Post on 27-Dec-2015
transcript
Data Storage
CPTE 433John Beckett
The Paradox
• “If I can go to a computer store and buy 1000 gigabytes for $50, why does it cost more in your server farm?”
• It isn’t about having storage• It’s about managing data through its
life-cycle• The new measurement is price per
gigabyte-month
Definitions
• Spindle, platters, heads– Physical arrangement of disk– Of little interest to us, except to help us
understand how new technologies will impact us
• Drive controller– On the hard drive itself– Connected to…
• Host Bus Adapter
RAID
Raid Level
Methods Characteristics
0 Stripe data acrossmultiple drives
Faster reads and writes; poor reliability
1 Mirrors copy of dataacross two drives
Faster reads; good reliability; failures tend to be catastrophic (JB & SA)
5 Distributed parityAny single disk may fail without loss
Faster reads; slower writes; more economical
10 Mirrored stripesRaid 0 group mirrored onto another group
Faster reads; best reliability; most expensive
Table 25.1
Dilemma: Can you add hardware without subtracting from reliability? (Only by using very high-quality hardware)
Where Is the Data?
• DAS – Directly Attached Storage (IBM: DASD), connected directly to the server– May be a RAID array
• NAS – Network-Attached storage– Uses a protocol to transfer data
• SAN – Storage-Area Network– Separate network segment for storage,
connecting servers and drives
A SAN is usually made out of NAS devices
Structure of a SAN
LAN
SANCtrlr
(Server)
SAN backbone
NASNAS
NASNAS
SANCtrlr
(Server)
Managing Storage
• Think of storage as a community resource– If it’s personal, does it have any
business on company equipment?• Determine storage needs of the
group• Identify an architecture that will
satisfy that need• Plan an upgrade path for growth in
the future• Implement inventory and spares
policy
Standardization
• Disk drives are as important to standardize as any other component– Spares issue– Warrantee service procedure– Ability to use obsoleted drives
• Drive lifetime issue:– A drive motor may become unreliable
after so many revolutions
The Storage SLA
• Availability• Response time
• Reliability is increased by RAID > 0– …only if monitored and maintained– …only if RAID method is preserved
• Network is a part of the reliability picture
Backup and RAID
• RAID is not a backup strategy• If >n drives fail, you lose data• Controller failure can cause data loss
• One possibility: RAID mirror as a backup– Requires disconnecting other drive on failure
• How about: Spare drive, auto backup each night– Maybe including incremental backups
Using RAID mirror to effectively speed-up backup
• Break the RAID pair• Back up• Re-connect the RAID pair
Monitoring
• How full– Rate of change
• Broken drives• How busy (especially network on
NAS)• Unused
SAN Caveats
• Benchmarks are problematical• Useful versus physical storage size• Product life-cycle issues
Pipeline Optimization
• Read – buffered and available immediately
• Write – buffered and done at leisure– Dangerous if drive fails before update is
posted
Sync
• Early versions of an OS usually don’t sync properly if shut down during “quiet” time– Novell – unscheduled shutdown could be
catastrophic– Windows – learned some lessons from
others
• Is it safe to turn off power during operation?– A mainframe will be able to handle this
Performance
• Locate simultaneously-used data on different spindles to minimize head thrashing– The more complex your data, the harder
this is to do– Restrict this technique to very heavily-
used data• Beware of compression
– Assumes your data is organized a certain way
– Assumes your CPU has spare time to spend
Disk Access Density
• I/O Operations per second per gigabyte of capacity
• How fast can you move the entire drive of data?
Fragmentation
• Don’t fill up your drives! • That makes defragging slow• Also slows online attempts at limiting
fragments
Continuous Data Protection
• Send a log of all changes somewhere other than your disk drive– Tape– Over the network to another location– Another disk drive
• Back-out and forward recovery