1. A Biologically-inspired QoS-aware Architecture for
Scalable, Adaptive and Survival Network Systems
Paskorn Champrasert and Junichi Suzuki
Department of Computer Science
University of Massachusetts, Boston
2. Content
• Goal
• Design Approach
• Overview of SymbioticSphere
• Simulation Results
• Conclusion
October 30 2008 DSSG Group Meeting 2/20
3. Motivation
• Large-scale network systems
e.g., Internet Data Center
• A tons of servers and network devices (e.g. router,
load manager) are connected through the high
speed network.
• A lots of users access several services (e.g. web
server) and data (web pages) that Internet Data
Center provides.
• Such network systems still
rapidly keep increasing in
their scale.
October 30 2008 DSSG Group Meeting 3/20
4. Goal
• Making network systems ( e.g. Internet data centers and grid
clusters) to be
– autonomous to avoid interrupting users/administrators frequently
– adaptable to various dynamic changes in network conditions
• e.g., network traffic and resource availability
• in order to…
– improve user experience (i.e. response time)
– expand system’s operational longevity
(e.g. users and administrators don’t want applications down for long time)
– reduce maintenance cost
(e.g. Save money and relieve developers from time-consuming maintenance)
October 30 2008 DSSG Group Meeting 4/20
5. Observation and Approach
• Observation
– Various biological systems have already
developed the mechanisms to achieve key
requirements of network systems.
• e.g. autonomy, adaptability
• c.f. bee colonies, bird flocks, fish schools, etc.
• Approach
– Apply biological concepts and mechanisms to
design network systems (i.e. application
services and middleware platforms).
October 30 2008 DSSG Group Meeting 5/20
6. SymbioticSphere
• SymbioticSphere is a biologically-inspired architecture for network
systems( network applications and middleware platforms)
• An application service (Agent)
– is implemented by an autonomous and distributed agent.
• an agent may implement a web service and contains web
pages.
• A middleware platform (Platform)
– runs on a network host and operates agents.
• Each agent/platform is designed as a biological entity.
– Some biological principles are applied to design agents and
platforms
October 30 2008 DSSG Group Meeting 6/20
7. Design Principles
Principles Biological Systems SymbioticSphere
No centralize entities in No centralized entities to control
Decentralization
biological systems. agents/platforms
Biological entities sense their Agents/platforms sense local
local environments and network environments (e.g.,
Autonomy autonomously invoke their resource availability) and
autonomously invoke behaviors
behaviors. (e.g., migration replication, death)
Biological entities seek and Agents/platforms seek and consume
Life Cycle and consume food for living. energy for living
Food chain -Reproduce (enough food) -Reproduce (enough energy)
-Die (lack of food) -Die (lack of energy)
Biological entities evolve so Agents/platforms evolve their genes
Evolution that the entities that fit better (i.e., behavior policies).
to the environment become Agents/platforms that have effective
more abundant behavior policies become abundant
October 30 2008 DSSG Group Meeting 7/20
8. Energy Exchange
• Human users = the sun
• have unlimited amount of
energy.
• Agents = producers
( e.g. shrubs)
• gain energy from users
• pay some of its energy
level to platforms to
utilize resources
• Platforms = consumers
(e.g. hares)
• gain energy from agents
• periodically evaporate
some of its energy level.
October 30 2008 DSSG Group Meeting 8/20
9. Agents and Platforms
Agent: Die
- Energy level Replicate
- Service (e.g., web pages) Reproduce
- Behaviors*
(reproduction, replication, Migrate
migration, death)** Reproduce
- Behavior policies
Replicate Agent
Platform
Host
Die
Platform:
- Energy level *When an agent/platform invokes a
- Middleware services behavior, it pays energy.
- Behaviors*
(reproduction, replication, death)** ** replication is a special case of
reproduction. Replications happens
- Behavior policies when an agent/platform cannot find a
mate to invoke reproduction.
