS
10,001,52
2 7,000,56
5
2,401,17 2,004,94
1 1,500,08
1 7
Huawei OceanStor Fujitsu Huawei OceanStor NetApp HDS IBM
Dorado 18000 V6 DX8900 S4 Dorado 18000 V3 A800 G1000 DS8888
2x Better Test conditions: 32-controller, SPC-1; test report released in 2020.
2x higher than competitors in databases

Than the Next-Best Player
640,000 IOPS
340,000
225,000
Huawei OceanStor
Vendor H high-end AFA Vendor E high-end AFA
Dorado 18000 V6
Test conditions: dual-controller, 7:3 read/write, 80% space
occupied, data reduction enabled, 1 ms latency
Huawei vs. Competition
Performance with value-added features enabled
Competitor AFA: 40% decrease; Huawei AFA: < 10% decrease
Uncompromising
Performance
In Different Scenarios
Data reduction GC RAID 6 Snapshot
Competitor AFA Huawei AFA
Test devices: dual-controller, 1 TB cache, 16 Git/s FC frontend, NVMe backend, 25 x 3.84 TB SSDs
Test conditions: hybrid workload, 8 KB I/Os, 7:3 read/write, 1 ms average latency, 8 LUNs
Online transactions: 5x more TPS*
OceanStor
Dorado 6000 V6 57,000 TPS
Vendor E high-end 11,500 TPS
AFA Test conditions: dual-controller, 40 x 3.84 TB SSDs,
SwingBench OE2 transaction simulation system
*TPS: transactions per second
Report queries: 33% shorter batch processing
5x Better 5.6 hr 3.8 hr

User Experience
Vendor E high-end OceanStor Test conditions: dual-controller, 40 x 3.84 TB
AFA Dorado 6000 V6 SSDs, report query simulation system, 3 TB data
VDIs: 80% faster application response

5s 5.1s
4.7s
Vendor E high-end AFA
1.8s OceanStor Dorado 6000 V6
1.5s
0.9s
Test conditions: dual-controller, 100 x 3.84 TB

SSDs, 8 TB per LUN, 50 GB per VDI
3,000+ DCs
10-year stable operation on the live network
Ever Solid Applications with 5 Reliability Layers
 Gateway-free cloud backup

Cloud backup  30x higher backup frequency
 20x faster backup speed
99.9999% Solution


Gateway-free active-active solution (1 ms latency)
FlashEver without data migration
system-level reliability
 Comprehensive enterprise-class features
System  Tolerance for simultaneous failures of 3 disks
 Reconstruction of 1 TB data in 15 minutes
99.99999% Architecture
 Tolerance for failures of 7 out of 8 controllers with
SmartMatrix fully-interconnected architecture
solution-level reliability  E2E active-active design
 Global wear leveling

SSD  Huawei-patented global anti-wear leveling
Disk Reliability
— 6 SSD Technologies Forging Industry-Leading Reliability
SSD wear leveling and Huawei-patented anti-wear leveling
LDPC + SmartFSP 3.0 for error correction granularity 10x

superior to competitors
Intra-disk DIF preventing silent data corruption
Data inspection algorithm preventing data distortion
Built-in dynamic RAID improving utilization
RAID at the SSD and system levels for solid reliability
Disk Reliability
— Global Wear and Anti-Wear Leveling
Late SSD life: Huawei-patented
Early SSD life: global wear leveling
global anti-wear leveling
RAID 2.0+ improves SSD reliability by evenly distributing The workload of one SSD increases to prevent service
data to SSDs with fingerprints for wear leveling. downtime from simultaneous failures of multiple SSDs.
Prolong SSD service life and improve reliability
Disk Reliability
— LDPC + SmartFSP 3.0 for 10-18 Error Correction Granularity
G1
G1
G1
Data LDPC
encoding
Adjusts read
calibration. NAND
Calibration DSP
flash
table logic
controller
LDPC Soft-bit
Data
decoding logic
H1 LLR table
H1
H1
10x superior to competitors (10-17)
 Error correction requires an advanced LDPC algorithm (132 bits/1 KB). Huawei-developed SSDs have controller enclosures
with an LDPC algorithm that equals the algorithm of enterprise-class SSD controllers.
 Huawei controller enclosures also integrate hard decision, soft decision, and DSP.
 SmartFSP 3.0 determines the optimal read voltage so that the correct data is read simultaneously for improved read reliability.
Disk Reliability
— Intra-Disk DIF Preventing Silent Data Corruption
Data block DIF

