Transcript Study on Adaptation of CDN Request-Routing to Scalable

Study on Adaptation of CDN Request-Routing
to Scalable Conference System
Toshiyuki KAWASAKI* Koji OKAMURA**
* Graduate School of Information Science and
Electrical Engineering (ISEE), Kyushu University
** Computing and Communications Center,
Kyushu University
Aug 28,2003
1
Contents
１．Background
２．Goal and approach
３．Evaluation
４．Consideration
５．Conclusion
2
１．Background－H.323 conference system
• H.323 conference system is popular
– Remote conference
– Remote medical care
– Remote education
Host A
Host B
• H.323
– Standard of video conference
on IP networks
Host A
Host B
• MCU（Multipoint Control Unit）
– Enable multipoint conference
– Mixing the video and audio streams
Host C
Host D 3
1．Background－Overload of MCU
• Problem of H.323 multipoint conference system
When hosts (participants) increase・・・
MCU may become overload
→load sharing is needed
←overload
MCU
• Existing method of load sharing
– Distribute the load to several MCUs
When more hosts increase
→MCUs may become overload after all
←overload
Lack of scalability for the number of hosts
4
２． Goal and approach －Load Sharing by Hosts
Proposal method
• Don’t use any MCUs
• Hosts connect with
each other
• Hosts play a role of
MCUs
Existing method
– Hosts distribute load
– increase of hosts
→increase of MCUs
Obtain scalability for the number of hosts
Proposal method
5
２． Goal and approach － Application Level Multicast
• Problem of the multipoint conference system without MCU
– Each host should select the host which it connects with.
• Existing method
– ALM ( Application Level Multicast)
• Multicast on logical network (overlay network)
• Hosts play the role of multicast router on logical network.
Lobby server
Make Make
links high
cost links
Entry
between
ASs
IP address list
New host
AS2
Overlay network
AS1
6
２．Goal and approach － ORIO and CDN Request Routing
• Proposal method ORIO（Optimal Routing Inter Organization）
– Using CDN Request-Routing
• CDN（Content Delivery Network）
– Network of severs to distribute content efficiently
– Load sharing by using many surrogates
• Request Routing
– Select the optimal surrogate for the clients
– Tell clients their optimal surrogates
contents
select
notify
Redirection server
notify
Delivery
requestcontents
Origin server
surrogate
surrogate
surrogate
surrogate
client
client
client
client
client
client
7
２． Goal and approach － Request routing using DNS
• Process of participation at ORIO
– Request Routing using DNS
• Prepare DNS to redirect requests.
• New participants write the certain host name for participation
into the packets destination address.
• DNS tell the optimal hosts IP address at name resolution.
→Connected with its optimal host automatically
AS 3
DNS server
AS 1
133.x.x.xxx
connect
select
conference.ac.jp（host name）
133.x.x.xxx（IP address）
New host
AS 2
8
３．Evaluation－Evaluation Environment
Simulation based evaluation using network simulator “NS２”
• Evaluation standard
– Maximum network load
• Comparing with the method
using ALM
• Experimental Network
– SuperSINET
– The number of ASs is １８
– The number of hosts per ASs is 64
• Environment
– Speakers※ １、Hosts※ １～２５６
• Change distribution of hosts
• Change the number of active AS
Experimental Network（SuperSINET）
– Speakers １～１２８、Hosts１２８
※Speakers = the number of source hosts, hosts = the number of destination hosts
9
３．Evaluation－simulation under the 3 distribution of hosts
• Speakers １, Hosts １～２５６
– Distribute hosts to 6 ASs
– Measure under the 3
distribution of hosts
Maximum use of
bandwidth (kbps)
1 area (ORIO)
2 area (ORIO)
3 area (ORIO)
• Concentrate at 1 area
• Concentrate at 2 areas
• Concentrate at 3 areas
• Result
1 area (ALM)
2 area (ALM)
3 area (ALM)
The number of hosts
– the increase amount of network load is decreasing as the number
of participants increase
→ Scalability to the number of hosts.
– ORIO is better than ALM at load sharing
10
3.Evaluation － Change the number of ASs
• Speakers1,Hosts１～２５６
– Changing the number of
ASs which hosts belong to
Maximum use of
bandwidth (kbps)
• Select hosts from 6 ASs
• Select hosts from 12 ASs
• Select hosts from 18 ASs
• Result
The number of hosts
– ORIO is better than ALM at load sharing
– ORIO’s efficiency of load sharing become worse when the
number of AS is small ( hosts per AS is big).
→ Policy of routing within AS should be improved.
11
3.Evaluation － Speakers
Maximum use of
bandwidth (Mbps)
• Change the number
of Speakers
Speakers １～１２８
Hosts １２８
– Select hosts randomly
• Result
The number of speakers
– Network load increase linearly at both ORIO and ALM.
→Don’t provide scalability in this case.
12
The number of speakers should be limited.
４．Consideration－Efficiency of ORIO and restriction of ALM
• ORIO provide more efficient load sharing than ALM
• Possible reason
・ ALM uses logical network for routing,
but ORIO doesn’t use it.
In ALM, there are two restrictions in receiving streams
• Receivable only from its neighbors on the
overlay network
• Receivable only from its upstream hosts
13
４．Consideration－restriction of ALM
• Receivable only from its neighbors on the
overlay network
• Receivable only from its upstream hosts
→ Promote the concentration of the load
speaker
Speaker Concentrate
Concentrate
the load
Host C
delay
5ms
Router
router
Router
router
Can’t use this
delay
delay
link
though
10ms
7ms
Host A it is useful Host B
the load
Concentrate the load
Concentrate the load
Speaker
speaker
Host C
delay
9ms
delay
5ms
delay
7ms
delay
9ms
delay
7ms
delay
delay
7ms
10ms
neighbor
is not
A’s
Host A BNot
Host B
upstream host
(B’s delay is bigger than A)
14
５．Conclusion
• Existing multipoint conference system
– MCU easily become overload
• Method of load sharing
– Any MCU is not used
– Hosts play a role of MCUs instead
• The system which tells the hosts
their connection partner is required
– CDN Request-Routing
• Result
– Prove the scalability of ORIO
• Future work
– Adding the function which support the hosts
exiting process
– Improve the policy of routing within AS
15