Course notes by Gregor v. Bochmann, University of Ottawa

Network Quality of Service

QoS at the network level

The quality parameters of data transmission services provided by networks are usually the following:

loss rate (by packet, by cell, or some other unit)
error rate (bit error rate, rate of erroneous packets, or other measures which may take into account that bit errors often occur in bursts)
average delay
delay variation, also called jitter (either standard deviation of delay distribution, or difference between minimum and maximum delay)

Clearly these parameters will depend on the throughput which is determined by the data source. Sometimes the available throughput is limited (then it may also be considered a QoS parameter of the network service, namely the maximum bandwidth available). Example: If an application sends data with a bandwidth of more more than 56kbits/sec over a modem network access link, then the loss rate will be high since packets must be dropped by the sending computer as soon as the allocated send-buffers overflow. This may occur for video transmission. In the case that the TCP protocol is used, this protocol provides flow control which forces the application to slow down when there are problems with the data rate, that is, when either the sink may be too slow, or some point in the network (e.g. the network access link) may be a bottleneck.

In many cases, the throughput required by an application is not constant, but variable (e.g. compressed video). This is called vaiable bit rate. People have designed various ways to characterize the variable throughput of applications (traffic characterization), using traffic models based on a leaky bucket or token bucket (see [Lu, Sect. 5.7 through 5.9]).

In the context of ATM, various classes of traffic have been defined, as well as the QoS parameters provided by an ATM network, which may be negotiated for each new ATM connection (see slides S1 , S2 , S3 ; for more details, see Chapters 6 on "QoS and ATM" in [Ferg 98] P. Ferguson and G. Huston, Quality of Service - Delivering QoS on the Internet and in Corporate Networks, John Wiley, 1998).

Congestion management, TCP slow-start and TCP-friendliness

In the Internet, packets get lost when a router is overloaded and the packet buffer for one of the outgoing links overflows. When the application uses the TCP protocol, this will lead to packet retransmission, which will further enhance the congestion of the network. Considering that (presently) most applications on the Internet use the TCP protocol, people have added a congestion management feature to TCP, called slow start and exponential back-off. It is implemented in most TCP implementations and has the effect that the effective throughput of a TCP connection reduces when packet losses are encountered within the flow of the TCP connection.

A so-called congestion window is introduced that limits the size of the transmission window, it has initially the value of one maximum segment size. Slow start has two stages. The first stage is actually not "slow": when all packets corresponding to one congestion window have been acknowledged, the window is doubled. When the window reaches the congestion threshold (initially set to 64Koctets), then the window is increased each time only by one maximum segment size (this stage is "slow"). However, when a packet is lost, that is no acknowledgement is received for a given packet for a certain time-out period, the slow start process starts again, but the threshold is set to half the last value of the congestion window. If there is a bottleneck in the network, all TCP connections routing their packet through that bottleneck will have a certain probability of having packet losses, and will therefore drastically reduce their throughput. This will alleviate the congestion status of that bottleneck (see example from Tanenbaum's book).

"TCP friendliness" means that UDP traffic is "friendly" with TCP traffic, that is, if there are losses, UDP traffic will also reduce its bandwidth for an amount similar to what TCP traffic would do. This means that there is no unfair competition between TCP and UDP traffic. Without this "TCP friendliness", UDP traffic would continue sending packets independently of the loss rate; as a result, the congestion will probably remain and the TCP traffic will be largely reduced, while the UDP traffic continues (possibly with unacceptable loss rate).

The standard congestion management feature of TCP does not work well at very high transmission bandwidth. Therefore various modifications to TCP have been proposed to make it suitable for hish-speed grid computing. One of these proposals is "Scalable TCP". A comparison of normal and scalable TCP was presented in the "DFN Mitteilungen" (March 2004) for some particular Grid session in Europe (note: the value 1000 corresponds to the available bandwidth in this example).

QoS in the Internet: Future developments

In contrast to ATM networks, the Internet currently only provides so-called "best-effort" service, that means, the QoS parameters of data transmission are not at all guaranteed. Applications have no choice, and the actually provided service quality depends strongly on the time of the day and other factors, including inpredictable congestion situations. In order to allow some form of QoS management, applications may use measurement data (obtained from monitoring the data transmitted by the application, or from a network monitor which may reside in the local host computer, or provided as a service by the local networking domain).

