虛擬化技術

Virtualization and Virtual Machines

Storage Virtualization
 Agenda
• Overview
 Introduction
 What to be virtualized
 Where to be virtualized
 How to be virtualized
• Case study
 On linux system
• RAID
• LVM
• NFS
 In distributed system
• Vastsky
• Lustre
• Ceph
• HDFS
 Overview
• Introduction
• What to be virtualized ?
 Block, File system
• Where to be virtualized ?
 Host-based, Network-based, Storage-based
• How to be virtualized ?
 In-band, Out-of-band
• Introduction
• What to be virtualized
• Where to be virtualized
• How to be virtualized
• Case study
STORAGE VIRTUALIZATION
 Introduction
• Common storage architecture :
 DAS - Direct Attached Storage
• Storage device was directly attached to a
server or workstation, without a storage
network in between.
 NAS - Network Attached Storage
• File-level computer data storage
connected to a computer network
providing data access to heterogeneous
clients.
 SAN - Storage Area Network
• Attach remote storage devices to servers
in such a way that the devices appear as
locally attached to the operating system.
 Introduction
• Desirable properties of storage virtualization:
 Manageability
• Storage resource should be easily configured and deployed.
 Availability
• Storage hardware failures should not affect the application.
 Scalability
• Storage resource can easily scale up and down.
 Security
• Storage resource should be securely isolated.
 Introduction
• Storage concept and technique
 Storage resource mapping table
 Redundant data
 Multi-path
 Data sharing
 Tiering
 Concept and Technique
• Storage resource mapping table
 Maintain tables to map storage resource to target.
 Dynamic modify table entries for thin provisioning.
 Use table to isolate different storage address space.
 Concept and Technique
• Redundant data
 Maintain replicas to provide high availability.
 Use RAID technique to improve performance and availability.
 Concept and Technique
• Multi-path
 A fault-tolerance and performance
enhancement technique.
 There is more than one physical path
between the host and storage devices
through the buses, controllers,
switches, and bridge devices
connecting them.
 Concept and Technique
• Data sharing
 Use data de-duplication technique to eliminate duplicated data.
 Save and improve the usage of storage space
 Concept and Technique
• Tiering
 Automatic migrate data across storage resources with different
properties according to the significance or access frequency of data.
 Example: iMac fusion drive

Storage Policies Access Group

• Introduction
• What to be virtualized
• Where to be virtualized
• How to be virtualized
• Case study
STORAGE VIRTUALIZATION
 What To Be Virtualized
• Layers can be virtualized

User Space
 File system
• Provide compatible system call Application
interface to user space
applications. System call interface
 Block device

Kernel Space
File System
• Provide compatible block device
interface to file system.
Block interface
• Through the interface such as
SCSI, SAS, ATA, SATA, etc. Device driver

