Transcript
© Copyright 2016 EMC Corporation. All rights reserved.
1
DSSD D5 TECHNICAL OVERVIEW DON WAKE – CONSULTANT TECHNICAL MARKETING ENGINEER, EMC | DSSD © Copyright 2016 EMC Corporation. All rights reserved.
2
AGENDA • Why DSSD D5?
• DSSD Hardware • DSSD Software
• HW + SW Data Protection
© Copyright 2016 EMC Corporation. All rights reserved.
3
WHY DSSD D5? INDUSTRY TRENDS DRIVING THE NEED FOR A RACK-SCALE FLASH PLATFORM © Copyright 2016 EMC Corporation. All rights reserved.
4
MORE DATA MEANS MORE COMPLEX RELATIONSHIPS TO ANALYZE IN REAL-TIME AT A LARGE SCALE
The Data Multiplier Effect
Amount of Data
Machine
44 Zb 4.4 Zb
2013
Video
Source: IDC 2014
Recording
Human
Sensors
16+ Zb Hot Data
2020
Sensor/ External Data
Satellite Imaging
VARIETY VOLUME VELOCITY
Enterprise/ External Data
Email Web Logs
VARIETY VOLUME
M2M Log Files Documents BioInformatics
Business
Social
Database Data VOLUME
1X
10X
100X
•
More Data Needs To Be Captured Faster
•
Existing Applications Are Taxed
•
Real-time Analytics For Business Insights
•
Evolving New Applications & Architectures
© Copyright 2016 EMC Corporation. All rights reserved.
5
SOLUTIONS FOR REAL-TIME WORLD
High Performance Applications Running on HDFS
High Performance DBs and DW • •
Enable Larger Data Sets, More Analysis Eliminate Excessive Tuning
Custom Applications
•
Enable Real Time Analysis
•
•
Independently Scale Compute And Storage
Provide Unprecedented Bandwidth Density
•
Reduce Time To Results and Increased Analysis
Enable Enterprise Analytics Workflow Requiring Combination Workloads Support Diverse Performance Profiles & Constraints Eliminate Need for Complex, Inefficient, Siloed Workarounds
© Copyright 2016 EMC Corporation. All rights reserved.
6
OTHER INFRASTRUCTURES USING FLASH SATA/SAS or PCIe
Fabric
SATA/SAS or PCIe
SATA/SAS or PCIe
Hybrid or All Flash Arrays • Enterprise storage features
BUT • Traditional network latencies & I/O stack bottlenecks
© Copyright 2016 EMC Corporation. All rights reserved.
Server Attached Flash • Flash hostage to individual servers • Stranded storage and data shuffling among servers • No enterprise storage features
• Limited capacity
7
DSSD – REIMAGINING INFRASTRUCTURE
Rack Scale Flash
Next Gen Hardware & Software Innovations
PerformanceOriented Architecture
© Copyright 2016 EMC Corporation. All rights reserved.
Dense & Shared Flash
Native & Flexible Data Access
Enterprise Reliability
8
DSSD HARDWARE THE DSSD D5 RACK-SCALE FLASH PLATFORM © Copyright 2016 EMC Corporation. All rights reserved.
9
DENSE AND SHARED FLASH DSSD D5 - 5U RACK SCALE FLASH PLATFORM
FLASH AND CMs 36 Flash Modules (FMs)
18 Flash Modules when Half Populated
2TB/4TB Flash Modules today Larger FMs on the roadmap
Dual Ported PCIe Gen 3 x4 per FM Dual-Redundant Control Modules (CMs)
PCIe Gen 3 Connected
© Copyright 2016 EMC Corporation. All rights reserved.
10
DENSE AND SHARED FLASH DSSD D5 - 5U RACK SCALE FLASH PLATFORM
IOMs, Fans, Power Supplies Redundant Power Supplies x4 Dual-Redundant IO Modules (IOMs) PCIe Gen 3 Connected
48 PCIe Gen 3 x4 Client Ports Per IOM Total of 96 PCIe Gen 3 x4 Client Port Connections per D5
Redundant Fan Modules x5
© Copyright 2016 EMC Corporation. All rights reserved.
