Download Document

www.ait.edu.gr
AN OPTICAL NETWORK
INFRASTRUCTURE SUITABLE FOR
GLOBAL GRID COMPUTING
D. Simeonidou, R. Nejabati & M. J. O’Mahony
University of Essex
Wivenhoe Park, Colchester CO4 3SQ, UK
A. Tzanakaki & I. Tomkos
Athens Information Technology Center
Markopoulou Ave., PO. BOX 68, 190 02 Peania, Athens, Greece
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Outline
• Global Grid scenario based on optical networks
• Appropriate switching paradigm
•Optical Burst Switching
• The functional blocks required:
•Core Router
•Edge router functionality
•Grid User Network Interface
•Grid Resource Network Interface
• Challenges and solutions in terms of functionality and
technology
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Global Computing
• Local computational resources cannot keep up
with the demands generated by some
users/applications:
– high-volume (long- or short-lived) jobs with high
demands for processing and storage
– large number of medium/small jobs requesting
distributed resources
• Distributed computing using the concept of a
global computational Grid is proposed
– It is not a new paradigm but until recently networks
could not support the features and capabilities to offer
efficient use of remote resources
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Infrastructures for Grids
• The Grid application characteristics and requirements influence the
choice of a suitable network infrastructure
– transport and switching format
– control signalling and management
– physical layer technology
•
Grid applications differ with respect to the following characteristics:
–
–
–
–
–
granularity of traffic flows
required data transaction bandwidth
QoS and acceptable delay
throughput and packet loss
storage capacity
– processing power etc.
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Applications
• Particle physics large international collaborations and experiments involve
enormous amounts of data requiring processing and analysis of Petabytes
per year
• Very Long Baseline Interferometry (VLBI) used by radio astronomers for
detailed images and experiments bring data from distributed instruments to
a central point to correlate the signals from individual telescopes
• High Performance Computing and Visualization focuses on adapting and
developing parallel codes for execution on parallel processors. Remote
visualisation of terabytes of data requiring high bandwidth links ~ 1 Gbit/s or
hundreds Mbit/s that increase with the number of remote observers
• E-Health, e.g. Mammography introduces increased capacity requirements
due to size and quantity of scan images
– For 100 patients to be screened remotely, the network would have to carry 1.2GB
of data every 30 seconds
– For this type of application speed of data transfer is important
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Grid Features and Characteristics
• Suitable network infrastructures are required to offer very different features
compared to traditional telecommunications infrastructures
• In telecommunications networks when traffic demands arise there is always a
predetermined pair of two discrete points that need to communicate
• In Grid networks particular end-users/applications require to access available
network resources for processing/storage and need to identify their availability
– Only the source is predetermined, while the destination has to be discovered and
identified using intelligent mechanisms supporting advanced signalling schemes
•Self organization
• In Grids bidirectionality and symmetry of connections unlike
telecommunications networks is not necessary. The two directions are
generally required, one to discover and access the network resources and
submit the job and the other to extract the results and deliver them to the user
– However, the two directions can be decoupled and set-up independently
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Grid Networks I
• The common requirements in this type of networks are
summarised as follows:
– High capacity for bulk data transfer, low cost bandwidth on
demand for short or long periods of time between discrete points
across the network (i.e. point and click provisioning)
– Service granularity at the wavelength and sub-wavelength level
– Multicasting capabilities
– Hardware flexibility to support wide range of different distributed
resources
– Resilience to different layers, from the application layer to the
wavelength layer
– Network security
– Ability to provide management and control of distributed network
resources to the user/application
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Grid Networks II
• In these networks resource request, discovery and allocation are
performed initially when a processing requirement arises. There is
no bandwidth reservation in advance - core routers decide on the fly
where to forward the data
• Core routers require some application and network-level awareness
to configure the resources best suited for the task. This can be
achieved utilising two types of information:
– The data needs to be accompanied by information on the nature of the
job e.g. estimated computational and storage capacity, QoS related
