View
216
Download
0
Tags:
Embed Size (px)
Citation preview
Dynamic Allocation in Honey Bee and Internet
Server Colonies
Sunil Nakrani, Computing Lab., University of Oxford, England, UK
Craig Tovey, ISyE, Georgia Institute of Technology, Atlanta, USA
Natural Systems Research & Education
Honey bee colony foraging (Bartholdi, Seeley, Tovey & VandeVate, J. Th. Bio. 1993); food storing to cue nectar intake (Seeley & Tovey, Animal Beh. 1993)
Dominance hierarchy formation (Chase, Tovey, et al., Proc. Nat. Acad. Sci 2002, Behaviour 2003); natural selection mechanism
Biomimetic heuristic for allocating resources in a web-hosting facility (Nakrani & Tovey, Proc. MASI II, 2003)
Time lags and overdiscounting of environmental costs, hedging value of environmental investments; replacement policies under technological change (Regnier, Sharp & Tovey, IE Trans.)
Assessing systems (Tovey, Ausenda); adjusting GDP for natural systems deterioration
Sustainability intro in sophomore course (2030); topics course on root causes of env. problems and sustainability (4833); stat and design sustainability projects
OR
->
BIO
BIO
->
OR
O
R -
> E
NV
Introduction
Web-Hosting Facility Rationale Benefits
Server Allocation Problem: allocate servers amongst web-apps to maximize revenue
Honey Bee Colony: allocate foragers amongst flower patches to maximize nectar intake
Introduction Approach: Honey Bee Heuristics-waggle
dance Map web-apps to flower patches, servers
to bees Solution Mapping: dance floor--> advert
board Algorithms and Simulation Model Results Conclusions, biological insight Future work
Web-Hosting Model
Benefits: Economy of scale: Resource sharing
means increase in utilization and better availability
Web-App shielded from over-provisioning
Web-Hosting Optimisation
Web-App: pay-per-use Service Level Agreement (SLA)
Hosting Center: Allocate servers among Web-Apps to “maximize” revenue (s.t. changeover downtime)
Users: Unpredictable and highly variable request pattern
Web-Hosting Optimisation Server
Allocation Problem: Allocate servers among web-Apps to “maximize” revenue
Server Allocation Problem Current Techniques: Threshold and Ad-
hoc Rule based, Continuous tracking of load metrics by large operations staff, Manual management Static provisioning altered approx. once a
month Current Literature– Jayram et. al. (2001),
Chase et. al. (2001) Commercial Domain: Proprietary methods
Honey Bee Colony
Typically requires 60 lb of honey per year to survive
25% of workers engaged in food collection (nectar, pollen)
Exploit food sources (flower patches) from surrounding countryside
Honey Bee Colony
Flower Patches: Availability varies daily and seasonally; Quality depends on exploitation, flower
type, micro-climate etc.. Round trip time (nectar collection time)
Colony: Exploit flower patches efficiently to satisfy nectar requirement
Forager Allocation Problem Forager
Allocation Problem: Allocate forager bees among flower patches to “Maximize” nectar intake
Problem Mapping Server
Allocation Problem: Single Server Web-Apps + User Group of servers
(cluster) serving users at one web-app
Forager Allocation Problem: Forager Bee Flower Patches Group of foragers
collecting nectar at a specific flower patch
Problem Mapping Server
Allocation Problem: Request service
time depends on Web-App
Find a user to serve
Forager Allocation Problem: Travel Time
depends on Flower Patch
Nectar collection time at the patch
Problem Mapping Server
Allocation Problem: Value-Per-
Request-Served Varying rates of
user request arrivals and balking behaviors
Forager Allocation Problem: Nectar quality
(sugar content) Varying flower
patch density, quality, and replenishment rate
Problem Mapping Server
Allocation Problem: Server Migration
Time (purge current Web-App and load new Web-App)
Forager Allocation Problem: Time to learn the
location of the flower patch and successful discovery (Seeley, T.D.)
