Challenges in modeling enterprise storage systems
Arif Merchant Hewlett-Packard Laboratories Contact:
[email protected] © 2006 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice
An enterprise storage system Multiple applications
App1 data1
App2
data2
data3
Primary site Running on resources at primary site
Host
Application data data5 objects
App3 data4
Data storage system
Host Storage-area network
Tape lib
Each app has its own data objects
Potentially using secondary site resources
Secondary site(s)
Disk Disk array array
Sites may serve as primary / secondary peers WAN link(s)
Tape transport QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
2
Storage System models Why should you care? •
Storage is a significant part of IT costs − Over 35% for enterprises with 5000+ employees (IDC 2002) − Large fraction of storage cost is management − Environment is complex: 100s of applications sharing petabytes of data − Stringent application requirements – failures can be catastrophic (Survey: $89K-6.4M/hr of downtime)
•
Good storage models are crucial for managing storage − Systematic, accurate models needed for informed choices
•
Uses of models − Capacity planning – green field & consolidation − Storage design − On-going management
QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
3
Storage management automation
Where storage models fit in [Anderson2002,Alvarez2001] Understand needs
Determine solution
•Workload characterization
•Select devices + configurations
•System components
•Model to verify config. Is good (Changing) business
requirements
•Assign load to devices
Design Design
Iterative feedback loop
Analyze Analyze
Implement Implement
Monitor usage Construct solution
•Measure workload, re-characterize
•Configure devices, LVM, …
•System response
•Migrate data
QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
4
Components of storage models App1 data1
App2
data2
data3
data4
Primary site
Workloads
Host
Application data data5 objects
App3
Data storage system
Host
Secondary site(s)
Storage-area network
Tape lib
Disk Disk array array WAN link(s)
Tape transport QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
5
Workload characterization • Need
− Compact characterization − Adequate for predictive storage system models • Minimal
parameter set:
− Size of data − IO rate, IO size distribution − Reads vs. writes − Sequentiality, spatial locality − Temporal locality (e.g., frequency of repeated access) − Concurrency (Simultaneously outstanding requests) QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
6
Measuring workloads •
How do we acquire workload parameters? − Users do not know workload details − Measurements on production system may add load − Storage devices have limited, vendor-dependent measurement points (although industry standards SMI-S are helpful!) − Workload behavior may depend on hardware & configuration − Open problems:
• Minimum perturbation methods for measuring workload parameters • Tradeoffs between inaccuracies of workload parameters and model accuracy IO trace
running system QEST, September 2006
workload analysis
workload description
workload analysis
Arif Merchant: Challenges in modeling enterprise storage systems
7
Workload attributes – access variability
How do you compactly capture this complex behavior? Sample results for a single run i3125om5, RC = 8GB, WC = 8GB; cache outcome percentages by I/O address range 100%
Coalesced write
average over full trace
90%
80%
Warm write
average for writes second half
70%
Cold write
60%
Delayed write
I/O's
cache hits for a traced openmail workload
50%
40% reads
Compulsory read miss
30%
Replacement read miss 20% read hits
10%
0% 0
400
800
1200
1600
2000
2400
2800
3200
trace time [seconds] QEST, September 2006
3600 ave steady state
Arif Merchant: Challenges in modeling enterprise storage systems
Read hit - reread Read hit from write cache
Graph Graph by by the the HPL HPL Sonora Sonora team team 8
Workload specification challenges • What
is an adequate workload specification? −Complex request arrival patterns
• Self-similar? Dependent on response? • Seasonality, trends
−Complex access patterns −Locality (spatial and temporal) −Concurrency variation −Correlation/interference between workloads • Synthesis
of workloads accurately representing real workloads
QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
9
Components of storage systems The basic storage device
App1 data1
App2
data2
data3
data4
Primary site Host
Disk drives
Application data data5 objects
App3
Data storage system
Host
Secondary site(s)
Storage-area network
Tape lib
Disk Disk array array WAN link(s)
Tape transport QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
10
Disk drives simplified
2.5” Seagate Savio
1 0 1 1 1 0 1 0 0 1 1 0 0 0 0 1
Block QEST, September 2006
Arif Merchant: Challenges in modeling enterprise storage systems
11
Disk drives specs
High end (enterprise) disk Seek
Rotation Transfer
3.5ms
• Typical
Seagate Cheetah 15K drive
2ms
1ms
64 KB random read service time
access times (seek + rotation):
− 5.5ms − Not improving quickly (limited by speed of physical movement) • Typical
sustained transfer rate:
− 58 to 96 MB/s or ~1ms for 64KB − Increasing quickly over time (increases with bit density) • Annual QEST, September 2006
failure rate: 0.62% Arif Merchant: Challenges in modeling enterprise storage systems
12
High disk access time
The root of many storage (modeling) problems • Random
IO throughput
