• No results found

DDS-Enabled Cloud Management Support for Fast Task Offloading

N/A
N/A
Protected

Academic year: 2021

Share "DDS-Enabled Cloud Management Support for Fast Task Offloading"

Copied!
25
0
0

Loading.... (view fulltext now)

Full text

(1)

DDS-Enabled Cloud Management Support for

Fast Task Offloading

IEEE ISCC 2012, Cappadocia Turkey

Antonio Corradi

1

Luca Foschini

1

Javier Povedano-Molina

2

Juan M. Lopez-Soler

2

1

Dipartimento di Elettronica, Informatica, e Sistemistica

Universit`

a di Bologna (Italy)

2

Departamento de Teor´ıa de la Se˜

nal, Telem´

atica y Comunicaciones

Universidad de Granada (Spain)

(2)

Agenda

Cloud Monitoring and Management

DARGOS

Data-Centric Publish-Subscribe

Architecture

Experimental Results

Testbed Description

Results

(3)

Cloud monitoring

Cloud monitoring systems can be categorized:

I

Architectural model: Centralized vs. Decentralized

I

Communication model: Pull vs. Push

(4)

Resource monitoring in Clouds

I

A typical approach: centralized pull

I

Central node queries and stores remote resource usage

I

pros: easy to implement

I

cons: central point of failure, request-reply, scalability in N:M

scenarios, support for different update rates, no notifications

(5)

Centralized Cloud management

(6)

Centralized Cloud management (II)

(7)

Types of loads in Clouds

I

Services

I

Long term duration

I

Load is (almost) stable (e.g. Web server, Databases, ...)

I

Tasks

I

Short duration (from seconds to few minutes)

I

Load of each task is unknown a priori

(8)

Cloud resource monitoring in dynamic scenarios

I

Short-mid tasks with dynamic load

I

Bag of Tasks (BoT)

I

Media transcoding

I

Computation offloading

I

Require an accurate and reliable snapshot of resources

available (real-time update)

I

CPU load, memory usage, system load, hypervisor,...

I

Different goals: maximize throughput, minimize power

(9)

DARGOS

I

Distributed Architecture for Resource manaGement and

mOnitoring in cloudS

I

A distributed monitoring system

I

“Argos Panoptes“: Argos the ”100 eyed“ guardian

I

Uses a Publish Subscribe approach

I

Used to collect real time monitoring data for taking

scheduling decisions

(10)

DCPS

Data-Centric Publish-Subscribe

I

Entities share a data model instead using interfaces

I

Producers publish data conforming this data model

I

Subscribers receive data matching their interests

I

Publishers and subscribers are decoupled in space and time

(11)

DCPS

Data Distribution Service (DDS)

I

OMG Specification for Data-Centric Publish-Subscribe

I

Data model

I

Wire protocol

I

Entities exchange Topics (e.g. temperature, 2D position, ...)

I

Topics are defined by their name and data type

I

Topic samples can contain key data to identify them

I

Publishers pushes Topic updates into Subscribers local cache

I

QoS control and management

I

Partition mechanisms

I

Unicast and multicast support

I

Adopted in time critical systems (avionics, stock exchange

quotations,...)

(12)

DCPS

(13)

Architecture

DARGOS Entities

I

DARGOS has two kinds of entities:

I

Node Monitoring Agent (NMA): collect and publishs local

resource usage

I

Installed at each node (e.g. CPU, system load, memory,...)

I

1 resource, 1 topic

I

Cloud Monitoring Supervisor (CMS): interested in remote

monitoring data

I

Discovers and subscribes remote resources

I

Define their own requirements (reliability, acceptable

deadlines)

I

Installed in every application interested in resource data

(schedulers, dashboard,...)

(14)

Architecture

(15)

Architecture

Node Monitoring Agents (NMA)

I

Collect local resource data and publishes as DARGOS Topics

I

DARGOS NMAs have two operation modes:

I

Periodic

I

NMA pushes periodically resource usage information

I

Maximizes Accuraccy

I

Event based

I

NMA pushes resource information under certain conditions

(e.g. resource usage delta exceeds threshold)

(16)

Architecture

Periodic vs. Event based

Periodic

I

Period=1 second

I

Samples published=10

Event-based

I

Samples sent when usage

changes range

(17)

Architecture

Cloud Monitoring Supervisor (CMS)

I

CMS discovers available nodes and their available sensors

(DARGOS Topics)

I

CMS subscribe to sensor information of interest (CPU,

memory,...)

I

Applications that use CMS: Cloud dashboards, schedulers

I

Each CMS define their own quality of service (QoS)

requirements

I

Reliability or best effort (RELIABILITY)

I

Maximum allowable delay between updates (DEADLINE)

I

Maximum refresh rate (TIME BASED FILTER)

I

CMSs establish subscription contracts with NMAs

(18)

Architecture

DARGOS-based Cloud management

(19)

Testbed Description

Experimental testbed

I

Testbed with DARGOS-enabled OpenStack Cloud

I

DARGOS based OpenStack scheduler

I

Server Consolidation

I

Load balancing

(20)

Testbed Description

Testbed description

I

OpenStack Cloud fabric

I

DARGOS enabled scheduler service

I

Three DARGOS enabled compute nodes

I

RTI DDS 4.5d middleware

(21)

Results

Results (VM per node)

OpenStack out-of-the-box scheduler

OpenStack DARGOS-based

scheduler (consolidation)

(22)

Results

Results (bandwidth)

(23)

Conclusions

I

Typical centralized Cloud monitoring systems are not suitable

for dynamic scenarios

I

DARGOS is a decentralized Cloud monitoring system suitable

for fast task

I

DARGOS can also satisfy service scenarios

I

DARGOS is more robust than centralized systems

I

The Data-Centric Publish-Subscribe model used by DARGOS

makes possible to manage task oriented Clouds accurately and

reliably

(24)

Future Directions

I

Include more sofisticated scheduling algorithm (e.g. include

historical data)

I

Extend the notification mechanism to include alarms or

complex events

I

Add customized action registration to automatically react to

certain events (e.g. live migrations)

(25)

Q&A

References

Related documents