11
CLIENT CONNECTIVITY
UP TO 48 HOSTS DUAL CONNECTED TO ONE D5 D5
Key Features
Client
• 1-2 DSSD Client Cards per host • DSSD I/O Cable (PCIe Gen 3 x4) BACK
connects Client Card ports to PCIe Gen 3 x 4
DSSD D5 I/O Module ports
• CRC Data Protection • Always-On Multipathing • Client connects to single D5
• Lengths of 1m, 2m, 3m & 4M copper CRC error correction between client and FMs © Copyright 2016 EMC Corporation. All rights reserved.
Client 12
DSSD FLASH MODULE (FM) • 2TB, 4TB • Future FMs will offer greater capacities and support 3D NAND and NGNVM • Industry’s first HotSwappable PCIe Gen 3 x4 based FM
• Industry’s first Dual Connected PCIe Gen 3 x4 FM • Industry’s most reliable FM • Power delivery to flash bursts up to 60 Watts
© Copyright 2016 EMC Corporation. All rights reserved.
13
DSSD FM VS OTHER FLASH STORAGE SIMPLER AND FASTER FLASH MODULES
Standard Flash Devices ASIC
ASIC
DSSD CM
ASIC Wear Leveling
Flash Physics
Garbage Collection
Cubic RAID
Defect Mgmt
FTL
ASIC
ASIC
ASIC
Vaulting
DSSD FM
ECC
Vaulting
DSSD FM
ECC
Vaulting
ECC
SSD Drive
Vaulting
FTL
Defect Mgmt
Flash Physics
SSD Drive
Garbage Collection
ECC
Wear Leveling
FTL
Vaulting
Defect Mgmt
Flash Physics
Garbage Collection
ECC
Wear Leveling
FTL
Vaulting
Defect Mgmt
Flash Physics
Garbage Collection
Wear Leveling
ECC
SSD Drive
D5
DSSD FM
• Complex firmware, limited power
• DSSD has simple, fast Flash Modules
• Independently managed media
• Control Module with rich resources implements advanced global algorithms
© Copyright 2016 EMC Corporation. All rights reserved.
14
DSSD I/O Module (IOM) – Data Plane • The IOM interconnects Client Ports, CMs and FMs • 12x PCIe Hubs
• 48 PCIe Gen 3 x4 connections to client • 36 PCIe Gen 3 x4 links to Flash Modules
• Dual-Redundant FRU • Service Module Daughter-Board
© Copyright 2016 EMC Corporation. All rights reserved.
15
DSSD Control Module (CM) – Control Plane • Manages the control plane, FMs, IOMs • Receives copy of all data writes but only for Cubic RAID calculations • Dual Redundant FRU • Appliance-wide view of all activity so more advanced flash management algorithms • Flood, the DSSD D5 software runs on the CM
© Copyright 2016 EMC Corporation. All rights reserved.
16
HIGH BANDWIDTH CUT-THROUGH DATA PATH 1
2
HOST-SIDE PCIe CONNECTIONS
48
100 GB/S TRAFFIC ASSIST
END-TO-END PCIe SWITCHED FABRIC
1
© Copyright 2016 EMC Corporation. All rights reserved.
2
CONTROLLER
36
17
PCIE MESH ARCHITECTURE I/O Module PCIe ports
Flash Modules
Control Module CPUs
I/O Module PCIe ports © Copyright 2016 EMC Corporation. All rights reserved.