Storage Device
 File System Level
• Data and Files
 What is data ?
• Data is information that has been converted to a machine-readable,
digital binary format.
• Control information indicates how data should be processed.
• Applications may embed control information in user data for formatting or
presentation.
• Data and its associated control information is organized into discrete units
as files or records.
 What is file ?
• Files are the common containers for user data, application code, and
operating system executables and parameters.
 File System Level
• About the files
 Metadata
• The control information for file management is known as metadata.
• File metadata includes file attributes and pointers to the location of file
data content.
• File metadata may be segregated from a file's data content.
• Metadata on file ownership and permissions is used in file access.
• File timestamp metadata facilitates automated processes such as backup
and life cycle management.
 Different file systems
• In Unix systems, file metadata is contained in the i-node structure.
• In Windows systems, file metadata is contained in records of file attributes.
 File System Level
• File system
 What is file system ?
• A file system is a software layer responsible for organizing and policing the
creation, modification, and deletion of files.
• File systems provide a hierarchical organization of files into directories and
subdirectories.
• The B-tree algorithm facilitates more rapid search and retrieval of files by
name.
• File system integrity is maintained through duplication of master tables,
change logs, and immediate writes off file changes.
 Different file systems
• In Unix, the super block contains information on the current state of the
file system and its resources.
• In Windows NTFS, the master file table contains information on all file
entries and status.
 File System Level
• File system level virtualization
 File system maintains metadata
(i-node) of each file.
 Translate file access requests to
underlining file system.
 Sometime divide large file into small
sub-files (chunks) for parallel access,
which improves the performance
 Block Device Level
• Block level data
 The file system block
• The atomic unit of file system management is the file system block.
• A file's data may span multiple file system blocks.
• A file system block is composed of a consecutive range of disk block
addresses.
 Data in disk
• Disk drives read and write data to media through cylinder, head, and
sector geometry.
• Microcode on a disk translates between disk block numbers and
cylinder/head/sector locations.
• This translation is an elementary form of virtualization.
 Block Device Level
• Block device interface
 SCSI (Small Computer System Interface)
• The exchange of data blocks between the host system and storage is
governed by the SCSI protocol.
• The SCSI protocol is implemented in a client/server model.
• The SCSI protocol is responsible for block exchange but does not define
how data blocks will be placed on disk.
• Multiple instances of SCSI client/server sessions may run concurrently
between a server and storage.
 Block Device Level
• Logical unit and Logical volume
 Logical unit
• The SCSI command processing entity within the storage target represents a
logical unit (LU) and is assigned a logical unit number (LUN) for identification
by the host platform.
• LUN assignment can be manipulated through LUN mapping, which
substitutes virtual LUN numbers for actual ones.
 Logical volume
• A volume represents the storage capacity of one or more disk drives.
• Logical volume management may sit between the file system and the device
drivers that control system I/O.
• Volume management is responsible for creating and maintaining metadata
about storage capacity.
• Volumes are an archetypal form of storage virtualization.
 Block Device Level
• Data block level virtualization
 LUN & LBA
• A single block of information is
addressed using a logical unit
identifier (LUN) and an offset within
that LUN, which known as a Logical
Block Address (LBA).
 Apply address space remapping
• The address space mapping is
between a logical disk and a logical
unit presented by one or more
storage controllers.
• Introduction
• What to be virtualized
• Where to be virtualized
• How to be virtualized
• Case study
STORAGE VIRTUALIZATION
 Where To Be Virtualized
• Storage interconnection
 The path to storage
• The storage interconnection provides the data path between
servers and storage.
• The storage interconnection is composed of both hardware and
software components.
• Operating systems provide drivers for I/O to storage assets.
• Storage connectivity for hosts is provided by host bus adapters
(HBAs) or network interface cards (NICs).
 Where To Be Virtualized
• Storage interconnection protocol
 Fibre Channel
• Usually for high performance requirements.
• Supports point-to-point, arbitrated loop, and fabric interconnects.
• Device discovery is provided by the simple name server (SNS).
• Fibre Channel fabrics are self-configuring via fabric protocols.
 iSCSI ( internet SCSI )
• For moderate performance requirements.
• Encapsulates SCSI commands, status and data in TCP/IP.
• Device discovery by the Internet Storage Name Service (iSNS).
• iSCSI servers can be integrated into Fibre Channel SANs through IP storage
routers.
 Where To Be Virtualized
• Abstraction of physical storage
 Physical to virtual
• The cylinder, head and sector geometry of individual disks is virtualized
into logical block addresses (LBAs).
• For storage networks, the physical storage system is identified by a
network address / LUN pair.
• Combining RAID and JBOD assets to create a virtualized mirror must
accommodate performance differences.

 Metadata integrity
• Storage metadata integrity requires redundancy for failover or load
balancing.
• Virtualization intelligence may need to interface with upper layer
applications to ensure data consistency.
 Where To Be Virtualized
• Different approaches :
 Host-based approach
• Implemented as a software
running on host systems.
 Network-based approach
• Implemented on network devices.
 Storage-based approach
• Implemented on storage target
subsystem.
 Host-based Virtualization
• Host-based approach
 File level
• Run virtualized file system on the
host to map files into data blocks, Block 1
Sub-file Block 2 Block 1
Sub-file Block 2Sub-fileBlock 1
which distributed among several 1 2 3

storage devices.
 Block level
• Run logical volume management
software on the host to intercept I/O
requests and redirect them to
storage devices.
 Provide services
• Software RAID
 Host-based Virtualization
• Important issues
 Storage metadata servers
• Storage metadata may be shared by multiple servers.
• Shared metadata enables a SAN file system view for multiple servers.
• Provides virtual to real logical block address mapping for client.
• A distributed SAN file system requires file locking mechanisms to preserve
data integrity.
 Host-based storage APIs
• May be implemented by the operating system to provide a common
interface to disparate virtualized resources.
• Microsoft's virtual disk service (VDS) provides a management interface for
dynamic generation of virtualized storage.
 Host-based Virtualization