information (feed-forward information)
– Information about the status of the network needs to be included. The
grid resources provide periodically information about their status and
availability e.g. free storage capacity, computational load, and network
resources (feed-back information)
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Optical Network Infrastructure for Grids I
User / Grid application
Computational & storage
resources
Resource-Network
Transport/Signalling
User-Network
Transport/Signalling
Optical router
Optical router
with
BGP functionality
User / Grid application
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Optical Network Infrastructure for Grids II
• Distributed computing becomes a realistic solution with
the recent advancements of optical networks based on
wavelength division multiplexing (WDM)
• Can support a distributed control plane to allow
control/access and even ownership of network resources
by the users/applications in contrast to traditional
telecommunications networks
• Set-up, control and tear-down of end-to-end lightpaths
across multiple domains can be provided through the
use of
– existing protocols like Generalised Multiprotocol Label Switching
(GMPLS)
– new protocols such as the Optical Border Gateway Protocol
(OBGP)
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Switching Paradigm for Grids
• Optical burst switching (OBS) supports multi-service
traffic, offering high granularity (sub-wavelength), high
spectral efficiency as well as bandwidth and low latency
• Offers transport for highly demanding Grid applications
and all-optical data transmission with ultra-fast
user/application-initiated light path setup
• Accommodates bursty traffic with improved network
economics and provide convergence of electronic and
optical technologies
• Enables control and management integration and
simplification offering a distributed control plane
supporting advanced signaling schemes
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Optical Burst Switching
• OBS assumes burst aggregation at the edge of the network with
burst lengths from 10s of kB to several MB
• OBS is based on the asynchronous operation mode with variable
length optical bursts depending on the nature of the application
• The control information is out of band and transmitted prior to the
burst
– It is at a lower data-rate than the data burst
– It is processed independently of the data burst mostly in the electronic
domain
– The value of the time offset may be used for offering QoS differentiation
•Resulting in reduction or elimination of optical buffering
• The data burst is transparently switched and routed through the
network without the requirement of any optoelectronic conversion
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
OBS Network Scenario I
Local
resource manager
User’s job
in form of an optical burst
Grid edge device
GRNI
Computational
resources
GUNI
signalling/transport
Local
resource manager
Grid edge device
GUNI
Intelligent
OBS router
Grid User Network
Interface
(Job submission gateway)
Grid edge device
GRNI
GRNI
signalling
Local
resource manager
Grid edge device GRNI
transport
GRNI
Grid Resource Network
Interface
Computational & storage
resources
Storage
resources
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
OBS Network Scenario I
• OBS can support a set of important requirements for Grid users
listed below:
•
•
•
•
•
•
Application initiated lightpath set-up
Resource discovery and allocation mechanisms
Delivery of jobs to the available resources
Delivery of processing results (if there are any) to the user
Dedicated network feedback mechanisms to user
Providing necessary flexibility in architectures to support both
carrier-owned and user-owned networks
• Supporting the requirements for both physical and application layer
QoS
• Light-trees & Optical Multicasting
• OBS can deal with a wide variety of applications and
accommodate small, medium and large jobs to support the
application requirements in terms of duration, latency etc
• OBS does not require resource reservation and does not impose
any strict requirement for symmetry in the two directions of
transmission
www.ait.edu.gr
Optical Network Infrastructures for Grid Computing
Optical Multicasting
G
A
E
C
F
B
D
•
Optical multicasting is the concept of 1:N light-trees
•
A light-tree is a clear channel originating at a given source node having multiple
destination nodes i.e. is a point-to-multipoint channel. The use of light-trees can
significantly reduce the number of hops (or lightpaths) that the data has to
traverse and therefore significantly improve the throughput of the network.
•
Optical multicasting can be applied in OBS network scenarios for grid
applications to improve network performance and efficiency. In general avoiding
multicasting at a higher layer:
•reduces requirements for optoelectronic conversions
•limits the need for store-and-forward functions
•enhances the virtual connectivity of the network
www.ait.edu.gr
Optical Network Infrastructures for Grid Computing
Core Routers - Optical Burst
Switches
Buffering
Switching
1
1
.
.
.
N input ports
Input
Processing
.
.
.
.
.
.
N
Output
Processing
.
.
.
