Transcript Chapter 4 Lecture Presentation

Applications for dynamically
shared GMPLS networks
Malathi Veeraraghavan
University of Virginia
[email protected]
Sept. 24, 2007

Outline



Quick summary of CHEETAH project
"Business" orientation
Technical details of CHEETAH
1
Router-to-router leased EthernetSONET-Ethernet or SONET circuits (red)
e.g., T640s
if not colocated
in same PoP
Regional (metro) network
Enterprise
networks
WAN-access router
Backbone network (e.g., Abilene)
if colocated
in same PoP
e.g., CD-CIs
Regional (metro)
network
Enterprise
networks
WAN-access router
Server-to-server circuits (rather
than router-to-router)

Cheetah studies:

Focused on the use of circuits from server to
server


Since servers only have Ethernet NICs, the
circuits were all Ethernet-SONET-Ethernet
circuits
Focused on enabling dynamic sharing of circuits

Leased lines between servers would likely be
unjustifiable (from cost perspective)
High-speed justification


For router-to-router circuits, "high-speed"
is required because of aggregation
For server-to-server, our justification was
for file transfers

Higher the rate, faster the transfer
Applications we developed for
experimentation with GMPLS networks
Given that a significant % of file transfers
involve the Web, we experimented with two
Web based file-transfer applications:



Simple Web client to Web server transfers

Goal: Use GMPLS network without changing Web client
or Web server software

Problem: GMPLS networks need to stretch end-to-end
Web proxy servers located at core-network PoPs

Goal: If GMPLS network can be only deployed in the
core initially, deploying proxies allows even nonconnected end hosts located in enterprises to benefit
from core GMPLS network's high speeds.
Quick summary of Cheetah project

Deployed a wide-area experimental SONET GMPLS network






Developed core software





Three PoPs: Raleigh, Atlanta, ORNL
Intercity OC192s purchased from NLR and ORNL
Colo services purchased from MCNC, SLR, ORNL
GbE interface cards for server connectivity
Located 2 to 3 servers + GMPLS switch (SN16000) at each PoP
RSVP-TE client for the server
Circuit-TCP for transport protocol on circuits
Modified Web applications to interface with the RSVP-TE client
to request circuit setup before transfers, and release after
Ran our Cheetah core software on HOPI
Interconnected Cheetah to HOPI testbed
Outline check

Outline


Quick summary of CHEETAH project
"Business" orientation



Revenues: potential market - applications
Costs
Technical details of CHEETAH
Business orientation


Choose applications [for development,
experimentation and demonstration] after
taking into account "business"
considerations
So we started by classifying applications
suitable for different types of GMPLS
network deployments and services
Services & applications (for "dynamic
circuit" networks)
GMPLS networks
Leased lines
•
Fine Grained Sharing
(FGS)
TCP/IP
Coarse Grained Sharing
•
•
•
•
Coarse Grained Sharing
(CGS)
High-bandwidth circuits, AND
"Long" holding times
Need Book-Ahead (BA) support in the control-plane
(scheduling or advance reservations)
Fine Grained Sharing
•
•
•
Moderate-BW circuits, and/or
Short holding time
Immediate-Request (IR) mode sufficient in the control-plane.
Services & applications (for "dynamic
circuit" networks)
DCS-network Dynamic circuit
scope services in the core
network ONLY
Bandwidthsharing modes
Dynamic circuit
services are
intraregional
Coarse Grained
Sharing
ISP router-torouter LongDistance (LD)
leased lines
ASP server-toserver LD lines?
•
Web services
(proxy, CDN)
IPTV/video
distribution (CDN)
Inter-SMTP server
•
•
•
Fine Grained
Sharing
•
•
•
•
•
•
•
Dynamic circuit
services involve
regional and core
networks
Disasterrecovery (DR)
Serverreplication
WAN accesslink rate change
•
Software-onthe-web
Backup-storage
WAN accesslink rate change
•
•
•
•
•
Business interconnect
eScience applications
Video-conferencing
Distance-learning
Business interconnect
Video-telephony
Row/column headings: define service types
Entries in the body cells: applications
10
Blue: router-to-router
One sample point

To support the case for providing GMPLS
network based dynamic circuit services
between PoPs


MCI network has 2500 PoPs throughout North
America and 2000 around the globe!
Are there SMTP servers, CDN servers and
other applications servers that need
interconnectivity?
Video and Content Delivery Network
(CDN)





The rise of You-tube and video is often
cited as a reason for growth in bandwidth
and network equipment sales
CDN example providers: Akamai
CDN servers placed in PoPs
Requests from clients served from closest
CDN server
Use high-speed GMPLS networks in the
core to move files between CDN servers
Catch?