8 bytes
CRC App tag REF tag
An REF tag is generally the first four bytes of a logical

block addressing (LBA). DIF uses the tag to prevent data
replacement faults.
Applications assign an application tag for data verification and

protection at the application level.
CRC protects the integrity of data blocks with cyclic redundancy checks.
Disk Reliability
— Data Inspection Algorithm Prevents Data Distortion
Channel 0 Channel 1 Channel 16 Channel P
Page 0 LBA 16 LBA 17 … LBA 100 Parity data
Page 3 LBA 30 LBA 128 … LBA N Parity data
Page N LBA 201 …
Checks full-disk data in the background.
Detects data correction limit and potentially distorted data.
Migrates data immediately.
Avoids data distortion and prevents the service from reading error data.
Disk Reliability
— Built-in Dynamic RAID Fully Utilizing SSDs
Conventional RAID Huawei dynamic RAID
The entire RAID group of a damaged block requires data Only a damaged block is shielded. The remaining blocks
migration and shielded space to recover data. reconstruct a new RAID group.
4. Wastes space by shielding the entire RAID group with the damaged block. 4. Reconstructs the remaining blocks into a new RAID group.
ch 0 ch 1 ch n-1 ch n ch P ch 0 ch 1 ch n-1 ch n ch P
PBA 0 16 17 … 100 60 P0 PBA 0 16 17 … 100 60 PPm+2
0
1. Detects a damaged block. 1. Detects a damaged block.
PBA 1 101 160 … 10 11 P1 PBA 1 101 160 … 10 11 P1
PBA m 3000 1280 … n n+1 Pm PBA m 3000 1280 … n n+1 Pm

2. Creates a new RAID group to store all the RAID group 2. Creates a new RAID group to store all the RAID group data.
data. P m+1
PBA m+1 PBA m+1 P m+1
3. Uses the RAID algorithm to recover the damaged block and 3. Uses the RAID algorithm to recover the damaged block and
migrates data in all blocks. migrates data in all blocks.
A large amount of flash space is wasted. Space utilization improves by 4.3%.
Disk Reliability
— RAID at the SSD and System Levels for Solid Reliability
SSD-level RAID 4: data reliability System-level RAID 5, 6, and TP:

tolerates simultaneous failures of
up to 3 disks
 SSDs at the end of their service life cannot implement disk recovery with only SSD-level RAID for faults (such as a
two-channel fault). System-level RAID is invoked to recover data.
 Faulty SSD data moves to a functional SSD to protect intra-disk OP, performance, and reliability.
Architecture Reliability
— SmartMatrix Sets a New Benchmark
BIM**
Global cache
CPU CPU CPU CPU

Ever solid architecture
FIM*
• Tolerates failures of • Tolerates failure of 1

up to 7 controllers controller enclosure
CPU
CPU
Global cache
Global cache
SmartMatrix
CPU
CPU
FIM
BIM
OceanStor
FIM
BIM
Tolerance Vendor E Vendor H
CPU
CPU
Dorado V6
CPU
CPU
1-controller
failure
2-controller
failure
FIM
CPU CPU CPU CPU 7-controller