• A typical example :
 LVM
• Software layer between the file
system and the disk driver.
• Executed by the host CPU.
• Lack hardware-assist for functions
such as software RAID.
• Independence from vendor-specific
storage architectures.
• Dynamic capacity allocation to
expand or shrink volumes.
• Support alternate pathing for high
availability.
 Host-based Virtualization
• Host-based implementation
 Pros
• No additional hardware or infrastructure requirements
• Simple to design and implement
• Improve storage utilization
 Cons
• Storage utilization optimized only on a per host base
• Software implementation is dependent to each operating system
• Consume CPU clock cycle for virtualization
 Examples
• LVM, NFS
 Network-based Virtualization
• Network-based approach
 File level
• Seldom implement file level Block 1 Block 2 Block 1 Block 2 Block 1

virtualization on network device.

 Block level
• Run software on dedicated
appliances or intelligent switches
and routers.
 Provide services
• Multi-path
• Storage pooling
 Network-based Virtualization
• Requirements of storage network
 Intelligent services
• Logon services
• Simple name server
• Change notification
• Network address assignment
• Zoning
 Fabric switch should provide
• Connectivity for all storage transactions
• Interoperability between disparate servers,
operating systems, and target devices
 Network-based Virtualization

• Techniques for fabric switch virtualization

 Hosted on departmental switches
• A PC engine provisioned as an option blade.
 Data center directors
• Should be able to preserve the five nines availability characteristic of
director-class switches.
• Dedicated virtualization ASICs provide
high-performance frame processing
and block address mapping.
 Interoperability between
different implementations
will become a priority.
 Network-based Virtualization
• Interoperability issue
 FAIS ( Fabric Application Interface Standard )
• Define a set of standard APIs to integrate applications and switches.
• FAIS separates control information and data paths.
• The control path processor (CPP) supports the FAIS APIs and upper layer
storage virtualization application.
• The data path controller (DPC) executes the virtualized SCSI I/Os under the
management of one or more CPPs
 Network-based Virtualization
• Network-based implementation
 Pros
• True heterogeneous storage virtualization
• No need for modification of host or storage system
• Multi-path technique improve the access performance
 Cons
• Complex interoperability matrices - limited by vendors support
• Difficult to implement fast metadata updates in switch device
• Usually require to build specific network equipments (e.g., Fibre Channel)
 Examples
• IBM SVC ( SAN Volume Controller ), EMC Invista
 Storage-based Virtualization
• Storage-based approach
 File level Block 1
Block 1 Block 1
• Run software on storage device to
provide file based data storage
services to host through network.
 Block level
• Embeds the technology in the target
storage devices.
 Provide services
• Storage pooling
• Replication and RAID
• Data sharing and tiering Replica
Block 1
Sub-file Replica
Block 1
Sub-file Replica
Block 1
Sub-file
1.bak
1 2.bak
2 3
 Storage-based Virtualization
• Array-based virtualization
 Storage controller
• Provide basic disk virtualization in the form of RAID management,
mirroring, and LUN mapping or masking.
• Allocate a single LUN to multiple servers.
• Offer Fibre Channel, iSCSI,
and SCSI protocol.
 Cache memory
• Enhance performance.
 Storage assets coordination
• Coordination between
multiple storage systems
is necessary to ensure high
availability.
 Storage-based Virtualization
• Data replication
 Array-based data replication
• Referred to as disk-to-disk replication.
• Requires that a storage controller function concurrently as both an
initiator and target.
 Synchronous vs. Asynchronous
• Synchronous data replication ensures that a write operation to a
secondary disk array is completed before the primary array
acknowledges task completion to the server.
• Asynchronous data replication provides write completion by the
primary array, although the transaction may still be pending to the
secondary array.
 Storage-based Virtualization
Synchronous Asynchronous
To preserve performance, synchronous data Asynchronous data replication is largely
replication is limited to metropolitan distances immune to transmission latency
 Storage-based Virtualization
• Other features
 Point-in-time copy ( snapshot )
• Provide point-in-time copies of an entire storage volume.
• Snapshot copies may be written to secondary storage arrays.
• Provide an efficient means to quickly recover a known good volume state
in the event of data from the host.
 Distributed modular virtualization
• Decoupling storage controller logic from physical disk banks provides
flexibility for supporting heterogeneous disk assets and facilitates
distributed virtualization intelligence.
• Accommodates class of storage services and data lifecycle management.
 Storage-based Virtualization
Distributed Modular Virtualization
Decoupling storage controller intelligence and virtualization engines from
physical disk banks facilitates multi-protocol block data access and
accommodation of a broad range of disk architectures.
 Storage-based Virtualization
• Storage-based implementation
 Pros
• Provide most of the benefits of storage virtualization
• Reduce additional latency to individual IO
 Cons
• Storage utilization optimized only across the connected controllers
• Replication and data migration only possible across the connected
controllers and the same vendors devices
 Examples
• Disk array products
• Introduction
• What to be virtualized
• Where to be virtualized
• How to be virtualized
STORAGE VIRTUALIZATION
 In-band Virtualization
• Implementation methods :
Control Control Control
 In-band Message Message Message