Akamai does a trade with regional
Research-and-Education Networks (RENs)
Place servers in regional REN PoP
Regional REN pays for collocation costs
(power, space, remote hands-and-eyes)
Regional REN gains by cutting the required
rate for the circuit it purchases for IP
connectivity from core IP service provider
"Dynamic CDN"



CDN service is comparable to "leased line"
service
A web service provider enters into an agreement
with a CDN provider to serve out its content
What about small-to-moderate sized
enterprises?


Can they recruit CDN servers located at a few PoPs if
they expect a sudden surge of traffic to their web
servers (e.g., slashdot phenomenon)?
If so, use dynamically setup high-speed circuit to copy
over the whole web structure (esp. with databases) to
dynamically recruited CDN servers
Storage

Three types of applications:



Disaster recovery (DR): backup of critical data
Server replication: e.g., of web servers (to
allow for quick switchover in case of failures)
Backup storage: of ordinary enterprise users'
data
DR and server replication


Typically, only these two types require network
connectivity outside the enterprise
Small-to-moderate sized enterprises only require
intraregional DCS services (if used)



general rule of thumb: 75-mile distance of backup site
hence listed in column 2 of services/applications
classification table
Fortune-500 companies with multiple locations
require DCS across regional AND core

hence listed in column 3 of services/applications
classification table
DR and server replication

Requirements



Few endpoints or users initiating these apps.
Few transfers a day
Is IP-routed network sufficient?
Backup storage, on the other hand

"Backup storage" application

If a new "storage ASP" emerged, which sold backup
storage services for "all" data in enterprises, then given



the large number of employees,
who could initiate backup at any time if they want to save an
important file as they make modifications,
could justify needing high-speed DCS networks
Is bandwidth cheaper than HR costs to hire engineers
to maintain backup storage at each enterprise?
"Blue" vs "black" applications in table
Applications
Listed in
Blue
Listed in Black
Endpoints
Router-torouter
Server-toserver
Target market for
"encroachment"
Leased line
services
IP services
Volume and price
Low volume; High volume
High perLow per-unit
unit price
price
Router-to-router circuits

Services (Verizon):



Provide network administrator web portal access to
explicitly request an increase in leased-line rate
 e.g., if GbE interface used, but rate capped with VLAN
rate-limiting, allow for rate limit to be increased
(signaling if leased line realized through SN16000s).
Software that reads SNMP MIBs to monitor usage on
leased line, and automatically issue signaling request for
bandwidth increase
Both ideas: aggregate traffic based
increase/decrease requests
Per-transfer based increases

Even if link is lightly loaded, a single file transfer
delay can be reduced by increasing the bandwidth
of the bottleneck (lowest-rate) link.



e.g., an enterprise has an OC3 WAN access link. Even if
this link is lightly loaded, this becomes the max. rate
that any single file transfer can enjoy.
By dynamically increasing this rate for a few
seconds, user can enjoy a higher transfer rate.
Need tools to determine if WAN access link is
the bottleneck link on an end-to-end path, and
then increase rate.
Costs

Started by seeing Internet2 fee structure

http://www.internet2.edu/network/fees.html
Why GMPLS in core network?


Because high-speed interface cards cost
less in SONET switches than in IP routers
For high switching capacity nodes, which
are mainly required in the core.
What is the major component of cost?

Service provider costs:

Same for IP-routed and SONET networks



HR costs
Bandwidth costs
Differ:

Equipment costs:


Mainly line card costs
If bulk of the costs are in HR and bandwidth,
then equipment cost differentials become less
significant
Summary




Opportunity to increase potential market for GMPLS
switches
We have access to three GMPLS testbeds on which we can
test applications and gain experience with R&E users
 Internet2's DCS, HOPI, Cheetah
Choose application(s) carefully
 with due consideration of business aspects
Looking for support:
 Student HR support to implement "glue" software to
make applications run on GMPLS networks, and to build
usage base
 Cheetah testbed annual maintenance charges
Outline check

Outline


Quick summary of CHEETAH project
"Business" orientation



Revenues: potential market - applications
Costs
Technical details of CHEETAH
Cheetah concept



Hybrid architecture: an IP-routed network AND a GMPLS network
Use dynamically setup circuits for file transfers
Send small files on IP-routed path and use GMPLS network only
for large files: call-setup overhead
IP-routed network
(1)
(2)
(3)
(4)
(5)
(10)
(9)
(8)
(7)
(6)
Circuit gateway
Circuit gateway
NIC1
NIC2
End
host
SONET
SONET
switch
switch
NIC1
NIC2
End
host
GMPLS network
Ethernet
Interface
SONET
Interface
Messages through Internet
Ethernet-EOS-Ethernet CHEETAH circuit
SONET
Interface
Ethernet
Interface
(1)-(5): RSVP-TE PATH messages
(6)-(10): RSVP-TE RESV messages
CHEETAH: Circuit-switched High-speed End-to-End ArcHitecture
CHEETAH End-host Software
Determines which path to use: IP-routed or Circuit
Optical connectivity service (uses DNS servers)
End Host
CHEETAH software
CHEETAH software
OCS Client
End Host
OCS Client
IP-routed network
Application
Routing Decision
Routing Decision
RSVP-TE client
RSVP-TE client
Application
SONET circuitswitched network
TCP/IP
C-TCP/IP
NIC 1
NIC 2
NIC 1
Circuit
Gateway
Circuit
Gateway
Circuit-TCP: TCP minus congestion control
NIC 2
TCP/IP
C-TCP/IP
CHEETAH End-host Software