failure
Global cache
BIM
*Front-end interconnect I/O module (FIM)
**Back-end interconnect I/O module (BIM)
— E2E Full Mesh for Service Continuity
Host I/Os
Network Network
adapter adapter
One FIM shared by 4 controllers
 A FIM connects to 4 controllers through PCIe ports to
access all the controllers in active-active mode using
FIM FIM FIM FIM multi-channel technology.
192 192 Fully interconnected controllers

cores cores
 The controllers in an enclosure are fully interconnected
through a passive backplane.
Shared Back-End Shared Back-End  Cross-enclosure expansion: 100 Gbit/s RDMA shared
BIM BIM BIM BIM 192 192
cores cores
interface modules connect to 8, 12, or 16 controllers.
2 controller enclosures connected to 1 smart

SSD enclosure
 A BIM is installed in a controller enclosure. All
controllers can simultaneously access an SSD
enclosure connected to the BIM.
 A smart SSD enclosure has 2 groups of uplink ports
that connect to 2 controller enclosures. The SSD
enclosure connects to 8 controllers.
— Tolerance of Failure of Up to 7 Controllers
FE Front-end interface
Concurrent 2-controller 1-controller-enclosure Consecutive 7-controller module
failure failure failure Ctrl Controller
BE Back-end interface
module
Cache data
Controller enclosure Controller enclosure Controller enclosure Controller enclosure Controller enclosure Controller enclosure
FE FE FE FE FE FE FE FE FE FE FE FE
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
Ctrl
BE BE BE BE BE BE BE BE BE BE BE BE
SSD enclosure SSD enclosure SSD enclosure
Industry-first:  3 cache copies on 3 controllers  3 cache copies on 2 controller  1-controller failure: cache image
3-copy across  2-controller failure: at least 1 enclosures reconstruction Industry-leading:
controller cache copy available  1-controller-enclosure failure: at  Tolerance of 7-controller failure continuous cache
enclosures least 1 cache copy available mirroring
— Industry-Leading Service Continuity
Vendor E Vendor H OceanStor Dorado
2 controllers, scale-out 4 to 8 controllers, scale-out
• I/O interface module connected to 1 • Interface module shared by controllers • Interface module shared by controllers
controller • Controller failure: 0 front-end link • Controller failure: 0 front-end link
• Controller failure: front-end link switching switching, 0 impact on hosts
switching
Front end
Controllers
Back end
• SSD enclosure connected to 2 • SSD enclosure connected to 4 • Global cache: continuous mirroring, 3 copies
controllers (1 controller enclosure) controllers (1 controller enclosure) across controller enclosures
• 2-controller failure: service disruption • 4-controller failure: service • SSD enclosure connected to 8 controllers (2
disruption controller enclosures)
• 1-controller-enclosure/2-controller/7-controller
failure: 0 service disruption
— Continuous Front-End Link in the Event of Controller Failure
Servers
Shared Shared Shared Shared
frontend frontend frontend frontend
Data
Controller Controller Controller Controller
Shared Shared Shared Shared FC or ETH

backend backend backend backend
network
• The FIMs are linked to servers independently regardless of FIM FIM FIM FIM
controller failures.
Data
• Each I/O goes to a specific storage controller through the

backplane. Back plane
• Faulty controller I/Os redirect to another functioning Controller Controller Controller Controller
controller for a continuous link between the FIM and the
server.
— Seamless Service Switchover in Seconds
Vendor E Vendor H Huawei
FE FE FE FE FE FE FE FE
Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl Ctrl
BE BE BE BE BE BE BE BE
IOPS
IOPS
IOPS
4s 6s 6s 9s > 9s 1s 1s 1s
Time Time Time
Source: Huawei lab FE: front-end interface module Ctrl: controller BE: back-end interface module
— Non-Disruptive Upgrade in Seconds with Higher Stability and Simplicity
Host
 Modular software architecture
 94% of components are in user mode and can be upgraded within 1s.
FC switch  Components on the I/O path can be upgraded within 1s.
 Continuous link and host
unaware of upgrade  FIM with the Huawei-developed Hi1822 chip:
 Connects 4 internal links to 4 controllers in an enclosure
 Provides 1 communication link to the host through the front-end port
 I/O components Hi1822  Switches the service link to another controller when any controller restarts
upgraded within 1s during an upgrade. The host is unaware of the upgrade, and the
communication link is uninterrupted.
 The firmware supports loose mappings and patch upgrades. In extreme

scenarios, the FIMs can ensure host linkup.
Competitor E Competitor H Huawei

I/O process I/O process I/O process I/O process
System mgmt. Device mgmt. Configuration