October 30 2008 DSSG Group Meeting 9/20
10. Behavior Policy
Factor ( F1 ) w1
∑ WFi i
> Threshold ?
w2 i
Factor ( F2 ) Invoke
. Threshold
behavior or not
. wn
.
Factor ( Fn)
• Each agent/platform has its own policy for each behavior.
• A behavior policy
defines when to and how to invoke a particular behavior.
• A behavior policy
consists of factors (Fi), which evaluate environment conditions.
• Each factor is given a weight (Wi) relative to its importance.
• A behavior is invoked if
the weighted sum of its factor values exceeds a threshold.
• Agents/Platforms periodically check weighted sum to invoke behaviors
October 30 2008 DSSG Group Meeting 10/20
11. Agent Behavior Policy
• Factors in agent reproduction behavior
1. Energy Level: ( the agent energy level )
• encourages agents to reproduce their
offspring in response to their high energy level.
2. Request Queue Length:
• The length of a queue, which the local platform
stores incoming service request to.
• encourages agents to reproduce their offspring
in response to high demands for their services.
When the weighted sum of the factors exceeds a threshold, an agent
seeks a mate from the local and neighboring platforms.
If a mate is found the agent invokes the reproduction behavior.
Otherwise, it invokes the replication behavior
October 30 2008 DSSG Group Meeting 11/20
12. Platform Behavior Policy
Factors in platform reproduction behavior
1. Energy Level: Platform energy level
• encourages platforms to reproduce their
offspring in response to higher energy level. Reproduce
2. Resource Availability Ratio: The ratio of resource availability
on a remote host to the local host.
• encourages platforms to reproduce their offspring on
healthier neighboring hosts.
3. The Number of Agents: The number of agents working on the
local platform
• encourages platforms to reproduce their offspring in
response to high agent population on them
October 30 2008 DSSG Group Meeting 12/20
13. Constraint-aware Evolution
• The weight and threshold values in behavior policies have significant
impacts on the adaptability of agents and platforms.
• It is hard to test all possible network conditions and find an appropriate
set of weight and threshold values for the conditions.
There are 18 weight and threshold variables in total. Assuming that
10 different values can be assign to each variable, there are 1018
possible combinations of weight and threshold values.
• SymbioticSphere allows agents and platforms to autonomously find
appropriate weight and threshold values through evolution.
• Behavior policies are encoded as genes for agents and platforms
October 30 2008 DSSG Group Meeting 13/20
14. Mating Partner
• When an agent/platform wants to invoke reproduction behavior, it
finds a mate.
• A mate is selected by ranking agents/platforms on the local and
neighboring hosts. A mate is the agent/platform in the first rank.
• Agents/platforms are ranked with the notion of constraint
domination, which considers:
– Optimization objective
• Energy Level
• The total number of behavior invocations.
• The resource availability of the underlying host.
– QoS constraint violation
• Response Time
(e.g., repose time must be less than 1 second)
• Throughput
The agent/platform that satisfies all of given constraints is said to
be feasible.
October 30 2008 DSSG Group Meeting 14/20
15. Constraint Domination
• Agent/platform i constraint-dominate another agent/platform j if
any of the following conditions are true.
1) i is feasible, and j is not
2) Both i and j are feasible but i dominates j in optimization
objective domain
3) Both i and j are infeasible, but i dominates j in constraint
violation domain
Case 2 Case 3
October 30 2008 DSSG Group Meeting 15/20
16. Genetic Operations
• When an agent/platform invokes reproduction behavior;
it perform genetic operations (crossover and mutation)
• If an agent/platform cannot find a mate, it
invokes replication behavior and perform mutation
October 30 2008 DSSG Group Meeting 16/20
17. Simulation Configurations
• A simulated network system is modeled
as a sever farm.
• 7x7 grid network topology.
– 49 network hosts
• Each agent implements a web service in
its body
• There is one agent and one platform on
each host at the beginning of simulation.