Forager Allocation Mechanism Active foragers
return to the hive with nectar and profitability rating of the visited flower patch
Interact with food-storer bees to offload nectar (waiting time provides feedback on nectar flow into the hive)
Forager Allocation Mechanism: Feedback sets
threshold for enlisting signal (Waggle Dance)
Profitability + signal threshold = Waggle dance duration
Forager Allocation Mechanism: Waggle dance
performed just inside the hive entrance (Dance floor)
foragers follow dance to learn flower patch location
Suboptimal allocation in static sense
fi(xi) ´ return from xi bees at patch i Max i fi(xi)
s.t. xi ¸ 0i xi · N
OPTIMUM fi
0(xi) = 8 i2 A xi = 0 8 i A equalize marginal
return at active patches
BEE HEURISTIC fi(xi)/xi = 8 i2 A xi = 0 8 i A equalize average
return at active patches
Properties of Heuristic Solution(from BSTV 93)
Usually not optimalFactor-2 approximation even under
very weak conditionsConvergence proved by potential
function argumentValidated experimentally in a honey
bee colony
Solution Mapping Server
Allocation Advert Advert Board Advert Duration Reading an Advert
Forager Allocation Waggle Dance Dance Floor Dance Duration Following Waggle
Dance
Simulation Model: Honey Bee
Advert Board
Post/Read Adverts
Post/Read Adverts
MigrateRepurpose
Web-App: B
Web-App: A
Web-App IDDuration Time
Web-App IDDuration Time
Users: A
Users: B
Simulation Model: Greedy
Users: A
Users: B
MigrateRepurpose
Web-App: B
Web-App: A
Compute optimal policyfor next interval based on present queue status, present allocation, and user arrival from lastinterval
New Policy
New Policy
Simulation Model: Greedy
St = state of world at start period t (customers,servers)
At = arrivals (times, types) in period tP(, S, A) = profit using from state S
with arrivals Af(,S,A) = next state of world using
from S with arrivals At
G = arg max P(, St, At-1)
St+1 = f(tG, St, At)
Simulation Model: Others
Users: A
Users: B
MigrateRepurpose
Web-App: B
Web-App: A
Offline OmniscientComputation
New Policy
New Policy
Simulation Model: Omniscient Optimum
S=state, A=arrival, P( )=profit, f( )=next state
A1,, An known
vn+1(Sn+1) = 0 (no salvage value)
vt (St) = max{P(,St,At) + vt+1(f(,St,At))}
tOpt(St) = arg max {P(,St,At) + vt+1(f(,St,At))}
Omniscient Optimum Computation
Parallel implementation runs in 24 hours
Discretized space of possible states Inner loop function that we maximize
is theoretically concave … … but not concave numerically
Simulation Model: Optimal-Static
S=state, A=arrival, P( )=profit, f( )=next state
A1,, An known
s.t. St+1 = f(, St, At)
Result: Synthetic User Load
02468
1012
Reve
nue
(100K)
2Web-App
3Web-App
Facility configuration
Relative PerformanceOmni.HoneyGreedyOpt-Stat
Result: Internet Service Trace Load
02468
101214
Revenue (
100K
)
2 Web-App
3 Web-App
Facility Configuration
Relative Performance Omni.HoneyGreedyOpt-Stat
Adaptability to Synthetic Variable Load
0
0.2
0.4
0.6
0.8
1
Norm
alise
d P
erf
.
(1:1) (10:1) (15:1) (25:1) (30:1)
HTTP Request Variability
Performance(3 Web-sites)
Omni.HoneyGreedyOpt-Stat
Synthetic Load: Low Variability
0
0.2
0.4
0.6
0.8
1
Norm
alise
d
Perf
.
(2:1
)
(3:1
)
(4:1
)
(5:1
)
(6:1
)
(7:1
)
(8:1
)HTTP Request Variability
Performance(3 Web-sites)
Omni.HoneyGreedyOpt-Stat
Conclusions
Bee heuristic: works well, effective in highly dynamic environment
Competitive against standard heuristics
Bee heuristic: Not tuned, Common sense scaling parameters used
Conclusions
Trade-off static optimality for responsiveness Static optimization requires equalization
of derivatives (marginal rate bee) Bee heuristic has no marginal “bee”
but, instead, has ability to migrate several “bees” at the same time and avoids problem of measuring f’ under variability
Future Work
Test to see if we were lucky or robust Scale up to more patches/web-apps Make autonomic --more feedback loops Power … imitate indolent bees? Convergence rates Compare with IBM’s online network
algorithm
Some other interesting stuff Dominance hierarchies: first
experimental validation of a self-organizing social structure in animals (Chase, Tovey, Martin & Manfredonia 02)
Time lags of environmental costs: mean 10 years vs. mean 5 years for other types. (Regnier & Tovey)
Opportunities for Sr. Design sustainability projects
Some Big OR Questions in Natural Systems
Individual versus group selection: classic argument against latter is essentially an OR proof, but why do forests thrive?
Discounting and EPV, intergenerational equity and intraperiod utility. Relationship to future growth? Intraperiod utility and discounting is almost equivalent to linear utility, Sobel 2000