Resource Allocation for stor - serv : Netw ork Storage Service w ith - PowerPoint PPT Presentation

Resource Allocation for stor - serv : Netw ork Storage Service w ith QoS Guarantees John Chuang chuang@ sims.berkeley.edu NetStore’99 October 14 1999

Outline � Introduction: what is stor - serv ? � Resource allocation: model & example John Chuang 1999 2

Introduction � Benefits of distributed network storage - Bandwidth savings - Latency reductions - Data availability/redundancy - Load balancing � Different approaches - Caching: network-centric (max hit rate) - Replication: publisher-centric (max publisher value) - Others: push, pre-fetch, differential caching, etc. John Chuang 1999 3

Caching � Pros: - Simple, adaptive (traffic-driven; best effort) - Object level granularity - Transparent to both publisher and consumer - Static cache hierarchy makes resource discovery easy � Cons: - Publisher has no control over placement or replacement (traffic-driven; best effort) - Cache misses possible - Publisher cannot collect access statistics - Static cache hierarchy difficult to change John Chuang 1999 4

Replication � Pros: - Publisher controls object placement & replacement - Advanced reservation and placement - Guaranteed data availability (no cache miss) - Easier to arrange for collection of access statistics � Cons: - High setup cost (entire sites, not individual objects) - Not as adaptive to changes - Not necessarily transparent to consumers (resource discovery non-trivial outside of cache hierarchy) John Chuang 1999 5

What is stor-serv ? � Unified network storage service framework � Inspired by intserv/diffserv � Publisher can choose QoS level - Best-effort caching - Differential caching - Push/pre-fetch - Guaranteed service object replication John Chuang 1999 6

QoS Dimensions: Placement & Replacement Service Class Custom Custom Placement Replacement Simple Caching Differential Caching [ Push/pre-fetch [ Replication [ [ John Chuang 1999 7

stor-serv Efficiencies � Standardized service semantics + automated resource allocation ⇒ low setup cost ⇒ adaptability � Statistical multiplexing - Resource reserved for object replication, etc. - Left-over capacity for best-effort caching John Chuang 1999 8

The Publisher stor - serv Performance Traffic Profile Requirements Framew ork Service Specification Resource Reservation Service Provision Network Topology & Resource Resource Mapping Availability Report back to Publisher Network Admission Conditions Control Metadata & Traffic Management Patterns Security Resource Realized Management Pricing and Performance: Payment latency, availability, Resource etc. Clients Discovery John Chuang 1999 9

Service Specification � Traffic Profile - storage capacity - time and duration (advance reservation) - data access pattern (if known) � Performance Requirements - delay, distance, jitter, availability, etc. - deterministic vs. statistical guarantee John Chuang 1999 10

Example Replication Services Deterministic 100kB storage capacity for 1 hour; maximum distance = 4 hops Average 1GB storage capacity for 1 day; 500ms average network latency Stochastic 1MB storage capacity for 1 hour; Probability[hops > 4] < ε Advance 1GB storage capacity for 1 hour; Reservation starting at 11:59pm, Dec 31, 1999; average distance = 2 hops John Chuang 1999 11

Resource Allocation � Resource Mapping - service specification � physical resource requirements - map into storage & transmission resources - facilities-location problem � Admission Control - accept/reject service requests based on utilization level John Chuang 1999 12

Netw ork Model � network G ( V , E ) � demand points V = { v 1 , v 2 ,..., v i ,...} � supply points S = { s 1 , s 2 ,..., s j ,...} ; S V � storage cost c S ( j ) � incremental transmission cost c T ( i , j ) John Chuang 1999 13

Traffic Profile � object collection Q = { q 1 , q 2 ,..., q k ,...} � object size = b ( k ) Σ � collection size, B corpus = b ( k ) q k Q � reservation start time T s and duration T d � data request rate λ � data request distribution g ( i , k ) - conditional probability that object q k is requested by some user at node v i given that there is an object request John Chuang 1999 14

Performance Requirements � worst-case delay: D max < τ max � average delay: D avg < τ avg � stochastic guarantee: P[ D > τ threshold ] < ε John Chuang 1999 15