– 49 agents and 49 platforms
100
00
Service Request Rate
(# of requests / min)
Input:
80
00
This service request rate is taken 60
00
from a workload trace of the 1998 40
00
Winter Olympic official website 20
00
0
0 2 4 6 8 10 12 1 16 18 2 22 24
4 0
Simulation time (hour)
October 30 2008 DSSG Group Meeting 17/20
18. Service Request Rate 100
00
(# of requests / min)
80
00
00
60
40
00
00
20
Simulation Results
0
0 2 4 6 8 10 12 14 1 18 20 22 2
6 4
Simulation time (hour)
350
3.5
300 3
The number of agents
Response time (sec)
250 2.5
200 2
150 1.5
100 1
50 0.5
0
0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Simulation time (day) Simulation time (day)
The biologically-inspired mechanisms in The biologically-inspired mechanisms in
SymbioticSphere allow agents to evolve SymbioticSphere allow agents to
their behavior policies and adaptively improve and maintain response time.
adjust their availability.
October 30 2008 DSSG Group Meeting 18/20
19. Simulation Results
• The response time constraint is given:
“An agent must operate a service request in 5 second.”
• After adding the response time constraint,
– The average response time is 15% decreased (1.95 1.66)
– The number of violations is 90% decreased (849 78)
October 30 2008 DSSG Group Meeting 19/20
21. Conclusion
• This paper
– describes how evolution happens in SymbioticSphere and how QoS is
maintained by using QoS-aware evolution.
• Simulation results show that
– agents and platforms autonomously adapt to dynamic environmental
conditions (e.g., user location, network traffic and resource availability) by
using their regular behaviors.
– a quality set of behavior policies can be obtained through evolution in much
shorter time than trial and errors.
Future works
• Dynamic network topology
• Service composition
– Multiple types of agent
October 30 2008 DSSG Group Meeting 21/20
22. Adaptability Measures
• Adaptability is measured as
– Service Adaptation
• Service Availability
– the number of agents
• Quality of Service
– response time of agents for processing service requests
from users
– Resource Adaptation
• Resource availability
– the number of platforms that makes resources available
for agents
• Resource efficiency
– indicates how many service requests can be processed per
resource utilization of agents and platforms.
October 30 2008 DSSG Group Meeting 22/20
23. Regular Behaviors without GA
100
00
Service Request Rate
Input:
(# of requests / min)
80
00
60
00
40
00
20
00
0
0 2 4 6 8 10 12 14 16 18 20 22 24
Output: Simulation time (hour)
400 60
# of agents 100%
Average response time (sec)
20
The number of platforms
The number of agents
# of platforms 50
300 80%
Throughput (%)
Average response time
40 15
Throughput
60%
200 30
10
20 40%
100
10 5 20%
0 0
0 0%
0 2 4 6 8 10 12 14 16 18 20 22 24
0 2 4 6 8 10 12 14 16 18 20 22 24
simulation time (hour)
Simulation time (hour)
Service availability (# of agents) and resource The biological mechanisms in
availability (# of platforms) change dynamically SymbioticSphere contribute for agents and
platforms to collectively retain response
time and throughput performance by
adjusting their populations and locations.
October 30 2008 DSSG Group Meeting 23/20
24. R: regular behaviors
S: regular + symbiotic behaviors
G: genetic operations
October 30 2008 DSSG Group Meeting 24/20
25. Other Results
• Adaptability GRIDNETS 05
Biologically-inspired mechanisms in SymbioticSphere contribute for agents and
platform to adapt to various dynamic changes in network conditions (such as
workload and resource availability); -- improve resource efficiency
• Scalability CIIT 05
Biologically-inspired mechanisms in SymbioticSphere contribute for agents and
platform to scale to large number of network hosts and user request rate.
• Power Saving and Load Balancing ICAS 06
SymbioticSphere saves nearly 50% power consumption at maximum, compared
with traditional network systems
• Self Healing (Survivability) COMPSAC 06
Biologically-inspired mechanisms in SymbioticSphere contribute for agents to
survive network link failures (data center failures) and maintain high
throughput for users.
