Distributed Control Algorithms

Slide 1:Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Michael Barry Andrew T. Campbell Andras Veres EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 2:The paper�

Investigates service differentiation in wireless packet networks Proposes solution based on 802.11 DCF VMAC (Virtual MAC) Estimates MAC � level statistics of the channel( delay, delay variation, packet collision, packet loss) VS (Virtual Source) Utilizes MAC to estimate application-level service quality Demonstrates through simulations that globally stable state may be achieved if the proposed algorithms are applied to admission control Discusses distributed service level management scheme built on top of the introduced algorithms EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 3:Motivation

TCP/IP will be the glue for all applications in mobile environments Desirable that wireless architecture supports the same quality assurances as the wireline Internet Wired internet � 2 main approaches. 1st approach Take circuit switching and extend with datagram service Very strict control over wireless/wireline resources Very rigid, complex EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 4:Motivation (cont�d)

2nd approach ( which this paper is based on) : Only minimal control and signalling Should accommodate many different (unforseen) wireless applications Good example : 802.11 Does not guarantee anything but the best-effort service (In either DCF of PCF mode) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 5:Motivation (cont�d)

Base on DCF because PCF is good, but � Not currently supported by most wireless cards Cooperation with DCF (which is used now) can lead to poor performance Distributed control (DCF) is more efficient for real-time services then the centralized one (PCF). EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 6:Brief DiffServ Introduction

Concept of Service Differentiation Differentiating among services provided by a network to different types of customers and applications IP-layer service differentiated service examples: value-added internet services VPNs application hosting/outsourcing packet telephony EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 7:Brief DiffServ Introduction (cont�d)

Proposed and spearheaded by IETF DiffServ task force Designed to help networks satisfy QoS req�ts. Allow different classes of service to be provided to streams on a common network infrastructure. One of two major models used in QoS The other one is Int-Serv Reservation-less (as opposed to Int-Serv) No state in routers is needed (as opposed to Int-Serv) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 8:Brief DiffServ Introduction (cont�d)

Push major state and work to the very edges of the networks. Fan-in and forwarding speeds of flows are smaller Packets carry their state in a few bits (DS CodePoint) Flows are policed and marked at the edge (when they enter the network) To support different classes of IP service : packet classifiers forwarding/per-hop-behavior traffic conditioning policies EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 9:Brief DiffServ Introduction (cont�d)

Services are built by applying rules Rules for how packets are marked initially Rules for how marked packets are treated at boundaries. Three elements work together to deliver a Diffserv service: Per-Hop Behaviors (PHBs) Deliver special treatment to packets at forwarding time Traffic Conditioners Alter packet aggregates to enforces rules for services Bandwidth Brokers (Policy Managers) Apply and communicate policy EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 10:Brief DiffServ Introduction (links)

This is where I found info used in these slides http://www.qosforum.com/docs/faq/ http://www.ietf.org/html.charters/diffserv-charter.html ftp://ftp.netlab.ohio-state.edu/pub/jain/courses/cis788-97/integrated_services/index.htm ftp://ftp.netlab.ohio-state.edu/pub/jain/courses/cis788-99/h_6qos/sld001.htm EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 11:Approach to providing differentiation

MAC layer must be differentiation-aware to be able to provide any kind of differentiated services in mobile environments Propose differentiation-aware MAC based on DCF Differentiation at radio level not sufficient to provide proper differentiation for different traffic types Propose VMAC (virtual mac) and VS (virtual source) which monitor MAC � level load. Base admission control on the estimates. EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 12:DCF

Distributed Coordination Function Covered in the EE206A notes CSMA/CA RTS/CTS for hidden-terminal Sense channel for DIFS then tx, else backoff and sense again ACKs are used EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 13:DCF � backoff strategy

Central to this paper Tbackoff = Rand(0,CW) * Tslot Tslot is slot time (contention window is broken into slots) Tbackoff initializes the timer. Timer is decreased only when nobody is tx�ing. Otherwise it is frozen. Timer is decreased every time the medium idle for DIFS (distributed interframe space) interval If no ACK after tx: double CW until you reach Cwmax EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 14:What should diffserv-enabled MAC be?

Distribute available radio resources between classes ensuring that different classes receive different level of service Adaptive and robust Decentralized There has been work done to achieve this, but proposed solutions have drawbacks, which are corrected (obviously) by this paper. EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 15:Backoff timers as means of differentiation

Idea: change the CW values to reflect priority. Even if collisions, CW values changed @ same rate CWmin � lower for higher priority traffic (smaller wait) Cwmax � also lower for higher priority traffic. Needed to handle time intervals with high congestion EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Drawback: No explicit guarantee of level of differentiation Only relative timings, and even those are based on random number generators.

Slide 16:Simulation and results

Ns-2 used with wireless extension by Monarch group Mix of best effort TCP and better-then-best-effort voice connections Vary CWmin for both traffic classes Increasing CWmin significantly delays TCP packets, and only slightly affects voice traffic Choice of CWmin significantly affects voice, but not best-effort traffice The above two are confusing, so here is the graph� EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 17:Simulation - graph

EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 18:Simulation - robustness

Previous slides prove that differentiation is achievable Can it be maintained over range of traffic conditions? Simulate increasing levels of traffic up to channel saturation. Yes. Delay goes up for both priority levels, but separation is maintained. TCP is not starved even at saturation point. That�s due to statistical and non-deterministic nature of service separation. EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 19:Simulation � robustness (cont�d)

EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 20:VMAC and VS � WHY?