• Also known as symmetric,

virtualization devices actually sit in
the data path between the host
and storage.
• Hosts perform IO to the virtualized
device and never interact with the
Control
actual storage device. Message

 Pros
• Easy to implement
 Cons
• Bad scalability & Bottle neck

Data Data Data

 Out-of-band Virtualization
• Implementation methods :
 Out-of-band
• Also known as asymmetric, Control Control Control
Message Message Message
virtualization devices are
sometimes called metadata
servers.
• Require additional software in the
host which knows the first request
location of the actual data. Control
Message
 Pros
• Scalability & Performance
 Cons
• Hard to implement
Data Data Data
 Other Virtualization Services
Pooling Heterogeneous
Heterogeneous Mirroring
Storage Assets
In a virtualized storage pool, virtual assets may be Heterogeneous mirroring offers more flexible options
dynamically resized and allocated to servers by than conventional mirroring, including three-way
drawing on the total storage capacity of the SAN mirroring within storage capacity carved from
different storage systems
 Other Virtualization Services
Heterogeneous Data Replication
Heterogeneous data replication enables duplication of storage data
between otherwise incompatible storage systems.
 Summary
• Storage virtualization technique :
 Virtualization layer
• File level and block level
 Virtualization location
• Host, network and storage base
 Virtualization method
• In-band and out-of-band
• Storage virtualization services
 Storage pooling and sharing
 Data replication and mirroring
 Snapshot and multi-pathing
• Introduction
• What to be virtualized
• Where to be virtualized
• How to be virtualized
• Case study
STORAGE VIRTUALIZATION
• Case-study, virtualization on linux system
• Block-based
• Redundant Array of Independent Disks (RAID)
• Logical Volume Management (LVM)
• File-based
• Network File System (NFS)

STORAGE VIRTUALIZATION
ON LINUX SYSTEM
 RAID
• RAID (redundant array of independent disks)
• Originally: redundant array of inexpensive disks
 RAID schemes provide different balance between the key goals:
• Reliability
• Availability
• Performance
• Capacity
 RAID level
• The most used:
 RAID0
• block-level striping without parity or mirroring
 RAID1
• mirroring without parity or striping
 RAID1+0
• referred to as RAID 1+0, mirroring and striping
 RAID2
 RAID3
 RAID4
 RAID5
• block-level striping with distributed parity
 RAID5+0
• referred to as RAID 5+0, distributed parity and striping
 RAID6
 RAID 0
• RAID 0: Block-level striping
without parity or mirroring
 It has no (or zero) redundancy.
 It provides improved performance
and additional storage
 It has no fault tolerance. Any drive
failure destroys the array, and the
likelihood of failure increases with
more drives in the array.