October 30 2008 DSSG Group Meeting 25/20
26. Request Forwarding
agents
Platform
request Q
User
access point Service daemon
Data center Network host
• When a user requests a service
– the user creates a request message and sends to the data center.
• When service request arrives a host.
– The service daemon checks whether there is a platform and any agents working on its.
• If there is no platform, service daemon sends request msg to neighboring hosts.
• If there is a platform and agents on the host
– Service request msg is placed in service request queue in the platform
– A request message in the queue will be taken by an agent running on the
platform
October 30 2008 DSSG Group Meeting 26/20
27. Agents and Platforms
Agent/platforms behaviors Agent:
Die
- Agent ID
When an agent/platform Replicate
- Energy level
invokes a behavior, it
pays energy.
Reproduce - Service name
- Service
Migrate
- Behaviors
Reproduce - Behavior policies
Replicate Agent
Platform
Host
Die
Platform:
SymbioticSphere service daemon
- Platform ID agents
- runs on network host
- Energy level platform - handles
- Middleware services Service daemon - platform reproduction requests
- Behaviors - host resource availability requests
Host - forward service requests from users
- Behavior policies when there is no platform
October 30 2008 DSSG Group Meeting 27/20
28. Agents/Platforms Cooperation
• Symbiotic behaviors are intended to augment the adaptability of agents and
platforms by allowing two species to cooperate for pursuing their mutual benefits
• Each symbiotic behavior is a sequence of regular behaviors that an agent and its
underlying platform perform in order.
• There are two type of symbiotic behaviors:
1) Agent-initiated symbiotic behaviors (A1 –A3)
– An agent proposes the underlying platform to perform symbiotic behaviors.
– The platform may accept the proposal and perform symbiotic behaviors.
2) Platform-initiated symbiotic behaviors (P1-P3)
– A platform proposes the agents working on it to perform symbiotic behaviors.
– The agent may accept the proposal and perform symbiotic behaviors.
• A symbiotic behavior policy is a behavior policy that each agent/platform
possesses to determine whether it invokes a particular symbiotic behaviors.
– when to propose/accept to perform symbiotic behaviors
[ ( Σ WSi FSi > threshold) and ( condition is true ) ]
October 30 2008 DSSG Group Meeting 28/20
29. Agent-initiated symbiotic behaviors A1:
Condition:
An agent wants to migrate - An agent wants to migrate to host that
toward a user close to user but there is no platform on
that host.
migrate
A A - A Platform has low resource availability
4
Action:
Energy for
Propose 1) Agent proposes to perform A1.
platform
replication 2) Agent gives destination host
1 2 information and pays energy to let
platform replicate.
3) Platform replicates on the host.
replicate 4) Agent migrates
Platform Platform
3
Mutual Benefit:
Low resource availability A platform replicated
Agent can migrate toward to user
closer to a user
-> Response time reduces
A host close -> high chance to get energy
to a user Platform increases resource avail.
-> reduce the chance to be crashed
October 30 2008 DSSG Group Meeting 29/20
30. Agent Behavior Policy
• Factors in agent migration behavior
1. Energy Level: ( the agent energy level )
• encourages agents to move in response to higher energy level.
2. Service Request Ratio:
• The ratio of # of incoming service requests on a remote platform to
the local platform
• encourages agents to move towards users.
3. Resource Availability Ratio:
• The ratio of resource availability (--CPU cycles, memory space,
etc.) on a remote host to the local host
• encourages agents to move to platforms running on healthier hosts
4. Migration Interval: Time interval to perform migration
• discourages agents to migrate too often
October 30 2008 DSSG Group Meeting 30/20
31. Genetic Operations
• When an agent/platform invokes reproduction behavior;
it perform genetic operations (crossover and mutation)
d1 d2
Parent1 Offspring Parent2
F1 F2
d1/d2 = F2/f1
Offspring
Normal distribution mutation
October 30 2008 DSSG Group Meeting 31/20