Page 1
Scaling capacity and performance without compromise using SGI® DMFTM
Robert Mollard
Senior Storage Specialist, APAC
* = Some names and brands may be claimed as the property of others
Capacity, Performance & Reliability
Intel Enterprise Edition of Lustre* HSM
Page 2
• Hierarchical Storage Management
• Lustre Scalability with DMF (HSM)
• Tiered Data Management
• DMF – Start small and grow
• DMF Direct Archiving
• JBFS Fast Mount Cache
• Summary
Agenda
2 ©2014 SGI SGI Company Proprietary
* = Some names and brands may be claimed as the property of others
Page 3
HSM | Data Migration Facility (DMF)
3 ©2014 SGI SGI Company Proprietary
Hierarchical Storage ManagementTransparently migrate data to Tape, MAID or Cloud
Lustre*
DMF
* = Some names and brands may be claimed as the property of others
Data life cycle management
- DMF manages the placement of data within multiple tiers of storage
Automated data migration
- From expensive, production disk to 2nd or 3rd tier storage
Transparent to user
- All data appears on line all the time
Key Benefits
- DMF reduces tier 1 disk investment
- DMF reduces power consumption
- DMF protects data long term
SGI® DMF™ 25 years in production Cloud
Page 4
SSD SSD…
DMF MDS
DMF Managed HSM Environment
Back-End I/O
TAPE …
Parallel DMF Data MoverBuilding Block
JBFS TAPE …
Parallel DMF Data MoverBuilding Block
JBFS
4 ©2014 SGI SGI Company Proprietary
TAPE …
Parallel DMF Data MoverBuilding Block
…JBFSTAPE …
Parallel DMF Data MoverBuilding Block
JBFS
Lustre* Clients
Lustre* MDS/MDTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* Clients Lustre* Clients
Scalability without compromise
Front-End I/O
… … …Scale Here
…Lustre* OSS/OST
Building Block
Scale Here
Scale Here
Lustre* MDS/MDTBuilding Block
Lustre* MDS/MDTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
More File Systems
TAPE …
Parallel DMF Data MoverBuilding Block
JBFS
* = Some names and brands may be claimed as the property of others
Lustre* Filesystem
Capacity, Performance & Reliability
DMF Direct Archiving
Page 5
Seamless Tiered Data Management
5 ©2014 SGI SGI Company Proprietary
The most recent and active data is “live” in Lustre* and mirrored
within DMF.ALL DATA APPEARS ONLINE to users.
“Overflow” data is stored and protected
within DMF on various cost-correct
media types
* = Some names and brands may be claimed as the property of others
*
*
Page 6
HSM perspective: regular file
User perspective: online file
Tiered Data Management
6 ©2014 SGI SGI Company Proprietary
JBFS
Lustre*
DMF
DMF
* = Some names and brands may be claimed as the property of others
Lustre*
DMF
Before migrating
Page 7
HSM perspective: dual-state file
User perspective: online file
Tiered Data Management
7 ©2014 SGI SGI Company Proprietary
JBFS
Lustre*
DMF
DMF
* = Some names and brands may be claimed as the property of others
After migrating
Page 8
HSM perspective: offline file
User perspective: online file
Tiered Data Management
8 ©2014 SGI SGI Company Proprietary
JBFS
Lustre*
DMF
DMF
* = Some names and brands may be claimed as the property of others
After freeing space
Page 9
HSM perspective: unmigrating file
User perspective: online file
Tiered Data Management
9 ©2014 SGI SGI Company Proprietary
JBFS
Lustre*
DMF
DMF
* = Some names and brands may be claimed as the property of others
Recalling file data from cache
Page 10
HSM perspective: unmigrating file
User perspective: online file
Tiered Data Management
10 ©2014 SGI SGI Company Proprietary
JBFS
Lustre*
DMF
DMF
* = Some names and brands may be claimed as the property of others
Recalling file data from cache
Page 11
DMF EvolutionStart small and grow
Page 12
Client Access Space
Expand Front-end I/O
H/A Node
High Availability
Remote disk target via DMF
FTP, NFS, Cloud
CXFS SAN Native Client
Nodes
CXFS
Client
Edge
Server
CXFS
Client
Edge
Server
NFS/CIFS NFS/CIFSFront-End I/O
Back-End I/O
DMF
ServerIntegrated
Data Mover
NFS/CIFS
Lustre* HSM with DMF, direct to Tier 2/3
Lustre* MDS/MDTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
Native Lustre* Clients
…
CXFS
Client Node
ParallelData Mover
CXFS
Client Node
ParallelData Mover
CXFS Node
CXFS MDS
H/A Node
DMF
ServerIntegrated
Data Mover
Expand Back-end I/O
* = Some names and brands may be claimed as the property of others
DMF Direct Archiving
DMF starts small and grows with you…Monolithic
H/A Node
Page 13