You cant exactly plan the wireless channel If you cant plan � adapt through measurement Most applications want absolute bounds, not relative EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 21:VMAC and VS � HOW?

Free capacity of the channel may be estimated by measuring idle channel time after DIFS Remember, nodes are required to sense channel for DIFS before even attempting transmissions in DCF Algorithms passive, operate in parallel to �real� MAC algorithms VS and VMAC are related: VS algorithm is actually a combination of Virtual Application, Interface Queue, and a Virtual MAC (VMAC) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 22:VMAC and VS � HOW? (cont�d)

Virtual Application generates packets Virtual packets timestamped and placed in the queue Scheduling on wireless channel: much as one would for real channel Sensing Backoff Collision detection (if you would tx a virtual packet on a busy medium) One difference from the real MAC: packets are not actually transmitted All aspects of real MAC are emulated (retries, etc) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 23:VMAC and VS � HOW? (cont�d)

�real� node would detect collision when it�s RTS timer would go off without receiving a CTS Virtual nodes detect collisions if they sense a channel busy, when they are ready to transmit a virtual packet Virtual nodes enter backoff procedure after the time equal to RTS timer of a real node MAC delay on no collision: D = tdefer+ tRTS+ tCTS+ tpacket+ tACK+ 3tSIFS+ 3 tau tau � maximum propagation delay EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 24:VMAC and VS - Performance

There is an argument which suggests that Virtual MAC is absolutely not resource intensive. (virtual and simulated MAC delay) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 25:VMAC and VS � Performance (cont�d)

Average Delay of a new voice source EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 26:VMAC and VS � Performance (cont�d)

Delay variation for a new voice source EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 27:Virtual Delay Curves

Another important part of the VS algorithm Average delay of virtual packets if the VS algorithm generates packets at the rate of Prate Example of a tradeoff: at the same bitrate Reduce application delay by increasing packet rate Yet, higher packet rates load channel more Applications can make use of this to shape their traffic Represented by a function d(Prate), where Prate is packet inter-arrival time Data bitrate is constant Psize * Prate = const EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 28:Virtual Delay Curves (cont�d)

Mobile host or base station runs VS algorithms with several Pratevalues in parallel From delay values obtained from VS you may construct a delay (and/or delay variance) curve From the curve a base station or a mobile host can choose optimal packet rate and packet size so that the application experiences minimum delay and delay variance EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 29:Distributed Admission Control

Based on VS and VMAC Natural location � at the base stations For better performance, mobiles should have that as well (since radio channel properties may be different in mobiles then in base-stations) Based on the idea that it�s better to drop then to delay past bound packets belonging to better-then-best-effort traffic Again, studied through simulation EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 30:Distributed Admission Control (cont�d)

Aggregate rates of traffic in simulation (notice burstiness of TCP) EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 31:Distributed Admission Control (cont�d)

Estimated delays by VS algorithms in base stations Few base stations stop admitting traffic when load goes way high. Some are in continuous accept. This is due to overlap- ping areas of coverage Bursty traffic handled well EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 32:Distributed Service-level Management in a Wireless Network

What does this have in common with the rest of the paper? Topic of admission control and the word �distributed� EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 33:Distributed Service-level Management in a Wireless Network

Mobile moves into the service area, and performs handoff you need to perform admission control. Base station needs to know SLA Service Level Agreement �Contract� between provider/user specifying (for example) what bitrates the user will supply, and what kind of performance the network can guarantee So, you need to transfer the SLA context somehow It also seems as if this is the very first time DiffServ actually kicks in. Before, nothing said proposed schemes could be made to work with Int-Serv EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 34:Distributed Service-level Management in a Wireless Network (cont�d)

First time the mobile enters the network make it obtain an SLA from some centralized control agent. SLA is transferred as a �token� SLA token is encrypted, and base-stations have keys Once Mobile Node obtains SLA it will submit it to base stations upon entering the cells. This way base-stations need not communicate, and are able to perform admission control locally EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 35:Critique?

Did not see too many things I did not like Admission control in base-stations � does it contradict DiffServ principles of pushing decisions to the edges? Seems performance would be impacted by virtual mac although maybe nothing major. If placed in base-stations, there will be much more then 20 (as the graph shows) nodes They argue its always better to drop packets then delay them for better-then-best-effort traffic. Should this not be left up to applications? Again, maybe contradicts the earlier �robustness� comment? Same statistical / non-deterministic nature that �does not allow TCP traffic to be starved� may very well starve it in some cases (or near-starve it) Distributed Service Level Management� does not seem to be very well integrated into the paper EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Slide 36:References

The EE206A slides (Spring �01) http://www.en.polyu.edu.hk/~cmsp/Events/doc/tutorial_QoS.pdf EE206A � SPRING 2001 � Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks