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

John Chuang 1999 2

Outline

Introduction: what is stor-serv? Resource allocation: model & example

John Chuang 1999 3

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 4

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 5

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 6

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 7

QoS Dimensions: Placement & Replacement

[ [

Replication

[

Push/pre-fetch

[

Differential Caching Simple Caching Custom Replacement Custom Placement Service Class

John Chuang 1999 8

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 9

The stor-serv Framew ork

Service Specification Realized Performance: latency, availability, etc. Resource Mapping Admission Control Resource Management

Resource Reservation

Service Provision Metadata Management Network Topology & Resource Availability Network Conditions & Traffic Patterns

Publisher

Resource Discovery

Clients Report back to Publisher

Performance Requirements Traffic Profile

Security Pricing and Payment

John Chuang 1999 10

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 11

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 12

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 13

Netw ork Model

network G(V,E) demand points V= { v1,v2,...,vi ,...} supply points S= { s1,s2,...,sj ,...} ; S

V

storage cost cS(j) incremental transmission cost cT(i,j)

John Chuang 1999 14

Traffic Profile

• bject collection Q= { q1,q2,...,qk ,...}
• bject size = b(k)

collection size, Bcorpus =

b (k)

qk Q

reservation start time Ts and duration Td data request rate λ data request distribution g(i,k)

• conditional probability that object qk is requested by some user

at node vi given that there is an object request

Σ

John Chuang 1999 15

Performance Requirements

worst-case delay: Dmax < τmax average delay: Davg < τavg stochastic guarantee: P[D > τthreshold] < ε

John Chuang 1999 16

Resource Mapping

Find optimal replication set Xh:

min where B(x) = b(k)

qk Qx

s.t. performance requirement(s)

B x

( )⋅ cS x ( )

x∈Xh

∑

Σ

(Storage requirement at node x) (Total Storage Cost)

John Chuang 1999 17

Admission Control

For each node x Xh , Ts < t < (Ts+ Td) test: B(x) + B0(x,t) < TSC (x,t)

where B(x) is requested storage capacity B0(x,t) is committed storage capacity at time t TSC (x,t) is total storage capacity

John Chuang 1999 18

Resource Mapping Example

14 22 32 25 8 41 24 26 27 3 17 16 42 1 9 11 15 31 43 2 29 23 6 39 36 47 38 37 40 19 18 20 46 33 12 21 5 44 45 35 34 10 13 7 4 28 30

ARPANET Number of nodes = 47 Number of links = 68 Average node degree = 2.89 Network diameter (hops) = 9

John Chuang 1999 19

14 22 32 25 8 41 24 26 27 3 17 16 42 1 9 11 15 31 43 2 29 23 6 39 36 47 38 37 40 19 18 20 46 33 12 21 5 44 45 35 34 10 13 7 4 28 30

Max Delay Bound = 4 hops

replica 1 hop 2 hops 3 hops 4 hops 2 replicas: average delay = 2.34 hops; maximum delay = 4 hops

John Chuang 1999 20

14 22 32 25 8 41 24 26 27 3 17 16 42 1 9 11 15 31 43 2 29 23 6 39 36 47 38 37 40 19 18 20 46 33 12 21 5 44 45 35 34 10 13 7 4 28 30

• Avg. Delay Bound = 2 hops

replica 1 hop 2 hops 3 hops 4 hops Average Distance = 1.87 hops; Maximum Distance = 4 hops

John Chuang 1999 21

Resource Mapping for Services w ith Max and Avg Delay Bounds

Resource Mapping for ARPANET

1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 Delay bound (hops) maximum delay bound average delay bound

Number of Replicas

John Chuang 1999 22

Non-Uniform Demand Distribution Improves Mapping Efficiency

Mapping for non-uniform demand distribution

1 2 3 4 5 6 7 8 1 2 3 4 Average delay bound (hops) uniform non-uniform

Number of Replicas

John Chuang 1999 23

Mapping Efficiency through Partial Replication

Full vs. Partial Replication for 4-Object-Collection

4 8 12 16 20 1 2 3 4 Average Delay Bound (hops) partial replication full replication

Number of Replicas

John Chuang 1999 24

Mapping into Storage & Transmission Resources

Minimize total cost subject to meeting

performance requirements

storage cost transmission cost

min

X h⊆S VL⊆V

B x

( )⋅c S x ( )+

λ⋅ g i, k

( )⋅ b k ( )⋅cT(i,Xk)

qk∈Q

∑

i∈VL

∑

x∈Xh

∑

      ⋅ dt

t= Ts Ts+ Td

∫

     

John Chuang 1999 25

Optimal Solution Depends on

Relative cost of storage v. transmission Object access frequency

10-5 10-4 10-3 10-2 10-1 100 101 102 10-3 10-2 10-1 100 101 102 103

1 replica Data Rate ( λ) 2 replicas 3 replicas 4 replicas

Storage to Transmission Cost Ratio (CS/CT)

John Chuang 1999 26

Optimal Combination of Storage and Transmission

1000 750 500 250 100 80 60 40 20 0% 20% 40% 60% 80% 100% Relative Cost Data rate ( λ λ) Storage to Transmission Cost Ratio

4 replicas 3 replicas 2 replicas 1 replica

Total Cost Relative to Storage-Only Solution

John Chuang 1999 27

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