18
NVME™ STORAGE PROTOCOL • NVMe is an industry-standard, low latency protocol for the fastest access to PCIe connected SSDs
MOTHERBOARD SAS
DISK or SSD
SATA/AHCI
DISK or SSD
– Within two years majority of flash SSDs will migrate to it
• DSSD NVMe Industry Firsts: – NVMe Shared Storage Platform – NVMe Access from user space – NVMe multipathing with link aggregation, hot-plug, and transparent fail-over and recovery
FLASH
The NVM Express™ design mark and NVMe™ word mark are trademarks of NVM Express, Inc. © Copyright 2016 EMC Corporation. All rights reserved.
19
DSSD SOFTWARE MAXIMIZE FLASH PERFORMANCE IN A WAY USEFUL TO APPLICATIONS - FLOOD © Copyright 2016 EMC Corporation. All rights reserved.
20
WHAT IS FLOOD? Deliver the full performance of a pool of flash in a way that is useful to applications • Flood is the Client Software, CLI, BUI
CLIENT app
• Flood is a Direct Memory API
FLOOD
• Flood is the Object Store supporting Key-Value, Directory, Block and File Object types
DSSD Client Card
PCIe
D5
• Flood is the Data Protector
IO MODULES
• Flood is the Data Manager
FLOOD
• Flood is the Appliance Manager, Appliance CLI
CONTROL MODULES Flash Modules
© Copyright 2016 EMC Corporation. All rights reserved.
21
OBJECT POOL
Everything is an Object
© Copyright 2016 EMC Corporation. All rights reserved.
22
Volumes Contain Other Objects D5 A client can attach to unlimited # of volumes
VOLUME OBJECT
© Copyright 2016 EMC Corporation. All rights reserved.
Multiple clients can attach to the same volume
23
Types of DSSD Objects VOLUME Created on the D5 then presented to the Client. It is the container that Clients create objects in.
DIRECTORY Directory (Dir) objects map names to objects. Dir is used for object management. Supports Flood ls, mkdir, rm
API
© Copyright 2016 EMC Corporation. All rights reserved.
API
KEY-VALUE Key-Value (KV) objects map keys to values. Supports Flood insert, remove, lookup, etc.
API
BLOCK Block objects are containers of fragments, I/O must be FLEN aligned, higher performance than File Objects
API and
DSSD Block Driver
FILE Similar to Block but: allow nonAligned I/O, POSIX compliance, lower performance than Block objects
API
24
NATIVE AND FLEXIBLE DATA ACCESS APPLICATION CENTRIC & PROVIDES MAXIMUM CHOICE
DSSD BLOCK DRIVER
FLOOD DIRECT MEMORY API
app
app
User Kernel
DSSD Block Driver
DSSD libflood
DSSD PCIe Client Card
DSSD PCIe Client Card
MENU
© Copyright 2016 EMC © Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
DSSD PLUG-INS app Common Application Interface DSSD PLUG-IN
DSSD PCIe Client Card
25
SOFTWARE PERFORMANCE ARCHITECTURE MEASURE PERFORMANCE TO THE APPLICATION Legacy
SOFTWARE
Application Libraries System Call
KERNEL
POSIX File System
Volume Mgr. Device Driver
HARDWARE
300µS to 5,000µS
Application Libraries
libflood
PCIe HBA
User DMA Port
SAS/SATA
PCIe Client Card
Device Controller
DSSD I/O Module
Disk/NAND
DSSD Flash Module
© Copyright 2016 EMC Corporation. All rights reserved.
<100µS
26
READ DATA FLOW – VIA FLOOD DIRECT MEMORY API Control Module
User
Application
DRAM
I/O Module
Memory
Flash Module
NAND
Flash Module
Flash Module
Flash Module
Flash Module
I/O Module
DMA I/O Cable
Control Module
OS Kernel DSSD Block Driver
DSSD Client Card
D5
Client
Reading data via libflood API accessed from an app – all in user space • DMA from FM to App memory- No software in the data path
© Copyright 2016 EMC Corporation. All rights reserved.