Resource Mapping � Find optimal replication set X h : ∑ ( ) ⋅ c S x ( ) B x min (Total Storage Cost) x ∈ X h Σ where B ( x ) = b ( k ) (Storage requirement at node x) q k Q x s.t. performance requirement(s) John Chuang 1999 16

Admission Control For each node x X h , T s < t < ( T s + T d ) test: B ( x ) + B 0 ( x , t ) < TSC ( x , t ) where B ( x ) is requested storage capacity B 0 ( x , t ) is committed storage capacity at time t TSC ( x , t ) is total storage capacity John Chuang 1999 17

Resource Mapping Example 23 14 2 43 31 9 20 27 29 15 6 39 40 19 18 41 10 13 30 47 38 37 7 36 8 32 4 34 11 25 3 28 17 ARPANET 24 26 35 Number of nodes = 47 16 Number of links = 68 12 21 5 Average node degree = 2.89 22 1 44 42 46 33 45 Network diameter (hops) = 9 John Chuang 1999 18

Max Delay Bound = 4 hops 23 14 2 43 31 replica 9 20 27 1 hop 2 hops 29 15 6 39 40 19 18 41 3 hops 10 13 30 4 hops 47 38 37 7 36 8 32 4 34 11 25 3 28 17 24 26 35 16 12 21 5 22 1 44 42 46 33 45 2 replicas: average delay = 2.34 hops; maximum delay = 4 hops John Chuang 1999 19

Avg. Delay Bound = 2 hops 23 14 2 43 31 replica 9 20 27 1 hop 2 hops 29 15 6 39 40 19 18 41 3 hops 10 13 30 4 hops 47 38 37 7 36 8 32 4 34 11 25 3 28 17 24 26 35 16 12 21 5 22 1 44 42 46 33 45 Average Distance = 1.87 hops; Maximum Distance = 4 hops John Chuang 1999 20

Resource Mapping for Services w ith Max and Avg Delay Bounds Resource Mapping for ARPANET 8 maximum delay bound 7 average delay bound 6 Number of Replicas 5 4 3 2 1 0 0 1 2 3 4 5 6 7 Delay bound (hops) John Chuang 1999 21

Non-Uniform Demand Distribution Improves Mapping Efficiency Mapping for non-uniform demand distribution 8 7 uniform non-uniform Number of Replicas 6 5 4 3 2 1 0 0 1 2 3 4 Average delay bound (hops) John Chuang 1999 22

Mapping Efficiency through Partial Replication Full vs. Partial Replication for 4-Object-Collection 20 partial replication full replication 16 Number of Replicas 12 8 4 0 1 2 3 4 Average Delay Bound (hops) John Chuang 1999 23

Mapping into Storage & Transmission Resources � Minimize total cost subject to meeting performance requirements   Ts + Td   ∑ ∑ ∑ ∫ ( ) ⋅ c S x ( ) + ( ) ⋅ b k ( ) ⋅ c T ( i , X k ) λ ⋅ g i , k ⋅ dt   min  B x  X h ⊆ S     x ∈ X h i ∈ V L q k ∈ Q t = Ts V L ⊆ V storage cost transmission cost John Chuang 1999 24

Optimal Solution Depends on � Relative cost of storage v. transmission � Object access frequency Storage to Transmission Cost Ratio (C S /C T ) 10 2 10 1 1 replica 10 0 2 replicas 3 replicas 10 -1 4 replicas 10 -2 10 -3 10 -4 10 -5 10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3 John Chuang 1999 25 Data Rate ( λ )

Optimal Combination of Storage and Transmission Total Cost Relative to Storage-Only Solution 4 replicas 3 replicas 100% 80% Relative Cost 60% 40% 0 20% 2 replicas 20 0% 40 1 replica 1000 60 750 80 Storage to 500 100 Transmission 250 λ λ ) Cost Ratio Data rate ( 0 John Chuang 1999 26

Conclusion � There is no one-size-fits-all storage solution - stor-serv provides unified QoS framework (from caching to replication) � Resource allocation should be able to support - Object level granularity - Short service durations - Placement (spatial) and replacement (temporal) control - Deterministic and stochastic guarantees - Mapping into storage and transmission resources John Chuang 1999 27