figure from: http://storage-system.fujitsu.com/jp/term/raid/

 RAID 1
• RAID 1: Mirroring without parity or
striping
 Data is written identically to two drives, thereby
producing a "mirrored set";
 A read request is serviced by one of the two
drives containing with least seek time plus
rotational latency.
 A write request updates the stripes of both drives.
The write performance depends on the slower of
the two.
 At least two drives are required to constitute such
an array.
 The array continues to operate as long as at least
one drive is functioning.
• Space efficiency
 1/N
• N=2
• Fault tolerance
 N–1
• N=2

figure from: http://storage-system.fujitsu.com/jp/term/raid/

 RAID 5
• RAID5: Block-level striping with
distributed parity
 distributes parity on different disk
 requires at least 3 disks
• Space efficiency
 1 − 1/N
• Fault tolerance
 1

figure from: http://storage-system.fujitsu.com/jp/term/raid/

 RAID 1+0 / RAID 5+0
RAID 1+0 RAID 5+0
RAID1(mirror) + Stripe RAID5(parity) + Stripe

figure from: http://storage-system.fujitsu.com/jp/term/raid/

 RAID Level Comparison
RAID level Reliability Write Performance Space efficiency
RAID 0 × ○ 2/2
RAID 1 ◎ ○ 1/2
RAID 1+0 ◎ ○ 2/4
RAID 5 ○ △ 2/3
RAID 5+0 ○ ○ 4/6

table from: http://storage-system.fujitsu.com/jp/term/raid/

 Logical Volume Management
• LVM architecture
 Logical Volume Management
• LVM project is implemented in two components:
 In user space
• Some management utilities and configuration tools
Ex. lvm , dmsetup
• Programming interface with a well-designed library
Ex. libdevmapper.h
 In kernel space
• Implement device mapper framework
• Provide different mapped device targets
Ex. linear , stripe , mirror …etc.
 Logical Volume Management
• Tools and utilities are in user space.
 Logical Volume Management
• lvm
 Command-line tools for LVM2.
• logical volume ( lv ) operations
• volume group ( vg ) operations
• physical volume ( pv ) operations
 Limited controllability
• Only can create logical volume with simple
mapping mechanisms.
• Do not allow cross machine mappings.

• dmsetup
 Limitations
• Still cannot provide cross machine mappings.
 Logical Volume Management
• dmsetup
 low level logical volume management
• Operate create, delete, suspend and resume …etc
• Work with mapping table file
 Logical Volume Management
• File system will build upon device mapper framework
by means of system calls.
 Logical Volume Management
• File system in operating system will invoke a set of block
device system calls.

Device Mapper framework

reload operation functions
 Logical Volume Management
• File system can be also implemented in the user space only.
 Logical Volume Management
• Device mapper framework implements a Linux kernel
driver for different mappings.
 Logical Volume Management
• Device mapper framework defines a set of target device
mapping interfaces.
 dm_ctr_fn ctr
• Initiator of each newly created mapped device
 dm_dtr_fn dtr
• Destructor of each removing mapped device
 dm_map_fn map
• Setup the mapping relations
 dm_ioctl_fn ioctl
• Exactly perform system IO invocations
 … etc.
 Logical Volume Management
• Develop a new mapped device target and add it
into device mapper framework.
 Improve scalability
 Network File System
• NFS architecture
 Network File System
• What is NFS ?
 NFS is a POSIX-compliant distributed file system
• Work distributedly as server-client model
 NFS builds on the Remote Procedure Call (RPC) system.
 The Network File System is an open standard defined in RFCs.
• Some features :
 Shared POSIX file system
 Common module in linux kernel
 well performance
 Network File System
• Dynamic Port and its handle
 In NFSv3, service listens on random tcp port.
 NFS use RPC(Remote Procedure Call) to get the port of service.
 Network File System
• Consistency and concurrency in NFS
 Lockd offers a write lock to handle concurrent update.
 Statd handles the consistency between server and clients.
• Case-study, virtualization in distributed system
• Block-based
• VastSky
• File-based
• Lustre
• Object-based
• Ceph
• HDFS