27
WRITE DATA FLOW – VIA FLOOD DIRECT MEMORY API Control Module
User
Application
I/O Module
Control Module
Memory
Flash Module
Flash Module
Flash Module
Flash Module
Flash Module
I/O Module
Multicast
DMA
OS Kernel DSSD Block Driver
DSSD Client Card
D5
Client
Writing data via a libflood API called from an app – all in user mode • • • •
DMA to multicast PCIe target including CM DRAM, 2 X FM DRAM No software in the data path Write acknowledged to app Data committed in grid page from FM memory to flash die
© Copyright 2016 EMC Corporation. All rights reserved.
28
DSSD BLOCK DEVICE ACCESS TO DSSD A BIT MORE LATENCY DUE TO KERNEL OVERHEAD
SOFTWARE
Application Libraries System Call
KERNEL
POSIX File System
Volume Mgr. Device Driver
HARDWARE
300µS to 5,000µS
Application Libraries
libflood
PCIe HBA
User DMA Port DSSD Block Driver
SAS/SATA
PCIe Client Card
Device Controller
DSSD I/O Module
Disk/NAND
DSSD Flash Module
© Copyright 2016 EMC Corporation. All rights reserved.
<120µS
29
WRITE DATA FLOW – VIA DSSD BLOCK DRIVER Control Module
User
Application
I/O Module
Memory
Flash Module
Flash Module
Flash Module
Flash Module
Flash Module
I/O Module
Multicast
Control Module
1 DMA
OS Kernel
2
DSSD Block Driver
DSSD Client Card
Client
D5 1
Application performs a block write to the block device presented by the kernel
2
• • • •
System Call to block driver to set up transfer DMA to multicast PCIe target including CM DRAM, 2 X FM DRAM Write acknowledged to app Data safe due to vaulting on FM
MENU
© Copyright 2016 EMC ©Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
30
HARDWARE + SOFTWARE DATA PROTECTION PROVIDE BETTER PROTECTION WITH LESS OVERHEAD AT FLASH SPEEDS © Copyright 2016 EMC Corporation. All rights reserved.
31
DSSD Hardware + Software Data Protection A SYSTEM OF RESILIENCY FEATURES FROM THE CLIENT TO THE FLASH 1. PCIe Multicast Write
2. Always-On Multipathing 3. CRC for Data in Flight 4. Dual Redundant Hardware FRUs 5. Flash Physics Control
Client
6. Space-Time GC
D5
7. Defect Avoidance 8. Enterprise ECC 9. Always-On Cubic RAID
CM0
IOM0
CM1
10. Vaulting 11. Dynamic Overprovisioning
FMs
IOM1
12. Resilvering
© Copyright 2016 EMC Corporation. All rights reserved.
32
Hardware + Software Resilience Flash Physics Control DSSD software extends the life of NAND, improves the performance of NAND and increases the efficiency of operations performed on NAND
10
00
00
01 11 NEW NAND VOLTAGES
© Copyright 2016 EMC Corporation. All rights reserved.
Vtotal
01
10 00
11
01 11
Adjust sensor levels
GOOD
10 BAD
Vtotal
GOOD
Vtotal
OLD NAND VOLTAGES 33
Hardware + Software Resilience Space-Time Garbage Collection
Incoming I/O Long-lived
Space: Improves on generic GC • Reduced Write Amplification through Fragment Level GC analysis • GC only valid data, maximum space efficiency
Time: Continuously segregates data by
observed lifetime: • Active Data (frequently-overwritten) • Stable Data (long-lived) • Active and Stable locations swapped over time by GC for data retention and wear leveling © Copyright 2016 EMC Corporation. All rights reserved.
Low Rate
Frequentlyoverwritten
High Rate
Wear Leveling
34
HARDWARE + SOFTWARE RESILIENCE Always On Cubic Raid • Cubic RAID has ~2x greater reliability of other RAID but has similar overhead (17%)
• Cubic RAID Grid is an interlocked, multidimensional array of multi-page “cells” of NAND die • High performance – always on
© Copyright 2016 EMC Corporation. All rights reserved.