NMS process
Length (s) 3–5 > 10 1
process process mgmt. process Component-
based √
Continuous host
User mode upgrade
link
Yes Yes Yes
Length (minute) 132 90 30

Kernel mode Rolling
upgrade Continuous host
No Yes Yes
link
Ctrl 0 Ctrl 1 Ctrl 2 Ctrl 3
Restarted during the upgrade
— E2E Active-Active Design for Ever Solid Load Balancing
1 Huawei-developed multipathing software
Distributes I/Os evenly to all front-end ports

Access load balancing
Front-end interconnect I/O module Front-end interconnect I/O module

Front-end interconnect I/O module
2
Distributes I/Os evenly to all controllers
A B C D A B C D Front-end load balancing
Global cache
3
All controllers process service requests
Controller load balancing
Back-end interconnect I/O module Back-end interconnect I/O module
RAID 2.0+
4
Distributes data evenly to all SSDs
Disk load balancing
System Reliability
— Fully Redundant System Architecture for Hot-Swappable Components
Power module Management module
Interface module
BBU
System subrack
Controller
Fan module
System Reliability
— RAID-TP Provides Customized Protection for SSDs
Large-capacity SSDs lead to double the capacity (up to 32 TB) and 5x to 10x the failure rate
15 minutes
RAID-TP
Simultaneous 3-disk failure without service interruption
5 hours
Traditional RAID
Reconstruction of 1 TB data
SSD failure toleration Data reconstruction

Traditional RAID: up to 2 SSDs Traditional RAID: 5 hours
Huawei RAID-TP: simultaneous 3-SSD failure RAID-TP: 1 TB of data within 15 minutes
System Reliability
— Comprehensive Enterprise-Grade Features
Periodic Snapshot Hybrid
Multiple disk
snapshots within cascaded Cross-domain
consistency domains or dual
3 seconds snapshot and clone
group controllers
clone
Wide range of enterprise- Enhanced hardware and

grade features software reliability
Protection Forward and DR star

LUN copy consistency group reverse incremental networking Cloud backup
copy
System Reliability
— ROW and Multi-Time-Point Tech for Snapshots With Seamless Performance
1. Read-only snapshot clone
1. Creates clone with read-only snapshot
Source Clone creation with HyperCDP
HyperClone_1
TP0 HyperCDP_0 2. Unlimited snapshot rollback
HyperClone_2 Point-in-time tech for cross-level snapshot rollback
TP1 HyperSnap_0
3. Hybrid cascaded snapshot and
TP0 HyperSnap_1 2. Cross-level rollback clone
TP1
Point-in-time redirection for cascaded snapshot
TP2 TP0 HyperSnap_2 and clone
TP2 TP1
3. Cascaded snapshot HyperClone_3
TP2 3. Cascaded clone
TP0 HyperSnap_3
TP1 HyperClone_3
Snapshots with perfect performance using ROW and multi-time-point technologies

No need to copy original Data availability anytime,
Only pointers modified 0 performance loss anywhere
data
Snapshot interval in Cascaded
Cross-level rollback Writable snapshot
seconds snapshot/clone
System Reliability
— Stable Performance with Snapshot Enabled
Snapshot applications: development, testing, analytics Stable performance when continuously activating snapshots
Performance
Development (40 TB) Testing (40 TB) Analysis (80 TB)
OceanStor Dorado
Servers
Multi-copy technology
Production (10 TB) Traditional storage
Storage Snapshot-based
Development (4 TB) Testing (4 TB) Analysis (8 TB) Snapshot 1 Snapshot 2 Snapshot 3 Time
Stable performance during:

 Avoids multiple copies to improve space utilization  Snapshot activation
 Copies less data to reduce performance loss  Continuous snapshotting (interval: seconds)
 The whole lifecycle
System Reliability
— HyperCDP Has Industry-Leading Interval of 3 Seconds
3-second interval
LUN TP TP+1 TP+2 … TP+N
CDP CDP 0 CDP 1 CDP 2 … CDP N

snapshots (Max = 60,000)
Snapshot copy creation

Rollback
Snapshot copy Snapshot copy
Periodic policies with shortest interval More intensive and persistent data protection
 HyperCDP schedules by day, week, month, or specific  HyperCDP only saves the corresponding time points,
interval for customized backup. surpassing common snapshots with stronger and longer
 The multi-time-point and ROW technologies create a data protection.
copy (at a certain point in time) of a source LUN at a  A single LUN supports up to 60,000 HyperCDP snapshots.
minimum interval of 3 seconds. The minimum snapshot interval is 3 seconds, providing
 Each CDP snapshot matches a time point of the source LUN. continuous protection for three days.
Solution Reliability
— Gateway-Free Active-Active Data Centers Solution
ERP Lightning-fast and rock-solid

CRM BI
• Gateway-free
Fewer nodes and simplified management
• Active-Active
OceanStor Dorado OceanStor Dorado load balancing between sites, RPO = 0, RTO ≈ 0
HyperMetro
Gateway-free active-active
Production center A Production center B
HyperReplication HyperReplication Easy-to-scale

OceanStor • Scalability to 3DC improves reliability.
OceanStor Dorado storage system • Serial, parallel, and star networking meets the
most demanding requirements for enterprise reliability.
• Interconnection with traditional storage builds
DR center
cost-effective disaster recovery systems.
— FastWrite of the Metro Solution Accelerates Cross-Site Transmission
Traditional solution FastWrite
Host Traditional Traditional Host OceanStor OceanStor

Host Host
storage storage Dorado Dorado
100 KM 100 KM
1 Write command 8 Gbit/s FC or 1 Write command 8 Gbit/s FC or

10GE 10GE
2 Transfer ready 2 Transfer ready
3 Data transfer 3 Data transfer
RTT-1 RTT-1
6 Successful
RTT-2 transfer
8 Successful
transfer
Site A Site B Site A Site B
 One interaction
 Two interactions  100 km link transmission: One RTT, and service performance improved
 100 km link transmission: RTT ≈ (1.3 ms) x 2
by 25%.
— FlashEver Ensures Zero Data Migration for Always-On Services
New upgrade model

Individual replacement of controllers Zero data migration and Free access to next-gen hardware for
or SSD enclosures instead of
zero service disruption a 20% performance increase
replacing the entire system
OceanStor Dorado Vx OceanStor Dorado Vx+1 OceanStor Dorado Vx+2
— FlashEver Grants Seamless Services for New-Gen Hardware Upgrades
Storage controller upgrade SSD enclosure upgrade
Data in place (DIP) upgrade for zero service SSD enclosure replacement without service interruption
interruption or data migration Legacy controller enclosure
Legacy controller enclosure

Step 1
Remove an existing
Legacy SSD enclosure New SSD enclosure
controller and install
New a new controller.
controller
I/O I/O
Heterogeneous virtualization and
third-Party storage reuse
Takeover from third-party storage devices and reuse of the legacy storage
Switch services to new controllers. system. Services smoothly migrate to OceanStor Dorado devices.
.
Legacy controller enclosure
Original path New path
Step 2
Repeat step 1 to
replace all other Takeover
New controller controllers. path
Data Data
Traditional storage eDevLUN Target OceanStor Dorado