STORAGE VIRTUALIZATION
IN DISTRIBUTED SYSTEM
 VastSky
• Overview
 VastSky is a linux-based cluster storage system, which provides logical
volumes to users by aggregating disks over a network.
• Three kinds of servers
 storage manager
• Maintaining a database which describes physical and logical resources in a system.
• e.g. create and attach logical volumes.
 head servers
• Running user applications or virtual machines which actually use VastSky logical
volumes.
 storage servers
• Storage servers have physical disks which are used to store user data.
• They are exported over the network and used to provide logical volumes on head
servers. (iSCSI)
 VastSky
• VastSky Architecture

XML-RPC Storage
Manager

iSCSI request

Storage Pool
 VastSky
• Logical Volume
 a set of several mirrored disks
 several physical disk chunks on different servers
Storage Pool

Logical Volume
Storage
Server1

Storage
Server2
There are 3 mirrored disks
and all of them are distributed Storage
in 3 different servers. Server3

Storage
Server4
 VastSky
• Redundancy
 VastSky mirrors user data to three storage servers by default and
all of them are updated synchronously.
 VastSky can be configured to use two networks (e.g. two
independent ethernet segments) for redundancy.
• Fault detection
 The storage manager periodically checks if each head and storage
servers are responsive.
• Recovery
 On a failure, the storage manager attempts to reconfigure mirrors
by allocating new extents from other disks automatically.
 VastSky
• Recovery Mechanism Storage Pool

Storage
Logical Volume Server1

Storage
data Server2

spare
Storage
crash
data Server3

Storage
data Server4
 VastSky
• Scalability
 Most of cluster file-systems and storage systems which have a
meta-data control node have a scalability problem.
 VastSky doesn't have this problem since once a logical volume is set
up, all I/O operations will be done only through Linux drivers
without any storage manager interactions.
 VastSky
• Load Balance
 With VastSky's approach, the loads will be equalized across the
physical disks, which leads that it utilizes the I/O bandwidth of
them. Logical Volume
Storage Pool

D1 D2 D3
D1 D3 D1 Storage
D2 D2 D2 D1 Server1
D3

D1 D1 D1 D3
D2 D2 D1 Storage
D3 D1 D3 Server2
D2 D3
D1
D2 D3 D3
D1 D1 D1 Storage
D3
D2 D2 D2 D3 Server3
D3
D1
D2 D2 D1 D2
D3 Storage
D3 D2 D3 D2 D1 Server4
D3
 Lustre File System
• What is Lustre ?
 Lustre is a POSIX-compliant global, distributed, parallel filesystem.
 Lustre is licensed under GPL.
• Some features :
 Parallel shared POSIX file system
 Scalable
• High performance
• Petabytes of storage
 Coherent
• Single namespace
• Strict concurrency control
 Heterogeneous networking
 High availability
 Lustre File System
• Lustre components :
 Metadata Server (MDS)
• The MDS server makes metadata stored in one or more MDTs.
 Metadata Target (MDT)
• The MDT stores metadata (such as filenames, permissions) on an MDS.
 Object Storage Servers (OSS)
• The OSS provides file I/O service, and network request handling for one or
more local OSTs.
 Object Storage Target (OST)
• The OST stores file data as data objects on one or more OSSs.

• Lustre network :
 Supports several network types
• Infiniband, TCP/IP on Ethernet, Myrinet, Quadrics, …etc.
 Take advantage of remote direct memory access (RDMA)
• Improve throughput and reduce CPU usage
Lustre File System
 Lustre File System
• Lustre in HPC
 Lustre is the leading HPC file system
• 15 of Top 30
• Demonstrated scalability
 Performance
• Systems with over 1,000 nodes
• 190 GB/sec IO
• 26,000 clients
 Examples
• Titan supercomputer at Oak Ridge National Laboratory
– TOP500: #1, November 2012
• System at Lawrence Livermore National Laboratory (LLNL)
• Texas Advanced Computing Center (TACC)
 Ceph
• Overview
 Ceph is a free software distributed file system.
 Ceph's main goals are to be POSIX-compatible, and completely
distributed without a single point of failure.
 The data is seamlessly replicated, making it fault tolerant.