System Wide Data Protection
35
D5 Hardware
Software
Data Protection
Maximizes flash performance in a space-efficient and highly available fashion
Maximizes flash performance with flash management stack in a way useful to applications - Flood
Maximize data protection with less overhead at flash speeds
© Copyright 2016 EMC Corporation. All rights reserved.
36
ADDITIONAL DSSD SESSIONS AT EMC WORLD Technology and Birds of a Feather (Discussion) Sessions Title
Speakers
DSSD D5 Rack-Scale Flash Overview: Unprecedented storage performance for mission-critical workloads
Matt McDonough, Sr. Director, Product Management and Marketing, EMC DSSD Brian Dougherty, Chief Technical Architect, CMA
DSSD D5 Technical Deep Dive: Discover the innovations that enable next-generation storage performance
Donald Wade, Technical Marketing Engineer, EMC DSSD
DSSD Ask the Founders Panel: Hear from the team that envisioned and built DSSD D5
Mike Shapiro, Co-founder and VP of Software, EMC DSSD Jeff Bonwick, Co-founder and CTO, EMC DSSD Kam Esghi, Director of Strategic Planning & Technology Alliances, EMC DSSD
Speed up and simplify your Oracle database architectures simultaneously with DSSD D5
James Morle, Principal Solutions Architect, EMC DSSD Vibhuti Bhushan, Solutions Product Manager, EMC DSSD Tony Palmer, Senior Analyst, ESG (Monday session only) Doug Babb, Chief IT Systems Architect, Systems Implementers Inc., Contractor Supporting the USAF
DSSD D5 and Cloudera: Innovations that make Hadoop faster and more efficient
Anand Bisen, Principal Solutions Architect, EMC DSSD Tony Wu, Solutions Architect, Cloudera
The impact of high performance workloads on the evolution of the modern data center
Peter Galvin, Principal Solutions Architect, EMC DSSD Maryam Sanglaji, Principal Product Manager, EMC DSSD Niall Gafney, Director of Data Intensive Computing, TACC
© Copyright 2016 EMC Corporation. All rights reserved.
Times
Location
Monday, May 2: 12:00 p.m. - 1:00 p.m.
Lando 4201A
Wednesday, May 4: 1:30 p.m. - 2:30 p.m.
Plazzo N
Tuesday, May 3: 3:00 p.m. - 4:00 p.m.
Lido 3001A
Thursday: 1:00 p.m. - 2:00 p.m.
Lando 4201A
Monday, May 2: 4:30 p.m. - 5:30 p.m.
Marcello 4403
Wednesday, May 3: 8:30 a.m. - 9:30 a.m.
Plazzo K
Monday, May 2: 1:30 p.m. - 2:30 p.m.
San Polo 3403
Thursday, May 5: 11:30 a.m. - 12:30 p.m.
Murano 3201A
Tuesday, May 3: 8:30 a.m. – 9:30 a.m.
Lido 3005
Thursday, May 5: 10:00 a.m. – 11:10 p.m.
Lido 3003
Tuesday, May 3: 1:30 p.m. – 2:30 p.m.
Plazzo L
37
© Copyright 2016 EMC Corporation. All rights reserved.
38
BACKUP CUBIC RAID © Copyright 2016 EMC Corporation. All rights reserved.
39
APPENDIX A:
CUBIC RAID DETAILS
MENU
© Copyright 2016 EMC © Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
40
TRADITIONAL RAID & CUBIC RAID
MENU
© Copyright 2016 EMC ©Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
41
STANDARD RAID DP CONSIDER 8 DISK DRIVES CONFIGURED AS RAID-6 WITH DOUBLE PARITY(DP)
D
D
D
D
D
D
P
Q
6 Disks hold data, the 7th holds Parity P and the 8th holds Parity Q This configuration can recover from any 2 disks failing
© Copyright 2016 EMC Corporation. All rights reserved.
