Application Network Deployment in the Internet PhD Dissertation Public Defense January 15, 2004 Oscar Ardaiz Villanueva Advisor: Leandro Navarro Moldes.

Application Network Deploymentin the Internet

PhD Dissertation Public Defense

January 15, 2004

Oscar Ardaiz Villanueva

Advisor: Leandro Navarro Moldes

Index

Towards a Thesis Background, problem statement, solution requirements

First contribution ”A framework for application network deployment in an Internet service programmable

infrastructure allows creating application networks which are provisioned dynamically, and coordinated for service control”

Hypotheses, validation methodology, experiment and results Second contribution

”A resource initiated allocation and multicast based deployment mechanisms is appropriate for application network deployment in a large-scale network as the Internet”

Hypotheses, validation methodology, experiments and results Conclusions

Contributions, socio-economical impact, further work

Towards a Thesis








Application Networks

Nodes: Server Links: TCP Virtual Circuits Application Layer Network

Squid Server Parent-Child comm. channel. Web Proxy Caching Hierarchy

IRC server chat distribution comm.channel. Chat Server Network

HTTP

Server

IP link

Host

Router

Server

Server

Server

ServerClient

HTTP

HTTP

HTTP

Overlay Networks: State of Art

ServiceImplementation

Layer 7 overlay Layer 3 overlay

Application Application networks: Proxy-caches hyrarchy, chat networks, P2P applications: file sharing,..

--

Network P2P network overlays: Narada, Abone, Alan,....

Network overlays: Mbone, 6bone, VPN,...

Motivations for Overlays

Testbed to experiment new services (Mbone, 6-bone,..),

Isolated environment (VPN,..), Provision distributed services (Web-caches,

CDNs,..), Sharing distributed resources (Gnutella

OceanStore,..).

Overlay Research Issues

Organization: load balancing, fault tolerance, …

Mapping to underlying network, Virtualization for support of multiple

overlays, Isolation among different overlays. Scalability to the Internet.

Problem Statement (I)

“How to introduce a new application network service?”

Today, it requires operators to perform manually: 1. Discovery of hosts and network topology. 2. Lay out application network to satisfy demand. 3. Coordinated installation, configuration and start-up of

servers. 4. Update name / redirector service.

Or servers are activated and connected by end users (Specially peer-to-peer networks).

Problem Statement (II) Spatial and temporal demand variation

require application network adjustment.

S

Service Provider

Service Demand Variation

Application Network

Adjustment

Solution requirements (I)1. Programmable infrastructure (nodes providing resources and programmability) for remote activation of new services.

Service Provider

Programmable Infrastructure

Nodes

Solution requirements (II)2. Deployment mechanisms for coordinated resource

allocation, server activation and connection and organization.

Service Provider

Programmable Infrastructure

Nodes

Application Network

Application Network Deployed

3. Clients access application network efficiently.

Service Provider

Application Network Clients

Clients

Clients

Programmable Infrastructures: State of Art

Application Service WebOS: OS for the Internet.Active Services: Mobile code (Java) used for server mobility.Grids: Distributed resource sharing.

Network Service Active Network: Mobile code (Java) used to program routers. Programmable network: API for programmability of network router.

Motivation: facilitate incorporation of new services.

Prog. Infra. = Internet-wide Resource pool + Programmability.

Solution Space

No programmable infrastructure permits application network deployment.

Only Xbone permits network layer overlay deployment.

Xweb

Programmable Infrastructures:

WebOS, Grids, Active Services

Web

Coordinated Servers & Resources

On the fly Activation / Configuration

Manual Modifications

Service Control (quality service, efficient resource utilization)

Programmability (temporal/ spatial/ functional variation

Isolated Servers

Application CDN, P2P

Networks: Radar, ASPs,

Ejasent

First Contribution

”A framework for application network deployment in an Internet service programmable infrastructure allows creating application networks which are provisioned dynamically, and coordinated for service control”

First Contribution


First contribution ”A framework for application network deployment in an Internet service

programmable infrastructure allows creating application networks which are provisioned dynamically, and coordinated for service control”





Hypotheses H1.1-1.3 Framework

H1.1 An application network deployment framework requires the following building blocks: a programmable Internet service infrastructure, resource availability, discovery and monitoring mechanisms, service specification mechanisms, resource-mapping and deployment-plan creation and control, and resource allocation, code distribution and service composition mechanisms.

H1.2 Framework functionality is divided between deployment managers, which are situated at service provider locations, and resources agents, which are placed at resource provider nodes.

H1.3 Code authentication and virtual machines with code verification are required for resource security. Party authentication, message integrity checks and communication encryption are required for communication security. Resource authentication, results integrity checks and encrypted program execution are required for applications’ security.

Hypotheses H1.4-1.7 Programmable Infrastructure

H1.4 The programmable Internet service infrastructure provides shared resources, virtualization support for multiple applications and an execution environment.

H1.5 Resource availability is determined by monitoring, notification or probing mechanisms.

H1.6 Local resource managers at resource nodes allocate resources to applications.

H1.7 Package installation utilities and system calls for program execution provide resource binding to permanent storage, memory and processor resources. Programs make use of other system calls to obtain resource bindings to network and other specific resources.

Hypotheses H1.8-1.14 Deployment Mechanisms

H1.8 Server coordination requires the creation of communication channels between server instances and the configuration of coordination rules at each node or in a centralized server.

H1.9 Resource discovery and monitoring is provided by proactive mechanisms that are based on multicast expanded ring searches.

H1.10 Application network specifications are made up of a number of session requirements, a service demand table, and service-wide constraints.

H1.11 Connected placement is a suitable mechanism for resource mapping.

H1.12 Deployment control mechanisms are required so that resource-allocation and service-composition operations are carried out atomically on multiple remote nodes.

H1.13 HTTP-based mechanisms for information retrieval provide code distribution.

H1.14 The deployment of application networks is faster than it is by manual operation.

Validation Methodology

Resources

Execution Environments

Rsrc Agent

CPU

Resource Allocation

Code Distribution

Application Installation

Application Activation

Deply Mgr

BW Disk / RAM

Win Linux JVM

Resource Availability

EE Availability

Application Coordination

Application Communication

App1

Code Server App1 Code

RsrcAgent

CPU

ResourceDiscovery

Deply Mgr

ResourceAvailabilityTable

CommandsSentTable

ResourceAllocation

UserResourceMapping

CodeDistributionServiceComposition

DeploymentPlan Creation& Control

EE

Resources

DeployRequest

Prototype Implementation: Deployment manager used by

service deployers. Implements Web-based interface.

Resource agents used at each programmable infrastructure node.

Communicate by TCP.

Validation MethodologyExperiments: Deployment of 3 different kind of application networks: web proxy

server, chat server networks and servlet application networks:

Set up:

Measurement of temporal response: tTotal_Deployment, tNode_Deployment tNode_Actions

Specifications Proxy cache

hierarchy Chat server

network Servlet CDN

Application code Squid IRCd HelloWorld.java Execution

environment Linux-i386 Linux-i386 Tomcat

Per-server storage 200 Mbytes 200 Kbytes 20 Kbytes Net per-client BW 10 Kbps 1 Kbps 1 Kbps

Client session duration

10 sec 100 sec 1 sec

Demand regions All All All No. clients per region 100 100 1000

Constraints None None None

Storage:200MbyteBW: 10Mbit

LINUXTOMCAT

App1:

srv2

App2:

srv3

App3:

srv2

1 Mbit100 Mbyte1 Mbit2 Mbit

50 Mbyte

10 Mbyte


LINUXTOMCAT

App1:

svr1

App2:

srv2

App3:

srv1


50 Mbyte

10 Mbyte


LINUXTOMCAT

App1:

srv3

App2:

srv1

App3:

srv3


50 Mbyte

10 Mbyte

Svr2 parentsSrv3

Svr2 parentsSrv1

Svr2 parentsSrv3

Svr3 parentsSrv1

Svr3 parentsSrv2

Svr1 parentsSrv2

Svr2 parentsSrv1

H1.4-1.7: Programmable Infrastructure Results

Multiple application networks per infrastructure node require resource virtualization abstractions: Socket ports, multitasking OS, virtual memory.

Strict resource partitioning at resource nodes should be provided by mechanisms such as Qlinux or resource containers.

Dynamic resource binding mechanisms are required: reconfigurable sockets, virtual containers, ..

H1.8-1.14: Deployment Mechanisms Results

Mapping of application network specifications to resources depends on performance model of resource.

The deployment of application networks is faster than it is by manual operation:

main overhead is code distribution: can be improved by content distribution networks, caching mechanisms, incremental code loading.

Decoupling deployment of communication channels might simplify framework.

Cache ALN Chat ALN Servlet CDN

tTotal_Deployment18275 ms 2502 ms N/A

tCode_Distribution10284 ms 1530 ms 100 ms

H1.1-1.3: Framework Results

Most security threads solved by public key security technology. Application Java virtual machines with code verification are required for resource security. However there is no solution for malicious hosts accessing/attacking programs.

Building blocks and architecture validated by implementation and experiments, thus:

”A framework for application network deployment in an Internet service programmable infrastructure allows creating

application networks which are provisioned dynamically, and coordinated for service control”

However large scale experiments could not be performed.

Simulation to evaluate behaviour in large scale. (Today would use a testbed: planetlab.org )

Deployment in a large scale InternetDrawbacks of previous mechanisms in the large scale: resource monitoring creates a network bottleneck

and requires large server. resource mapping algorithm are computational

intensive. Resource contention among different deployers is

possible.

Resource allocation in the large scale had to be studied further.

Resource allocation mechanisms SoACentralized server initiated allocation: total control, but not scalable (Darwin,

Globus, ..)

Deployer Resource

“ServiceSpec”-> ::mapping -----“Allocation Command”--------------------> ::CheckAvaila

Resource initiated allocation: Distributed operation, scalable (RSVP,…)

Deployer Resource

“ServiceSpec”-> :: ---------------“Resource Requirements”-----> :: Mapping :: Allocation

Resource initiated allocation for distributed services has not been proposed.

Second Contribution


Second Contribution








H2.1 Service-initiated allocation deployment is not scalable since it requires high bandwidth at deployers location.

H2.2 Service-initiated allocation deployment is not scalable because of resource contention among deployer due to stale information.

H2.3 Multicast injection requires no resource discovery or monitoring mechanism.H2.4 Multicast injection uses simple deployers.H2.5 Multicast injection requires service specifications to be made public.H2.6 Multicast injection requires resource provider to interpret service specifications.H2.7 Multicast injection permits resources to control its usage.H2.8 Multicast injection can create resource over provisioning.H2.9 Multicast injection implementing resource-initiated allocation has high success

allocation ratio.H2.10 Multicast injection implementing resource-initiated allocation has high total

network bandwidth.H2.11 Multicast injection implementing resource-initiated allocation creates high

percentage of unused allocations.

Hypotheses H2.1-2.11

Validation Methodology

Implementation Ns-2 Process.tcl and Agent.tcl module are extended

to simulate deployment manager and resource agent. Simulation

Topology similar to Inet, smaller scale.

Parameters: resources quantity and distribution, application network specifications.

Variations in resource load, quantity of deployment entities and deployment rate.

S

S

S

SS

S

S

SS

S

S

S

SS

SS

S

S

S

SS

S

S

S

S

4 Mbits / 20 msec. link



D D

D

D

D

D

S Surrogate

Deployer

H2.3-2.8 Results

Deployer

1) Deployer mcast Spec:

Deploy “AppId, Regions: C1;C2;C3, Demand 1;2;1; Max distance 1, TopoLevel 2”

2) Nodes accept servicing:

Node 4,6,7,8,9,10,11,12,14

“OK; I allocate resources”

3) Nodes self-organize:

•Node 14 “parent 8”, Node 7 “parent 8”, Node 10 “parent 8”, Node 9 “parent 4”, Node 8 “parent 4”

Infrastructura Programable

H2.3-2.8 Results

Multicast injection uses simple deployer and no resource discovery or monitoring mechanism.

Common definition of service specifications (f.e. Internet regions, performance parameters) is very important.

Some service providers might not want to make public its service specifications, therefore secure multicast is required.

Resource overprovisioning happens and is not desired: multicast damping is a solution.

S

H2.3-2.8 Results

3) Nodes create multicast group, and exchange Info:

•Node 6,7“distance 1 C1”

•Node 9,10 “distance 1 C2”

•Node 11,14 “distance 1 C3”

•Node 4,8 “distance 2 C1,C2,C3”

QUIT QUIT

QUIT

4) Nodes Quit (unusing resources):

•Node 6,14 “QUIT”

•Node 12 “QUIT”

Multicast damping eliminates resource overprovisioning.

Deployer

Nodes Allocate after Multicast

Multicast injection has higher success ratio.(Centralized allocations suffer contention).

Multicast injection creates unused allocations, but not critical Input limiting filters will eliminate any overload.

H2.2-2.9-2.11 Results

0

50

100

150

200

250

300

350

400

0 0,5 1 1,5 2 2,5 3 3,5

Resource Load (resource units requests per second)

Ap

ps

Dep

loye

d S

ucc

essf

ul

MI CC

0

1

2

3

4

5

6

7

8

9

10

0 0,5 1 1,5 2 2,5 3 3,5

Resource Load (resource units requests per second)

Un

use

d A

lloca

tio

ns

(%)

MI

CC

Polinómica (MI)Polinómica (CC)

H2.1-2.10 Results

0

50

100

150

200

250

300

350

400

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7

Aggregate Deployment Rate( Deployment requests per second)

De

plo

ye

r B

an

dw

idth

(B

yte

s/s

)

MI CC

Multicast injection requires less network bandwith at deployers locations.

Multicast injection uses higher total network bandwidth create multicast regions to

diminishes traffic. 0

2

4

6

8

10

12

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7

Aggregate Deployment Rate (Deployment requests per second)

To

tal

Net

wo

rk B

and

wid

th (K

by

tes

/s *

ho

ps)

MI

CC

Conclusions








Contributions

Initial Hypotheses about “application network deployment in the Internet”.

Validated by implementation, experiments and simulation, thus the thesis:

“A framework for application network deployment in an Internet service programmable infrastructure allows creating application networks which are provisioned

dynamically, and coordinated for service control””A resource initiated allocation and multicast based deployment mechanisms is appropriate for application

network deployment in a large-scale network as the Internet”

Socio-economical impact

Less costly to set up application network services more people can provide them, many different services will be provided.

Application networks can adapt to demand variations applications with higher quality of service,

Virtual infrastructures creation is facilitated more virtual organizations shall be created.

Research Lines

Virtualization of other distributed systems incorporating physical resources.

Application to other programmable infrastructures: grids.

Economic based resource allocation mechanisms.

Application Network Deploymentin the Internet

PhD Dissertation Public Defense

January 15, 2004

Oscar Ardaiz Villanueva

Advisor: Leandro Navarro Moldes

Application Network Deployment in the Internet PhD Dissertation Public Defense January 15, 2004 Oscar Ardaiz Villanueva Advisor: Leandro Navarro Moldes.

Documents

application network

largescale network

underlying network

network topology

networkp2p network overlays

application networks

work slide

overlay networks