www.ict.csiro .au Self-Tuning End-2-End QoS Internet ICT Centre Presented by: Zvi Rosberg 21 March 2007
Dec 19, 2015
www.ict.csiro.au
Self-Tuning End-2-End QoS Internet
ICT CentrePresented by: Zvi Rosberg
21 March 2007
www.ict.csiro.au
A Non-fictional Dialogue between a scientist and a network operator / VC
on a new pioneering idea of a
“self-tuning end-2-end QoS Internet”
www.ict.csiro.auA Non-fictional Dialogue
I have this great idea for implementing end-2-end
Quality of Service in the Internet and Enterprise networks
Come on, who needs it?
Quality of Service or QoS: A set of performance levels required by some Internet applications.
www.ict.csiro.auA Non-fictional Dialogue
Well …, Internet-Phone (VoIP), Video-On- Demand (VOD), Internet
Gamers…
Indeed, but Internet gives you 100-250 msec packet delay using JUST “best-effort ” service, why bother?
best-effort service : flows without
any performance guarantee
www.ict.csiro.auA Non-fictional Dialogue
QoS is not just packet delay, what about jitter, bandwidth and packet
loss guarantee?
We have enough bandwidth !
www.ict.csiro.auA Non-fictional Dialogue
Coming to money, so why not make more money by deploying a more
economical solution ?
Now you talking !
www.ict.csiro.auA Non-fictional Dialogue
Firstly, how would you prevent shaping of streaming traffic during congestion, if you
use just best-effort flows ?
No worries !
Data use TCP, TCP flow control reacts to congestion - that’s OK, data can wait.
Streaming use UDP, UDP ignores congestion - that’s OK too, streaming needs priority.
TCP: A transport communication protocol used
by data applications in the Internet such as Email and
Web browsing .
UDP: A communication protocol used by streaming applications in the Internet
such as VoIP and VOD.
www.ict.csiro.auA Non-fictional Dialogue
Aha aha…, due to firewalls most streaming traffic is “tunnelled” via
open port 80 of HTTP
HTTP: An application protocol used between Web Servers and Browsers
www.ict.csiro.auA Non-fictional Dialogue
DiffServ guarantees only per-hop-behaviour and requires complex router configuration along with a
network provisioning tool.
IntServ can guarantee Intra-domain end-to-end QoS, but it is not
scalable
OK, so why aren’t the IETF RFC’s of IntServ and DiffServ
catching up ?
IETF: All-volunteer open standards organization that set the Internet
protocols. Each protocol is specified in an RFC document.
IntServ and DiffServ are two short names of QoS standards standing
for Integrated Services and Differentiated Services.
www.ict.csiro.auA Non-fictional Dialogue
Everything that DiffServ and IntServ don’t have – A Self-Tuning end-to-end
QoS Internet
So, what are you proposing ?
www.ict.csiro.auA Non-fictional Dialogue
• Bandwidth, delay and packet loss guarantee
• … not just per-hop, but end-to-end and across domains
• … and an extra bonus, “fairness”.
www.ict.csiro.auA Non-fictional Dialogue
Fairness is tricky. For some it means they get all bandwidth, for others it means they get just what they ask
for, etc.
I like fairness, customers were asking about it, but
what is it?
www.ict.csiro.auA Non-fictional Dialogue
There is a fairness concept that spans a spectrum of “fairness levels”, and we can tune our algorithms to any
level you like
So which one do we use?
www.ict.csiro.auA Non-fictional Dialogue
We add a new “Resource Management Protocol” (RMP) at the IP network layer which computes and probes
(similar to ATM).
The connection layer will use its information to shape the user flows.
That’s a lot, how do you implement it ?
www.ict.csiro.auA Non-fictional Dialogue
We will experiment with it in our research network and then bring it
to IETF
Hmm…, it’s not standard!
www.ict.csiro.auA Non-fictional Dialogue
Indeed, we do want it in the routers, but for experimental purposes we will place it in our own Network Processor Units and
attach them to the routers, …. as illustrated in the next diagram.
But your network layer RMP must run in the routers, CISCO
wouldn’t let you in.
CISCO: A networking equipment vendor that develops and
manufactures the majority of network devices comprising the
Internet.
www.ict.csiro.au
EdgeRouter
10Eth
erne
t
CoreRouter
11
CoreRouter
12
EdgeRouter
20
CoreRouter
13
CoreRouter
14
CoreRouter
15
EdgeRouter
30
Ethernet
Ethernet
User Devices
User Devices
User Devices
1
2
3
51
52
53
3
4
RMPEdge
Device
RMPCore
Device
RMPCore
Device
RMP Component Layout & Operation Principles
RMPEdge
Device
RMPEdge
Device
RMPCore
Device
RMPCore
Device
• RMP packets traverse along flow paths and
• calculates distributively the flows’ fair rates
• Also used for probing packet delay and loss
flow path
Operation Principles
www.ict.csiro.auA Non-fictional Dialogue
It means that RMP adapts to traffic changes and does also Connection
Admission Control (CAC)
What is this “self-tuning” thing?
www.ict.csiro.au… A Non-fictional Dialogue
:
• We proved stability of some of the algorithms.
• We demonstrated stability of the entire system by simulation, and got what we have expected.
[1] G. Rogers, J. Chan, and D. Agahari, “Rate Control of Elastic Traffic with QoS Guarantees: a Stability Analysis & Experimental Implementation,” in Proc. IFIP/IEEE NOMS 2006, Apr. 2006.
[2] Z. Rosberg and M. Zukerman, “Multi-Service Flow Control and Fairness for All.”
(Submitted to IEEE Trans. On Networking.)
[3] Z. Rosberg, “Self-Tuning Multi-Service Flow Control.” (Submitted to INFOCOM’08)
• However, we also want to: • test it in our research network where real reliability and performance
issues could affect the stability of the algorithms …
• prove stability of other elements of the algorithm.
• extend it to multicast protocols
Will it work?
www.ict.csiro.auWhat could we learn from this dialogue ?
• There are still open problems in the Internet and enterprise networks that need investigation
• Even if you derive solid solutions for a problem, substantiated by mathematical proofs and simulation, it is not sufficient to make an impact – you need to demonstrate it in a real network.
• Internet flow control is a hot and competitive research topic being investigated by world-class researchers from Cambridge, Caltech, UC Berkeley, University of Illinois and others.
• Our research network will give us an edge over Universities