Transcript Flow - MMLab

Microflow-based QoS Transport
Technology
Dr. Lawrence G. Roberts
Founder & Vice Chairman Caspian Networks Inc.
Seokjae Ha
Contents
Introduction
QoS
Flow
Caspian Networks flow-based router
background
Architecture
Comparison with other technologies
Conclusion
QoS(Quality of Service) Overview
“Best Effort” service
DropTail Queue, FIFO scheme
Opposite to “Best Effort” service
Guaranteed, Assured Service
A numerous QoS enabling technologies are suggested
ATM
IP QoS
intServ. diffServ
MPLS
Necessary Architecture & Protocols for QoS
Signaling Protocols(Q.2931, RSVP, CR-LDP,…)
Per flow queuing & packet scheduling algorithm(WFQ, WRR, …)
Different queue management schemes(RED, …)
However….
QoS is not popular
Per flow overhead & scheduling complexity
Not so scalable enough to apply large networks like
WAN
End-end QoS – network solution
Internet research is now focuses on Fairness and
Service differentiation(CoS)
Its HW requirements are not so large compared to QoS
technologies
“Better than” Best effort service
Near-future solution
But…..
QoS technology is not thrown away because
It gives so much good features
QoS technology recovers its strength when HW
technology is fully matured
Telecommunications Protocols and QoS
Flows / Trunk
QoS
Objective : Per flow control of rate, loss and delay
10 M -
Full Quality of
Service –
All flows shaped
1M100 K -
End of
Life
10 K 1K100 -
10 NCP
11970
IP-DiffServ
Class of Service
Group or Class
shaped
No QoS — No State
TCP/IPv4
1980
1990
2000
2010
Flow
A stream of data between on user/system and another in a given
interval
Web Access. VoIP call. File Transfer, P2P transfer, etc.
In IPv4 it is uniquely identified by the 5- tupple (D-Add. S-Add.
Protocol. D-port. S-Port)
In IPv6 it is uniquely identified by the 3-tupple (D-Add. S-Add,
Flow Label)
Flow in traffic is fully exploited in
Firewall
QoS based router
Service differentiation
Caspian Networks Technology
Is Best Effort IP Really Economic?
Router Price with 8 High Speed Ports
500
$K
400
300
Memory Cost
For Flow
State
200
100
1970
Ports are of Highest Speed Available in that Year
1975
1980
1985
1990
1995
2000
2005
2010
Link speed (50kbps in 1969, ARPANET) is up to 10Gbps
# of flows is from 100 to 2,000,000
cost of flow state memory became economic in the late 1990’s.
0.6% of system cost is memory(really?)
Routing History – Byte, Packet, Flow
Routing Technology Improvement
Less Decisions / bit reduces routing cost, not port cost
Cost
1
TDM – One Byte per Decision
1 Byte
Decisions per Bit
0.1
$
1969
40:1
0.01
ATM – 1 cell / decision
40 Bytes
52 Bytes
0.001
TDM Switching
0.0001
500 Bytes
ATM Switching
2003 14:1
Packet Routing
0.00001
1960
Flow Routing
1970
7000 Bytes
1980
1990
2000
2010
What is Flow Routing?
All routers were packet routers from 1969 to 2003
They examine only the packet and keep no history about the
flow
This allow them to route the packet, dropping by priority
(DiffServ)
They cannot determine the duration, rate or byte count of the
flow
Flow routers are also IP routers and can be
intermixed with packet routers
They keep Flow State about all packets is each active flow
There are about 100 K flows/Gbps. This requires memory
However, they can determine the duration, rate or byte count
of the flow
Thus, they can identify flow types and control the rate and
delay per flow
Flow-State Routing: The Technology(I)
Route each packet
Switch to output
Class-based QoS
Flow-State Router
Hash for flow identification
1.5M flows/s and 6M flows per 10 Gig
RAM
RAM
Route Each Packet
Queue (Class) & Forward
RAM
RAM
RAM
RAM
Switch Fabric
Conventional Router
Route, switch, filters, stats
Per-flow QoS behavior
Leverage flow state for advanced QoS
Shape, police, CAC, congestion control
No Flow Management Issues
Simple profile definitions
Hash, Lookup State, Route,
Store, WFQ/Flow, Switch
Lookup State, Store, and
WFQ/Flow
RAM
RAM
Switching
Network
Create “soft” state or look up
Flow-State Routing: The Technology(II)
Flow Router
Conventional Router
1. Hash 5-tuple for Flow
1. Route Every Packet
2. Lookup State or Create
2. Switch to Output
3. Route once per Flow
3. 8 Queues – Drop Tail
4. Discard per Flow
Lots of SRAM to keep Order
5. Calendar Schedule
Only DDRAM – Low Cost
Queue and Forward
SRAM SRAM SRAM
Switch Fabric
Route Each Packet
DDRAM
Hash, Lookup State, Route,
Store, WFQ/Flow, Switch
Lookup State, Store,
WFQ/Flow
DDRAM
DDRAM
Switching
Network
DDRAM
SRAM SRAM SRAM
Flow-State Routing vs. Other Flow Concepts***