In the future, two approaches for providing some choices to applications are foreseen (for more details, see paper by Xiao ):

Integrated Service: This approach uses the RSVP protocol to make reservations of resources throughout a path through the network for a given flow of data. This reservation is done before data can be sent. (disadvantage: not scalable)
Differentiated Service: This approach defines different classes of service and an application may select the service class it will use. Usually, the user has negotiated a Service Level Agreement (SLA) with the network service provider which states what maximum bandwidth will be available for each service class (which will necessarily be smaller than the available bandwidth of the network access point). SLAs may be negotiated statically, that is, for a longer period, or dynamically (for the period of a particular application, e.g. a teleconference). (disadvantage: in most cases no guarantee of specific QoS parameters for any of the service classes, only relative performance). << For an overview of what information is included in current SLAs between (larger) customers and service providers, see here. >>

Here are some interesting further readings, which are not part of the official course content, relating to possible future developments of the Internet:

Gigabit Ethernet (and future higher speed versions), directly running over optical wavelengths may provide the right framing structure for sending IP packets over the future high-speed networks, as explained << here >>.
An approach to providing QoS in the future Internet, based on flow-oriented reservations, is explained in the paper << "Internet traffic, QoS and Pricing" >>.

Switching versus Routing

A switch for time-division multiplexing works as follows: The switch is connected to N input lines and N output lines. In one multiplexing period, it receives on each input line M blocks of data belonging to the M different logical channels that are multiplexed over the input line. The switch has a table (which I will call forwarding table) which indicates for each input block characterized by (n, m) [input line n and time unit m] at what position (n', m') output line n' and output time unit m'] the block should be output.

In a network, the forwarding tables at all switching nodes must be established in such a manner that the blocks of data are sent along the switched connections that are established end-to-end through the network. When a new connection is established, the tables along the path of that connection are updated accordingly. All data over a given connection follows the same path which is determined when the connection is established.

The path for a new connection is either determined by a routing algorithm executed in a central control unit which then signals the forwarding table updates to each node, or it is determined in a distributed manner, usually through a signalling protocol which forwards a connection establishment request from one node to the next and each node determines the next hop of the connection path by looking up the best choice in a local routing table. The establishment of these routing table is a matter for a distributed routing protocol.

The switching function is very similar in the case of ATM switching, MPLS (Multi-Protocol Label Switching), and and so-called Lambda Switching used for Dense Wavelength Multiplexing (DWDM) in optical networks. Here is a correspondence:

Technology Unit of data processed Channel characterization method for updating forwarding tables in the nodes of the network method for determining the end-to-end routes realized by the forwarding tables

Time-division multiplexing block time unit in which block is received / sent SS7 signaling protocol

ATM switching cell (56 octets) virtual path and virtual circuit identifiers in cell header PNNI signaling protocol PNNI routing protocol

MPLS packet value of the label added to the packet Label Distribution Protocol (LSP) or RSVP-TE (adaptation of RSVP) IP routing protocol (e.g. OSPF, BGP, etc.)

Lambda Switching modulated lightflow wavelength of lightflow different proposals (GMPLS with LSP or RSVP-TE, OBGP, etc.)

Optical space switching all light within one fiber physical fiber port different proposals

Created: November 2002; last revision: October 14, 2004

Technology	Unit of data processed	Channel characterization	method for updating forwarding tables in the nodes of the network	method for determining the end-to-end routes realized by the forwarding tables
Time-division multiplexing	block	time unit in which block is received / sent	SS7 signaling protocol
ATM switching	cell (56 octets)	virtual path and virtual circuit identifiers in cell header	PNNI signaling protocol	PNNI routing protocol
MPLS	packet	value of the label added to the packet	Label Distribution Protocol (LSP) or RSVP-TE (adaptation of RSVP)	IP routing protocol (e.g. OSPF, BGP, etc.)
Lambda Switching	modulated lightflow	wavelength of lightflow	different proposals (GMPLS with LSP or RSVP-TE, OBGP, etc.)
Optical space switching	all light within one fiber	physical fiber port	different proposals