RSVP-TE client software architecture
End Host
circuitrequestor
bwlib
Sig_proc
OCS Client
Connection Admission Control:
check if bandwidth is available
on the UNI from the host to the
switch (multiple VLANs)
DNS
server
CAC
Data-plane
Configuration
Parsing/
Construction
read
RSVPD
RSVP-TE
messages
Configuration file
Configure IP routing and ARP table
since remote host is reached directly on
the newly setup circuit
CHEETAH testbed
GbE
Raleigh,
NC
ORNL, TN
SN16000
SN16000
Zelda4/5
OC192
OC192
3xGbE
Zelda1/2/3
Wukong/
Wuneng
SN16000
Atlanta
• Long-distance OC192s purchased from NLR and ORNL
• Collocation services purchased from MCNC in NC, SLR in Atlanta
• Zeldas and wukong/wuneng: Linux Dell PCs
30
Interconnection of CHEETAH to
US-wide HOPI experimental testbed
PC3
Chicago
HOPI
Force10
PC3
10GbE
NYC HOPI
Force10
10GbE
Seattle
HOPI
Force10
PC3
HOPI
10GbE
10GbE
PC3
PC3
Washington
HOPI
Force10
LA HOPI
Force10
10GbE
GbE
Zelda4/5
SN16k
NC
ORNL
OC192
SN16k
OC192
GbE
Zelda1/2/3
SN16k
ATL
Wukong/
Wuneng
NxGbE
CHEETAH
HOPI: Hybrid Optical/Packet Infrastructure: Internet2 supported testbed
31
Force10 E600s used to dynamically setup and release VLANs (virtual circuits)
Tech. transfer: CHEETAH control plane
software modified for HOPI
LOSA
10GbE 10GbE GbE

CCSA
pc1


CCSA
pc2


CCPM
Force10
CCPM: CHEETAH Control-Plane Module
CCSA: CHEETAH Client System Agent

pc3
OSPFD
RSVPD
Force10 programming module


RSVPD
CHEETAHD
Circuit-requestor
Circuit setup procedure

losa-pc1:


Internet

losa-CCPM:

STTL
10GbE 10GbE GbE

CCSA
pc1


Force10
pc3

Programs sttl-Force10 for that VLAN
losa-ccpm:


Configures VLAN, programs ARP and route tables
Sends back RESV message
sttl-ccpm:

CCPM
Route extract, Local CAC and VLAN ID check
sttl-pc1:

CCSA
pc2
Route computation, CAC, VLAN ID assignment
sttl-CCPM:


Use circuit-requestor to initiate setup to sttl-pc1
sends PATH meesage
Programs losa-Force10 for that VLAN; sets rate policing
losa-pc1:

Configures VLAN, programs ARP and route tables
32
Setup a circuit from losa-pc1 to sttl-pc1
33
Automatic configuration on the end host
34
Setup multiple circuits to the same remote end host
Request exceeding the
available bandwidth is
rejected.
35
Internet2's new Dynamic Circuit
Services (DCS) network
Yellow nodes: Ciena CD-CI SONET switches
Blue nodes: Juniper T640 IP routers
Courtesy: Rick Summerhill
(2006)
Testbeds
 Three
"GMPLS" wide-area
testbeds are available for testing
and demonstrating new
applications for GMPLS networks
Cheetah
 HOPI
 Internet2's DCS network

Application: WebFT
Web server
Web client
Web Browser
(e.g. Mozilla)
URL
Response
RSVP-TE
daemon
Web Server
(e.g. Apache)
CGI scripts
(download.cgi &
redirection.cgi
WebFT sender
WebFT receiver
RSVP-TE API
C-TCP API
Control messages
via Internet
Data transfers
via a circuit
Cheetah end-host software APIs
and daemons
OCS API RD API
OCS daemon
RSVP-TE API
RD daemon
C-TCP API
RSVP-TE daemon
Cheetah end-host software APIs
and daemons
PROBLEM: Need GMPLS networks to be deployed within
regional and enterprise networks, not just the core
Application: circuit-aware web
proxy servers
IP-routed network
HTTP
messages
squid
Web client
Original HTTP messages
HTTP
messages
squid
Core-only
GMPLS
network
Web server
HTTP and ICP
messages
• A web proxy software package: squid
• "Circuit-aware" by integrating RSVP-TE & CTCP
• Dynamic circuit setup triggered by web client request
• Use of circuits transparent to human users
• Use Internet path while circuit is being setup
39