PULSTORE : Automated storage management with QoS Guarantee in LargeScale Virtualized Storage Systems

1
PULSTORE Automated storage management with QoS
Guarantee in Large-Scale Virtualized Storage
Systems

• Review by
• Ajay

2
Key Points

• Autonomic management system of SANs (storage area
  networks) based on QoS and Cost metrics
• Growing need for SANs and storage virtualization
• Need for on-demand utility provisioning to suit
  to dynamically varying workload
• Migration model most cost efficiency integrated
  with a number of migration strategies that can
  provide the requested QOS

3
Storage Area Networks

• Storage shared over heterogeneous networks
• Storage accessed over a network
• Makes storage independently managed and
  administered
• Storage Controller
• Storage virtualization
• Static vs Dynamic

4
Typical IO Workload

• Varying from time to time
• Huge magnitude difference
• On-demand Utility Provisioning
• Optimize mapping of physical to virtual memory to
  satisfy IO Response time requirements
• Moving data in storage hierarchy under the
  following constraints
• Capacity/performance
• Cost

5
Model

• Hierarchical storage structure
• Higher the storage, higher the cost
• Storage QoS goals (QoSS)
• IO Latency used as metric for performance
  prediction
• Function of workload and disk config i.e. fn of
  time
• L(t) lt LQoS
• Resource Utility Cost
• Minimize

6
Design - PULSTORE

• Goal Provide migration sequence to minimize
  cost and satifsy QoSS specs
• Single Migration
• Provides a safe time zone
• Based on workload fraction remaining on source
  destination
• Optimal Migration
• Sequence of migrations
• Assumption All future workload completely known
• Approximate Migration
• Assumption Workload with fixed size prediction
  window
• Down migration early
• Up migration late
• Tunes accordingly as prediction window moves

7
Discussion

• Notable Features
• Cost utility model
• Model emphasizes time-based workload prediction
• Different schemes for data migration based on
  system knowledge
• Discussion
• Info on Implementation Results. And comparison
  with existing traditional migration schemes
• Need to incorporate users knowledge and
  prediction model . Eg seasonal prediction
  (increased sales during Christmas/New year)