LUN
Cloud Reliability
— Converged Data Management for the Most Comprehensive DR Convergence
DC 1 DC 2 Cloud DR center
• Active-active data protection for • No backup software or gateway for • Cloud backup and recovery
uninterrupted data services less investment • Unified cloud management
• Continuous data protection within • Snapshot copy available immediately simplifies O&M
seconds for zero data loss for service recovery in minutes
Service
recovery
Second-level Minute-level Hour-level
Zero service downtime, zero data loss

3-Layer, Intelligent Data Management for Confident Investment
Planning Fault fingerprinting eService

O&M Cloud AI
Prediction Analytics Online training Intelligent decision-making
Multi-device statuses Intelligent decision-making
Automation engine Prediction engine DJ

Management Center AI
Policy mgmt. engine Analytics engine Automatic full-lifecycle management
Single device status Intelligent decision-making
Service configuration Performance indicators

Configuration Device AI DeviceManager
SLA monitoring Firmware status Simplified configuration Intelligent O&M
Streamlined, AI-Powered Device Management for Bigger Profits
One-click service provisioning for Intelligent capacity prediction (365 days) and on-demand
1 7 steps faster O&M and lower OPEX 2 capacity expansion for more precise IT investment
Create LUN Group Advanced
* Name SQL
Collection of real-time performance data for intelligent
New LUN Existing LUN 3 decision-making on the cloud
Now * Storage Pool Storage Pool_001
* Application Type Default
LUNs:
* Name Prefix * Capacity per LUN * Quantity

LUN001 500 GB 10
Map To Host group Please Select Create
OK Cancel
Automated AI Management for the Entire Lifecycle Raises Efficiency
Planning Deployment Allocation O&M Optimization
90 days in advance 50% higher

for precise planning 5x faster deployment 10x faster provisioning 90% risk prevention
resource usability
AI management automatically allocates resources throughout the entire lifecycle
OceanStor DJ — Automatic Full-Lifecycle Management
Model training
Remote O&M
Bare metal VM Container HUAWEI CLOUD 3rd-party private cloud
RESTful API
eService
Model
pushdown
Intelligent risk prediction Intelligent fault analytics
Generation Usage Protection Archiving Flow
Rest/CLI/Cinder Converged One GUI, message, and model
Automatic Provisioning and O&M
SAN NAS Object

Intelligent Prediction, analysis, and optimization
Edge-Cloud AI Synergy for Mass Data and Intelligent Resource Utilization
eService
Data Data Model Intelligent

collection cleansing training decision-making
Data reporting Predictions
Capacity 365-day capacity
Performance 60-day performance bottlenecks
Alarm information 14-day disk risks
...
190,000+ storage systems on the live network
Automatic Data Management System — AI Anywhere
Cloud AI Center AI
200,000+ storage systems 2 PB+ feature data 1,000+ app scenarios
Planning Deployment Allocation
O&M Optimization
Model pushdown
Global learning Intelligent Expert experience
decision-making
OceanStor DJ (Automatic full-lifecycle
eService (Intelligent cloud O&M system)
management platform)
Application analytics and evaluation Automatic service provisioning

Automatic training model optimization shortens deployment time
Mobile O&M anytime, anywhere

Performance bottleneck 10x 50%
prediction 60 days in advance 93% fault location rate Service provisioning efficiency Higher resource utilization
Success Stories
CITIC Bank
— 2x Faster ODS for Big Data Analytics and Decision-Making
China CITIC Bank: one of the first commercial banks in China, ranking No.27 in Tier 1 capital on the Top 1000 World
Banks 2018 list
Data source
Integrated
Operational Product Risk control Supervision and
core system
analytics management reporting
Credit card
Credit system Extraction Consolidation Analytics
Online banking Operational Data Store (ODS)
00:00 08:00 a.m. 00:00 2x faster ODS system