42
Add Column Parity To Protect More Devices ROW PARITY
COLUMN PARITY
D
D
D
D
D
D
P
Q
D
D
D
D
D
D
P
Q
D
D
D
D
D
D
P
Q
D
D
D
D
D
D
P
Q
D
D
D
D
D
D
P
Q
D
D
D
D
D
D
P
Q
P
P
P
P
P
P
Q
Q
Q
Q
Q
Q
If we want more capacity we add more rows of disks The last rows provide parity for the above columns of data for more protection – 2 dimensions of parity MENU
© Copyright 2016 EMC ©Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
43
Other’s Multidimensional RAID Reconstruction C0
C1
C2
C3
C4
C5
P
Q
R0
D
D
D
D
D
D
P
Q
R1
D
D
D
D
D
D
P
Q
R2
D
D
D
D
D
D
P
Q
R3
D
D
D
D
D
D
P
Q
R4
D
D
D
D
D
D
P
Q
R5
D
D
D
D
D
D
P
Q
P
P
P
P
P
P
P
Q
Q
Q
Q
Q
Q
Q
Normally RAID 6 recovers from up to 2 bad blocks Here read of fails
© Copyright 2016 EMC Corporation. All rights reserved.
44
Other’s Multidimensional RAID Reconstruction C0
C1
C2
C3
C4
C5
P
Q
R0
D
D
D
D
D
D
P
Q
R1
D
D
D
D
D
D
P
Q
R2
D
D
D
D
D
D
P
Q
R3
D
D
D
D
D
D
P
Q
R4
D
D
D
D
D
D
P
Q
R5
D
D
D
D
D
D
P
Q
P
P
P
P
P
P
P
Q
Q
Q
Q
Q
Q
Q
Read rest of row or column to reconstruct with row or column RAID 6 parity We’ll try row R2
© Copyright 2016 EMC Corporation. All rights reserved.
45
Other’s Multidimensional RAID Reconstruction C0
C1
C2
C3
C4
C5
P
Q
R0
D
D
D
D
D
D
P
Q
R1
D
D
D
D
D
D
P
Q
R2
D
D
D
D
D
D
P
Q
R3
D
D
D
D
D
D
P
Q
R4
D
D
D
D
D
D
P
Q
R5
D
D
D
D
D
D
P
Q
P
P
P
P
P
P
P
Q
Q
Q
Q
Q
Q
Q
Fix using double parity – Done Fix using double parity - Done
© Copyright 2016 EMC Corporation. All rights reserved.
46
Other’s Multidimensional RAID Reconstruction C0
C1
C2
C3
C4
C5
P
Q
R0
D
D
D
D
D
D
P
Q
R1
D
D
D
D
D
D
P
Q
R2
D
D
D
D
D
D
P
Q
R3
D
D
D
D
D
D
P
Q
R4
D
D
D
D
D
D
P
Q
R5
D
D
D
D
D
D
P
Q
P
P
P
P
P
P
P
Q
Q
Q
Q
Q
Q
Q
Normal RAID 6 cannot handle an intersection of triple failures Neither row parity or column parity can recover
© Copyright 2016 EMC Corporation. All rights reserved.
47
DSSD CUBIC RAID
MENU
© Copyright 2016 EMC ©Corporation. Copyright 2015 All rights EMC Corporation. reserved. All rights reserved.
EMC CONFIDENTIAL
48
FLOOD MULTI-DIMENSIONAL CUBIC RAID® • 1st Dimension:
Divide each FM into 32 Fault Domains based on the channels in the FM that communicate with the NAND –
Rack-scale System Wide Data Protection
Survives channel failure, random page failure
• 2nd Dimension:
Manage the 32 Fault Domains across 18 FMs – –
Survives whole-FM failure and field upgrade Repairs errors that traditional RAID-6 could not
© Copyright 2016 EMC Corporation. All rights reserved.
49
1ST DIMENSION: DIVIDE EACH FM INTO 32 FAULT DOMAINS
© Copyright 2016 EMC Corporation. All rights reserved.