• Release
 On July 3, 2012, the Ceph development team released Argonaut, the
first release of Ceph with long-term support.
 Ceph
• Introduction
 Ceph is a distributed file system that provides excellent
performance ,reliability and scalability.
 Objected-based Storage.
 Ceph separates data and metadata operations by eliminating file
allocation tables and replacing them with generating functions.
 Ceph utilizes a highly adaptive distributed metadata cluster,
improving scalability.
 Using object-based storage device (OSD) to directly access data,
high performance.
 Ceph
• Objected-Based Storage
 Ceph
• Goal
 Scalability
• Storage capacity, throughput, client performance. Emphasis on HPC.
 Reliability
• Failures are the norm rather than the exception, so the system must have
fault detection and recovery mechanism.
 Performance
• Dynamic workloads  Load balance.
 Ceph
• Three main components
 Clients : Near-POSIX file system interface.
 Cluster of OSDs : Store all data and metadata.
 Metadata server cluster : Manage namespace (file name).
 Three Fundamental Design
1. Separating Data and Metadata
 Separation of file metadata management from the storage.
 Metadata operations are collectively managed by a metadata server
cluster.
 User can direct access OSDs to get data by metadata.
 Ceph removed data allocation lists entirely.
 Usr CRUSH assigns objects to storage devices.
 Separating Data and Metadata
• Ceph separates data and metadata operations
 Separating Data and Metadata
• CRUSH(Controlled Replication Under Scalable Hashing)
 CRUSH is a scalable pseudo-random data distribution function
designed for distributed object-based storage systems .
 Define some simple hash functions.
 Use hash functions to efficiently map data objects to storage
devices without relying on a central directory.
 Advantages
• Because using hash functions, client can calculate object location directly.
 Separating Data and Metadata
• CRUSH(x)  (osdn1, osdn2, osdn3)
 Inputs
• x is the placement group
• Hierarchical cluster map
• Placement rules
 Outputs a list of OSDs
• Advantages
 Anyone can calculate object location
 Cluster map infrequently updated
 Separating Data and Metadata
• Data Distribution with CRUSH
 In order to avoid imbalance (OSD idle, empty) or load asymmetries
(hot data on new device).
→distributing new data randomly.
 Use a simple hash function, Ceph maps objects to Placement
groups (PGs). PGs are assigned to OSDs by CRUSH.
 Dynamic Distributed Metadata
Management
2. Dynamic Distributed Metadata Management
 Ceph utilizes a metadata cluster architecture based on Dynamic
Subtree Partitioning.(workload balance)
 Dynamic Subtree Partitioning
• Most FS ,use static subtree partitioning
→imbalance workloads.
→simple hash function can get directory.
• Ceph’s MDS cluster is based on a dynamic subtree partitioning. →balance
workloads
 Reliable Distributed Object Storage

3. Reliable Autonomic Distributed Object Storage

 Replica.
 Failure Detection and Recovery.
 Client
• Client Operation
 File I/O and Capabilities