Off-peak Peak
Off-peak data consolidation lowers costs

!
Before 7 hours and shortens the process from 7 to 3
hours for 0 impact on big data analytics
Now and decision-making.
3 hours
BYD
— Smart Manufacturing Accelerates Data Extraction by 67%
BYD: China's largest private carmaker and new-energy vehicle leader
4.5 hours
67%
1.5 hours
BW data extraction:
4.5 hours → 1.5 hours Before Now
60 minutes
83%
10 minutes
Transfer of 1000 spare parts:

1 hour → 10 minutes Before Now
71
Bring digital to every person, home, and organization for a fully
connected, intelligent world.
Copyright © 2020 Huawei Technologies Co., Ltd.

All Rights Reserved.
The information in this document may contain predictive

statements including, without limitation, statements regarding
Thank You
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.

101-Huawei OceanStor Dorado All-Flash Storage Systems Pre-Sales Training V1.5

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

101-Huawei OceanStor Dorado All-Flash Storage Systems Pre-Sales Training V1.5

Transféré par

Droits d'auteur :

Formats disponibles

Huawei OceanStor Dorado V6

All-Flash Storage Systems

Connections Enterprise wireless FTTx and WTTx WiFi Campus networks

Unleashed data potential and value

Block File Object HDFS

Data access Diverse data connections

x86 Arm GPU NPU

Larger-scale financial Smoother 5G One-stop

Peak-time transactions Billing processing time Government service efficiency

300% 67% 50%

Huawei innovations make inclusive all-flash storage affordable

SmartMatrix architecture Power Maintenance

Data reduction for higher

All-flash storage market share

Source: Gartner Magic Quadrant for

Interop Grand Prize Verification by the ESG Lab Magnitude-9.0 earthquake

Proving that OceanStor

World's fastest OceanStor Dorado 5000,

2009 to 2013 2014 to 2016 2017 2018 2019

Ever Fast Ever Solid AI-Powered

Industry's highest SmartMatrix fully- Intelligent full-lifecycle

Seamless storage upgrades for industry-leading

FlashLink® SSD-controller synergy with intelligent algorithms for

Kunpeng 920 processor

4 Storage SSD controller chip

• Industry-leading interface chip for on-

• Offloading protocol parsing from general-

Acceleration Offloads compute-intensive storage

• Analyzes and learns I/O rules of various application

• Improves the read hit ratio and increases system

• Faster data access on SSDs with core FTL algorithm

• Write latency halved from 40 μs to 20 μs on NVMe

• Unified and intelligent fault management

• 93% fault locating accuracy

E2E Every model

*Front-end NVMe over 100 Gbit/s RDMA will be available

SAS Target NVMe 1. NVMe write command

Controller with 5 chips

Reads and Data exchange Protocol Data

CPU partitioning in controllers Core grouping in CPUs Core-based I/O binding

2x CPU processing efficiency, 20% lower latency

Controller CPU Core group

Controller enclosure Smart SSD enclosure with intelligent chips

Back-end Back-end Back-end Back-end Impact on controller

30 min/TB 15 min/TB 15% 5%

Smart SSD enclosure + Kunpeng Smart SSD enclosure + Kunpeng

80 MB/s 200 MB/s 70%

Real-time collection Online learning Feedback optimization

Private Cloud VDI scenario

Scenario: 10 linked clone VMs in VM reboot

Combined storage of metadata Separate storage of metadata

Metadata Other data

垃圾回收 1st 垃圾回收 5th

Read and write I/Os are processed first to

2x Better Test conditions: 32-controller, SPC-1; test report released in 2020.

2x higher than competitors in databases

Competitor AFA: 40% decrease; Huawei AFA: < 10% decrease

Data reduction GC RAID 6 Snapshot

Competitor AFA Huawei AFA

5x Better 5.6 hr 3.8 hr

VDIs: 80% faster application response

Test conditions: dual-controller, 100 x 3.84 TB

 Gateway-free cloud backup

 Global wear leveling