13 ©2014 SGI SGI Company Proprietary
DMF Servers
OSS 4
MDS 1 MDS 2
OSS 1
OSS 2
OSS 3
DMF Managed Data
DMF Direct Archiving | Data Flow
Lustre* Clients
Lustre* OSS
Storage
Lustre* MDS
…
DMF MDS 1
DMF MDS 2
DMF Data Mover
* = Some names and brands may be claimed as the property of others
Metadata
Data
Logical PathPhysical Path
PrimaryStorage
TAPE
JBFS DMF Tier2
DMF Tier3
Page 14
Lustre* HSM | Communication & Data Flow
Lustre* HSM CommunicationsDMF CommunicationsDMF Metadata updateDMF Data
TAPE …
Parallel DMF Data MoverBuilding Block
TAPE …
Parallel DMF Data MoverBuilding Block
TAPE TAPE…
Parallel DMF Data MoverBuilding Block
… TAPE …
Parallel DMF Data MoverBuilding Block
Lustre* HSMClient Agent
PolicyEngine DB
SSD SSD…
DMF MDS
DMF Managed HSM Environment
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
…Lustre* OSS/OST
Building BlockLustre* OSS/OST
Building Block
copytool
Lustre* MDS/MDTBuilding Block
Coordinator
Lustre* Clients Lustre* Clients
… …
Lustre* Clients
…
Parallel Data Mover Option• Data migration from multiple
parallel servers
• Scales I/O performance
• Add Additional data movers
as required
* = Some names and brands may be claimed as the property of others
JBFS JBFS JBFS
DMF Direct Archiving
Page 15
JBFS | The OpenVault VTL for DMF
15 ©2014 SGI SGI Company Proprietary
JBFS is an acronym for JBOD File System
JBFS provides mounting services
- Serialised access to disk media
- Independent from Linux disk mounts and file systems
Disks treated like tapes mounted in tape drives
The primary advantages
- Mount performance
- Low-cost scalable data throughput performance
- Power management via ZeroWatt™
Any Number of LUNs
Page 16
XFS(or any typical file system)
DMF Preferred JBFS Provides
x A large number of objects Small number of objects Small number of objects
x Object sizes change Object sizes fixed Object sizes fixed
x Flexible object organisation Fixed object organisation Fixed object organisation
x Primarily random access Primarily sequential access Primarily sequential access
x Bursty access Sustained access Sustained access
x Mount/dismountinfrequently
Mount/dismount frequently
Mount/dismount frequently
JBFS | Why A New File System?
16 ©2014 SGI SGI Company Proprietary
Page 17
• Recoverability
• Data Assurance
• High Performance
• Flexibility
• Power Management with Zero-Watt™
• JBFS API (Same as SGI Copan)
JBFS | Additional Benefits
17 ©2014 SGI SGI Company Proprietary
OpenVault includes a new DCP and a new LCP to manage JBFS volumes
Page 18
JBFS | Disk Structure
18 ©2014 SGI SGI Company Proprietary
Preparing a disk device for use with JBFS consists of three basic steps:
1. Apply the GPT labels
2. Apply the XVM labels
3. Apply the JBFS format
XVM volume name must be "JBFS_{lib}_{PCL}”
– JBFS – Fixed
– {lib} – OpenVault Library Name
– {PCL} – unique 6-character value [0-9A-Z]
Page 19
JBFS JBFS JBFS
JBOD
JBFS JBFS JBFS
JBOD
JBFS JBFS JBFS
JBOD
JBFS JBFS JBFS
JBOD
JBFS JBFS JBFS
JBOD
JBFS JBFS JBFS
JBOD
TAPE TAPE
TAPE TAPE
TAPE TAPE
TAPE TAPE
TAPE TAPE
TAPE TAPE
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* MDS/MDTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* OSS/OSTBuilding Block
Lustre* Native Clients
SAS Switch
FC Switch
SAS Switch
FC Switch
* = Some names and brands may be claimed as the property of others
Lustre* Filesystem
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
pDMF Data Mover
SSD SSD…DMF HA MDS
Lust
re*
Net
wo
rk
cop
yto
ol
Page 20
• New Lustre* and DMF features allow cost effective scalability without compromising performance
• SGI DMF provides a high performance parallel HSM for Lustre* with direct archiving to tier 2/3 storage targets
• SGI DMF – JBFS delivers a tier 2 fast mount cache with built in power management capabilities
• The Result:
– Cost effective capacity, reduced TCO (low cost/power storage tiers)
– Proven long-term data protection (DMF – 25 years in production)
– Improved operational procedures (simplified access to data)
– Scalable performance within archive tiers (parallel DMF)
Summary and Key Points
20 ©2014 SGI SGI Company Proprietary
* = Some names and brands may be claimed as the property of others
$ = on supported hardware
$
Page 21
Questions & Responses
21 ©2014 SGI SGI Company Proprietary
Robert MollardSenior Storage Specialist, Asia [email protected]
Page 22
©2014 SGI SGI Company Proprietary22