50
2nd Dimension: Manage 32 Fault Domains Across 18 FM
© Copyright 2016 EMC Corporation. All rights reserved.
51
Two Dimensional Raid Grid: 18 FM X 32 Fault Domains Per FM There are many or “n-grids” per D5 protecting the data
© Copyright 2016 EMC Corporation. All rights reserved.
52
Two Dimensional Raid Grid: 18 FM X 32 Fault Domains via FM 32 FAULT DOMAINS PER FM Data
P Q
Column Parity Row Parity
Defect
18 FMs
Intersection Parity
Each Square Is A Flash Die Multipage Cell Packed With Fragments
Recovered Data
P Q
© Copyright 2016 EMC Corporation. All rights reserved.
53
2D Cubic RAID Reconstruction
2D Cubic RAID CAN recover from this because we interlock dimensions First repair any recoverable rows © Copyright 2016 EMC Corporation. All rights reserved.
54
2D Cubic RAID Reconstruction
Now recover the columns © Copyright 2016 EMC Corporation. All rights reserved.
55
2D Cubic RAID Reconstruction – worst case unrecoverable failure pattern 32 FAULT DOMAINS PER FM Data
P Q
Column Parity Row Parity
Defect
18 FMs
Intersection Parity
Each Square Is A Flash Die Multipage Cell Packed With Fragments
Recovered Data
P Q
Three intersecting rows and columns with three failures each © Copyright 2016 EMC Corporation. All rights reserved.
56
2D Cubic RAID Reconstruction – recoverable failure pattern!
Incredibly powerful. This is recoverable. © Copyright 2016 EMC Corporation. All rights reserved.
57
2D Cubic RAID Reconstruction
First find and repair columns with 2 or fewer bad cells © Copyright 2016 EMC Corporation. All rights reserved.
58
2D Cubic RAID Reconstruction
Now recover rows with 2 or fewer bad cells © Copyright 2016 EMC Corporation. All rights reserved.
59
2D Cubic RAID Reconstruction
Now more columns are recoverable (reduced errors to <=2) © Copyright 2016 EMC Corporation. All rights reserved.
60
2D Cubic RAID Reconstruction
Back to the rows © Copyright 2016 EMC Corporation. All rights reserved.
61
2D Cubic RAID Reconstruction
A final iteration on the columns © Copyright 2016 EMC Corporation. All rights reserved.
62
2D Cubic RAID Reconstruction
Two dimensional Cubic RAID recovers everything © Copyright 2016 EMC Corporation. All rights reserved.
63
2D Cubic RAID Reconstruction
Two dimensional Cubic RAID recovers everything © Copyright 2016 EMC Corporation. All rights reserved.
64
Understanding Cubic RAID – 3 Dimensions
GRID PARITY We can stack these grids of flash die into a “cube” - again with the last 2 grids reserved for parity of the grids stacked above them © Copyright 2016 EMC Corporation. All rights reserved.
65
FLASH DIE FAULT DOMAINS (16)
Understanding Cubic RAID – Protection • Cubic RAID interlocks 3 dimensions to protect data • With Cubic RAID, a perfect cube of 27 flash cells is the smallest failure scenario • Other three dimensional RAID systems do not interlock dimensions and therefore the smallest failure scenario is only 7 flash cells • D5 ships with 2D Cubic RAID FLASH MODULES (18)
If deployed in three dimension mode, Cubic RAID does stripes across Fault Domains within an FM, across FMs, and down flash die within each Fault Domain
© Copyright 2016 EMC Corporation. All rights reserved.
66
A More Realistic View Of A RAID Grid 32 FAULT DOMAINS PER FM Data Column Parity
Intersection Parity Defect
18 FMs
Row Parity
Each Square Is A Flash Die Multipage Cell Packed With Fragments
Column, Row & Intersection parity are scattered within rows and columns for wear leveling
© Copyright 2016 EMC Corporation. All rights reserved.
67