N output ports
N
Electronic Control
Routing Table
Input processing:
Switching:
Output processing:
Buffering:
Routing table:
Control:
power equalisation, activates control
provides path between switch input port and required output port
optical burst conditioning
not always required, stores bursts when contention arises
stores routing information
control switch fabric, with info from control packet and routing table
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Optical Burst Switches I
• An optical burst switch architecture and the appropriate switching
technology should offer advanced features such as
•
•
•
•
•
•
•
dynamic reconfiguration with high switching speed ( s)
strictly non-blocking connectivity between input and output ports
multicasting capabilities
capability to address contention issues and QoS differentiation
scalability
upgradeability
minimum performance degradation for all paths
• To resolve contention in the optical burst switch the following options
and their combination can be used:
• Wavelength dimension: wavelength conversion
• Space dimension: deflection routing, in which optical bursts are
diverted through a different route to their destination nodes e.g. hot
potato scheme
• Time dimension: optical or electronic buffering
www.ait.edu.gr
Optical Network Infrastructures for Grid Computing
Optical Burst Switches II
1
..
.
Optical
fabric
1
..
.
..
.
Optical
fabric
..
.
Optical
multicasting
..
.
m
..
.
..
.
..
.
m
• Multicasting operation can be achieved using passive splitters and couplers
• Alternative solutions may be also attractive in terms of performance,
functionality and scalability such as:
•
•
•
multicasting switches with fast response
wavelength converters supporting multicasting capabilities
other optical/optoelectronic devices and subsystems offering similar features
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Grid User-Network Interface (GUNI) I
The Grid edge device with GUNI functionality is responsible for:
• User job pre-processing and transmission entity construction
•Job classification, aggregation (grooming)
•Optical burst assembly
•Flexible bandwidth allocation
• Pre-processing the results coming back from Grid resources
•Send back results to the users
• Anycasting the optical burst to the core nodes
•λ-selection for anycast
• User authentication and security check, job acceptance or
rejection
• Grid service billing and accounting
• Fault detection, protection and restoration
www.ait.edu.gr
Optical Network Infrastructures for Grid Computing
Grid User-Network Interface (GUNI) II
Application
Layer
Application Layer
Collective
Collective
Resource
Resource
GUNI signalling
Connectivity
Connectivity
architecture
protocolArchitecture
GRID
Protocol
Grid
GRID middleware
GRID
middleware
GUNI
Functionalities
Optical Burst
OBS
Control
Control
OpticalDATA
Burst
OBS
Data
Fabric
layer
Fabric
layer
Optical Burst
OBS
DATA
Data
Optical Burst
OBS
DATA
Data
www.ait.edu.gr
Optical Network Infrastructures for Grid Computing
Grid Resource-Network Interface (GRNI)
GUNI
GUNI transport
lling
igna
S
I
GUN
Grid
User/Application
Grid
Edge
Grid Enabled
Optical Network
GRNI S
ignalling
GRNI
Grid
Edge
Local
resource manager
GR
NI
tr
• In GRNI is responsible for:
• Pre-processing of the incoming optical bursts
an
spo
rt
Computational & storage
resources
• Optical burst segregation
• Job submission to local Gird resources
•
Advertising state of the local resources:
• Broadcasting of the available processing/storage capacity to the optical network
• Sending back results of a completed job
• Optical burst assembly
• Bandwidth allocation and light-path setup
Optical Network Infrastructures for Grid Computing
www.ait.edu.gr
Conclusions
•
A variety of applications and the requirements they imposed on global Grid networks
have been discussed
•
A novel Grid network scenario based on optical infrastructures has been proposed
•
The solution is based on the optical burst switching paradigm to fulfil the Grid
application specific traffic requirements and offer efficient sharing of network
resources
•
The fundamental functional blocks needed are the Core Router, the Grid User
Network Interface and the Grid Resource Network Interface
•
The core router based on optical technologies is able to support routing of the optical
bursts on the fly and provide advanced features such as optical multicasting
•
The optical GUNI is able to support fast and dynamic burst assembly and wavelength
allocation per burst
•
The GRNI provides simple signalling between the local resources and the optical
network while it offers a data transport mechanism between Grid resources and the
optical network
Thank you!
[email protected]