Flow: Def’n is flexible,
but generally any unique
combination of 5 tuples:

State: Managing and
maintaining the following
information for each flow:
 Source Address
 Switch Fabric Route
 Destination Address
 Nexthop
 Source Port
 Class
 Destination Port
 Rate
 Protocol
 Delay Variation
Forwarding Info
QoS Info
 Byte Count
 Flow Duration
Netflow
Cisco CEF/Riverstone
 Packets Received/Dropped,
Bytes Received/Dropped,
etc.
Statistics Info
Other vendors only manage subsets of Caspian’s flow state data
Flow State – Identify & Control
File
Transfer
Dimensions
• Rate
• Total Bytes
• Total Time
• Packet Size
• Port
• Protocol
• DiffServ
P2P
500
450
400
Rate
(Kb/s)
350
P2P with Adaptive Discard
300
Controls
• Rate
• Priority
• Delay Var.
• Loss Rate
250
200
150
HTTP
100
Skype
50
0
0
0.5
1
VoIP
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Time (Minutes)
Flow State provides information about each flow.
This permits the router to identify the type of traffic over time
Enables control of the rate, delay, or loss allowed for that flow.
Comparisons of Protocols
Protocol
Calls/Sec
per Gbps
BW
Guarantee
Delay
Guarantee
Data*
Utilization
TDM Voice
50
Yes
Yes
25%
IP/Diffserv
25,000
No
No
40%
Frame
Relay
1,000
Composite
No
80%
ATM
1,000
Yes
Yes
80%
MPLS
1,000
No
No
40%
Optical
Circuit Net
1,000
Yes
Yes
25%
Flow
Routing
25,000
Yes
Yes
80%
Capabilities and Applications
Flow-State QoS solution for IP networks
Includes per-class/aggregate/LSP capabilities
Per-Flow Capabilities Include:
Per-flow rate guarantees
Per-flow policing
QoE: Non-Interactive Traffic
Control
Reducing P2P problem traffic, or
turning it into revenue
Per-flow Adaptive Drop Probability
Other customer-definable fairness algorithms
Per-flow Statistics
Premium IP Services
New technology = new services =
new revenue
Per-flow shaping
Per-flow CAC
No complex signaling and reservation schemes
OR complicated flow-management
Videoconferencing over IP
Enabling cost effective and
deterministic IP Videoconferencing
Unique Characteristics of a Flow Router
Dynamic Load Balancing
Fast Error Recovery
Guaranteed Bandwidth
Guarantees for Flow Groups
Maximum Rate Traffic (UDP) CAC Control
TCP Slow Start Improvement
TCP Fairness and multiple SLA’s
High Trunk and Fabric Utilization
IPv6 QoS Signaling Standard (TIA Aug 04)
Sender AR=100, GR=2
AR=30, GR=2
AR=30, GR=2 Router AR=35, GR=2
Router
AR=35, GR=2
Router
Router
AR=30, GR=2
AR=30, GR=2
Receiver
Router
AR=30, GR=2
First Packet – Negotiate Rates
First Response – Return Rates
AR=30, GR=2
•TCP Rate Feedback
• Faster WWW, Files
• Guaranteed Rate Setup
•No Loss Video, Voice
•Uses IPv6 header options
•Signaling is between adjacent
routers
•Each router confirms or
changes the requests for a
specific rate and delay
• Sets up a “best available
bandwidth” path
•Enables a very high probability
of delivery
IPv6 QoS Signaling Value add
TCP Time to Get 1MB Page
Cros s Country - RTT=100 m s
IPv6 rate ne gotiate d of 32 Mbps
1,200
IPv4
With QoS Signaling
and
32 Mbps agreed
1,000
IPv6/QoS With TCP Slow-Start
32 Mbps TCP
Rate Negotiated
Typical TCP Slow Start
Kilo Bytes
800
600
400
200
0
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
Se conds
Up to 10x improvement in TCP download speeds
Guaranteed rates for real time applications like video
“Precedence” on IP for critical applications such as emergency calls
Summary
Flow-State QoS is Essential for Premium IP Networking
Network optimization vs. over-provisioning reduces TCO
Service Level Guarantees vs. Agreements leads to improved customer
satisfaction
Enhanced margin vs. falling ARPU requires efficient delivery of valueadd services (triple play)
The only technology to build true next-generation
networks
IP Scalability, IP Security, IP Control, IP Performance
Compliments existing IP infrastructure
IP convergence evolution
Immediate, Medium and Longer Term Network
Applications
References
Dr. Lawrence G. Roberts, “Is Best Effort IP Really Economic?”, IPv6
Newsletter, June 2004
Dr. Lawrence G. Roberts, “The Next Generation of IP - Flow Routing”,
SSGRR 2003S International Conference, L’Aquila Italy, July 29, 2003
www.caspiannetworks.com
C. Barakat, P. Thiran, G. Iannaccone, C. Diot, P. Owezarski, “A flow-based
model for Internet backbone traffic”, Proceeding of IMW 2002, ACM Press,
Marseille France, November 2002
Any Questions or comments?