Storage Systems: Management of Persistent Data

1
Storage Systems Management of Persistent Data

• Randal Burns

2
What distinguishes storage problems?

• What makes something a storage system?
• Any software that manages persistent data.
• Why is storage an interesting area?
• Storage holds the crown jewels of every
  organization Information
• Storage is the performance bottleneck in almost
  all systems
• Cause of the longest lasting and most expensive
  failures

3
The Storage Systems Chainsaw

• Horizontal cut across
• Database
• Operating systems
• Web serving
• Multimedia serving
• Mobile and wireless devices
• Discipline is broad
• Studies same class of problems in entirely
  different systems

4
The Storage Systems Auger

• Vertical cut through
• Applications
• Middleware
• Operating systems
• Logical volume
• Device
• Discipline is deep
• Persistence needs to be considered at all levels

5
What are we going to do?

• Look at the evolution of storage system
  technology
• Disk
• Tape
• Future I/O devices
• Sample a couple of research areas
• Volume managers
• File system recovery

6
The History of Innovation Devices

• Punch cards 1884
• Used to store information for calculators
• Card readers allowed data to be reloaded
• In computer systems 1950s
• Load programs and program data
• Computers had no memory or storage

7
The History of Innovation Devices

• Tape (1952)
• Used to store information for batch applications
• Use same data set during multiple runs
• Routine library shared by all applications
• Reduce load time
• Made programs smaller because not all information
  needs to be on punch cards
• What happened to tape?
• Became used as part of a hierarchical on-line
  store
• Used for backup/archive, off-line data only
• Has been declared dead several times

8
The History of Innovation Devices

• IBM 7494 at Houston

Modern Tape Robots
9
The History of Innovation Devices

• Disk (1956)
• Introduced as a random-access device
• Disks were small, hold runtime data, tapes were
  used for persistent storage
• Disks were a memory
• The evolution of disk drives
• No longer a random-access device
• Evolved into a store, with RAM as a memory
• Starting to be used as archival storage
• Recently has been proposed to be dieing

10
The History of Innovation Devices
Microdrive
RAMAC
11
The History of Innovation Devices

• Non-volatile memories (1995)
• Used as small, stable memories for crash recovery
  and synchronous I/O, in front of disks
• Same relationship between disk and tape in 1958
• Evolution of NV memories
• Stable storage for portable devices
• Cell-phones, PDAs, and handhelds

12
Conclusions Devices

• Some historical observations
• Device technologies move down the memory storage
  hierarchy as capacity increases
• Devices are replaced by faster technology
• Not all emerging devices fit the old model

13
Emerging Technology Devices

• MEMS storage
• Micro-electromechanical systems
• Not rotational sled translates in XY direction
• Performs much like a disk, but with different
  cost/density profile

14
Emerging Technology Devices

• MRAM magnetic RAM
• Stable storage without power
• Access speed, cost, and density of DRAM (memory
  used in computers today)
• Will not replace disk right away, because costs
  per unit will be high
• First stable memory since disk drives

15
Volume Managers

• External large-scale block storage
• Highly reliable
• Central repository of information is easy to
  manage
• Backup and restore
• User quotas
• Configurable
• Extend and move volumes
• RAID redundant array of inexpensive disks
• Add an extra disk to a system of disks so that
  data are preserved when a disk fails

16
Volume Managers

• EMC Symmetrix
• IBM Enterprise
• Storage Server

17
RAID Adding a Parity Disk

• RAID is the most important research idea
  underlying volume managers
• A parity disk allows a storage systems to survive
  any single failure
• Data can be read/written after a failure

18
RAID How does it go?

• Observe that if then
• Example
• P stands for parity

19
RAID Handling a Single Failure

• Block A can be rebuilt from B and P
• Block B can be rebuilt from A and P

20
RAID Adding a Parity Disk

• Data becomes highly available
• Disk MTTF 5 years
• 5 disks, MTTF 1 year
• RAID 5 P
• Assume MTTR of 1 day, add a new disk and rebuild
  the contents
• MTTF 11,000 years
• Chance of another disk failing during the day it
  takes to recover from the first failure

21
RAID Benefits

• Parallelism can write to n disks at the same
  time
• Good when writing a whole stripe of data

22
RAID Drawbacks

• Small write problem
• To update data smaller than a whole stripe, must
  read data from the disks before writing

23
RAID Drawbacks

• Small write problem
• Parity must change when writing just one block

24
RAID Drawbacks

• Small write problem

25
RAID Tradeoffs

• Small write problem plagues RAID systems
• Advanced techniques attempt to mitigate it
• RAID performs well in high-performance
  environments with large data
• Scientific computing
• Supercomputing
• Availability benefits overshadow any performance
  drawbacks

26
File System Recovery

• When your (or any) computer crashes, the
  operating system must check the file system
• Problem is called restart recovery
• Some file system data in memory was lost when
  system crashed
• File system integrity is in question
• Lost blocks of storage
• Pointers to nowhere
• Bad data in files
• Unreachable files, unnavigable namespace

27
File System Checking

• Traditional approach to file system recovery
• Check all of the file system data structures,
  make sure that
• Every file is in a directory
• Every directory can be reached from the root
• All disk blocks are either in files or empty
• For small disks and file systems, no problem
• Windows and UNIX
• For large-scale storage systems, file system
  checking can take hours

28
A Simple Operation
rename /home/randal/files /home/bob

• Before
• After

29
Updating the Namespace

• Changes do not occur all at the same time
• System crash can occur at any point

30
Need for File System Checking

• Several inconsistent states
• All are recoverable by searching file system

31
Transactional Storage

• For each operation that modifies the file system
• Record actions in a log describing how the file
  system is going to be changed
• Change the file system
• The log contains all information needed to fix
  the file system
• If a failure occurs during the log-write, the
  action never happened
• If a failure occurs while changing the FS, the
  action can be redone based on the log

32
Updating the Namespace

• After a system crash read the log
• Specifies actions that may be incomplete

LOG action action mv /home/randal/files
/home/bob action
33
Performance of Transactional Storage

• Logging can reduce latency and improve
  performance for writes to disk
• Writing changes to log alone is sufficient
• File system can be changed in the background
• Log writes can be done sequentially
• Does not reduce overall I/O, but can make single
  operations more responsive
• File system recovery takes seconds rather than
  hours

34
Summary

• Motivated storage systems as a broad and deep
  research area
• Overview of device technology that drives the
  field
• Disk, tape, NV RAM, and futuristic devices
• Two exemplary research problems
• RAID volume managers
• File system recovery