Request Translate file

name into
Client MDS inode(inode
(open file) number, file
owner, mode,
Check OK, return size, …)
Direct
access
Return inode number,
map file data into
objects(CRUSH)
OSD
 Client
• Client Synchronization
 If multiple clients (readers and writers) use the same file, cancel
any previously read and write requests until OSD check OK.
• Traditional: Update serialization. → Bad performance
• Ceph: Use for HPC (high-performance computing community) can read
and write different parts of same file (diff objects).
→ increase performance
 Metadata
• Dynamically Distributed Metadata
 MDSs use journaling
• Repetitive metadata updates handled in memory.
• Optimizes on-disk layout for read access.
 Per-MDS has journal (usually a circular log in a dedicated area of
the file system), when MDS failure another node can quickly
recover with journal.
 Inodes are embedded directly within directories.
 Each directory’s content is written to the OSD cluster using the
same striping and distribution strategy as metadata journals and
file data.
 Replica
• Replication
 Data is replicated in terms of PGs.
 Clients send all writes to the first non-failed OSD in an object’s PG
(the primary), which assigns a new version number for the object
and PG; and forwards the write to any additional replica OSDs.
 Failure detection
• Failure detection
 When OSD not response → sign “down”
 Pass to the next OSD.
 If first OSD doesn’t recover →sign “out”
 Another OSD join.
 Recovery
• Recovery and Cluster Updates
 If OSD1 crashes → sign “down”
 The OSD2 takes over as primary.
 If OSD1 recovers → sign “up”
 The OSD2 receives update request, sent new version data to OSD1.
 HDFS
• Overview
 HDFS(Hadoop Distributed Cloud Applications
File System).
 Reference from Google File
System. MapReduce Hbase
 A scalable distributed file
system for large data analysis. Hadoop Distributed File System
 Based on commodity (HDFS)
hardware with high fault-
tolerant.
 The primary storage used by A Cluster of Machines
Hadoop applications.
 Hadoop
• Introduction
 An Apache project
 A distributed computing platform
 A software framework that lets one easily write and run
applications that process vast amounts of data
• From three papers
 SOSP 2003 : “The Google File System”
 OSDI 2004 : “MapReduce : Simplifed Data Processing on Large
Cluster”
 OSDI 2006 : “Bigtable: A Distributed Storage System for Structured
Data”
 Hadoop Features
• Efficiency
 Process in parallel on the nodes where the data is located
• Robustness
 Automatically maintain multiple copies of data and automatically re-
deploys computing tasks based on failures
• Cost Efficiency
 Distribute the data and processing across clusters of commodity
computers
• Scalability
 Reliably store and process massive data
HDFS Architecture
 HDFS Architecture
• NameNode
 The file content is split into blocks (default 128MB,3 replica).
 Files and directories are represented on the NameNode by inodes
(permissions, modification and access times, namespace and disk
space quotas).
 Namespace is a hierarchy of files and directories.
 NameNode maintains the namespace tree and the mapping of file
blocks to DataNodes.
 Three components
• Image: the inode data and the list of blocks (name system).
• Checkpoint: the persistent record of the image (file system).
• Journal: the modification log of the image (file system).
 HDFS Architecture
• Image and Journal
 When Startup NameNode
1. Load checkpoint.
2. Use journal.
• CheckpointNode
 The CheckpointNode periodically combines the existing checkpoint
and journal to create a new checkpoint and an empty journal.
 The CheckpointNode downloads checkpoint and journal from
NameNode and return a new checkpoint and an empty journal.
• BackupNode
 The BackupNode always follows journal to keep NameNode latest
version.
 If NameNode fails, use BackupNode untill NameNode recovers.
 HDFS Architecture
• DataNode
 Each block replica on a DataNode is represented by two files:
• Data
• Block’s metadata (checksum, generation stamp)
 When startup DataNode, NameNode performs handshaking:
• Verify the name space ID.
• Verify the soft version.
 A new DataNode will receive namespace ID .
 After the handshaking the DataNode registers will get Storage ID.
 A DataNode identifies block replicas in its possession to the
NameNode by sending a block report (block ID, generation stamp).
(1 time/hr)
 Hearbeats: 1 time/3 sec
HDFS Client
 HDFS Client
• File Write
 HDFS implements a single-writer, multiple-reader model.
 The writing client periodically renews the lease by heartbeat:
• Soft limit: client fails to renew the lease, another client can preempt.
• Hard limit: client fails to renew the lease, the client quit and close file.
 Form a pipeline.
 A full packet buffer is pushed
to the pipeline.
 HDFS Client
• File Read
 When a client opens a file to read, it fetches the list of blocks and
the locations of each block replica from the NameNode.
 Read from the nearest replica first. If fails, read from the next
nearest replica.
 REPLICA MANGEMENT
• Block Placement
 When a DataNode registers to the NameNode, the NameNode runs
a configured script to decide which rack the node belongs to.
 The default HDFS block placement policy provides a tradeoff
between minimizing the write cost, and maximizing data reliability.
 The default HDFS replica placement policy:
• No DataNode contains more than one replica of any block.
• No rack contains more than two replicas of the same block.