Patent Grant
8045471
This invention relates generally to packet transmission networks and more particularly to methods for controlling packet flow on such networks.
As is known in the art, one type of packet transmission network is a wireless transmission network. As is also known, Wireless Local Area Networks (WLAN) are becoming more and more popular nowadays due to their easy deployment and wide spread of WiFi interface cards. A Wi-Fi Alliance report finds that 1.2 million 802.11 chipsets will be produced in 2006. Parallel to technological development, a flurry of analytical studies appeared in communication literature. Experimental results and theoretical studies show that wireless networks may enter a saturation regime characterized by a highly suboptimal medium utilization. More specifically, standard rate adaptation mechanisms reduce transmission rates when multiple packet loss occurs. Yet if the packet loss is due to collision rather than bad channel (which is the working assumption for the rate adaptation mechanism) then the controller induces a higher probability of collision which snowballs in turn into an even lower throughput. Such a mechanism is used by the Automatic Rate Fallback (ARF) algorithm used in WLAN-II products from Lucent which assumes all packet loss are due to bad channel.
In experiments with voice over wireless LAN, saturation induces catastrophic failures of the access point and clients. Clients drop voice calls. These can occur as early as 10-12 handsets and as late as 16-18 handsets, depending on the configuration, channel quality, interference at the time of experimentation. In
In accordance with the present invention, method and apparatus are provided to inspect wireless traffic parameters and, based on such input, to drop packets or influence admission control to avoid wireless saturation. In particular, the document discloses specific methods, called mitigation methods, to control or shape traffic, in conjunction with information about when the wireless channel is pre saturated. The procedure can be applied either in an access point (AP) or a client.
As described in two other co-pending patent application, one entitled, “Method for Congestion Detection in Packet Transmission Networks” filed on the same date as this application, and assigned to the same assignee as the present application, assigned Ser. No. 12/206,069, the entire subject matter thereof being incorporated herein by reference, and the other entitled “Method and Apparatus for Estimating Collision Probability in a Wireless Network”, filed on the same date as this application, and assigned to the same assignee as the present application, assigned Ser. No. 12/206,059, the entire subject matter thereof being incorporated herein by reference, two statistical methods are described for detecting the potential for wireless saturation by inspecting traffic parameters or features: The first using high level features of the radio channel accessible in the station (i.e. access point); and the second estimating directly the collision probability for a given (or all) access categories based on fine grain channel usage statistics. In the first case features of access to the radio channel are, for example, time utilization of the channel, number of retries, and delay. They can be recorded for a large variety of wireless data transmission situations (including voice over WLAN, when the interest is in voice, i.e. access category 3). A classification machine can automatically separate the regimes of saturation and normal operation. The classification logic can be incorporated for on-line, real-time operation in the firmware of a wireless station. In the second case, fine grained statistics can include the number of slots generated during a transmission, the total number of deferrals, and the total number of unsuccessful transmissions.
This document describes mitigation solutions working in conjunction with a solution of saturation prediction. Mitigation intervenes once pre-saturation is detected, in order to influence admission control or determine which packets should be disregarded (dropped) in order to alleviate saturation problems.
In accordance with the invention, a method is provided for controlling packet flow in a packet transmission network. The method includes determining a parameter representative of packet congestion on the network, and adjusting a flow of packets onto the network in accordance with such parameter.
In one embodiment, the parameter is a measure of saturation level of the channel.
In one embodiment, saturation level is a function of packet arrival rate at a receiver on the channel and total packet throughput on the channel.
In one embodiment, the function is that if there is a set of decreases in the packet arrival rates at each receiver that produces an increase in the total throughput, the channel is at the saturation level of the channel.
In one embodiment, the adjusting is a function of the time history of the parameter and the time average of such parameter relative to predetermined threshold levels.
In one embodiment, the adjusting selects one of a plurality of states, transitions between the states being a function of the time history of the parameter and the time average of such parameter relative to predetermined threshold levels.
In one embodiment, the parameter is a function of at least one of: time delay between transmission starts of a station on the channel and termination of the previously transmitted packet from such station; the fraction of time the channel is busy with transmissions, regardless of the origin of the transmission, or whether packets were correctly transmitted and received; and, average number of packet transmission retries on the channel.
In one embodiment, a method is provided for controlling packet flow in a packet transmission network. The method includes: during a training mode, generating a mathematical relationship between the degree of packet congestion on the channel and a plurality of measurable features of the network over a plurality of network conditions; during a subsequent normal operating mode, periodically measuring the plurality of measurable features and applying the generated mathematical relationship to such periodically measured plurality of measurable features to determine actual degree of congestion on the channel; and adjusting a flow of packets onto the network in accordance with such parameter.
In one embodiment, the degree of congestion is saturation level of the channel.
In one embodiment, the parameter is a function of at least one of: time delay between transmission starts of a station on the channel and termination of the previously transmitted packet from such station; the fraction of time the channel is busy with transmissions, regardless of the origin of the transmission, or whether packets were correctly transmitted and received; and, average number of packet transmission retries on the channel.
The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.
a)-2(c) are used to define interference for three cases: The first case
Like reference symbols in the various drawings indicate like elements.
The method and apparatus in this invention is defined by the following three steps:
1. Track special statistics in real time on top of underlying IEEE 802.11 chipset, e.g. air time utilization, number of retries, throughput;
2. Classify state of AP based on a classifier trained and pretuned using extensive statistics from real experiments; and
3. Control a WiSAT state machine implemented in the AP (client) and mitigation policy (packet drop) in congested or saturated situations.
The definition and integration of these components is described below:
1. Classifier:
Parameters: Window size W (×100 ms), Learning rate α (between 0 and 1),
Coefficients a1,a2,a3, and ρ
Inputs: GoodTxPkts, DrpTxPkts RCC, CC, TotDelay, RTxPkt
Output: Functional r=r(t) defined as follows:
i) If GoodTxPkts(t+W, AC=0)=GoodTxPkts(t, AC=0) then:
a. If DrpTxPkts(t+W, AC=0)>DrpTxPkts(t, AC=0) then r=−1 (saturation). STOP.
b. Else r=1 (non-saturation). STOP
ii) Else:
a. Compute (adjust for overflow/rollover as needed):
b. Compute distance to classification separation surface at any time t. s(t) is also called the WiSAT function:
s(t)=a1TU(t)+a2TD(t)+a3RT(t)−ρ
c. Compute smoothed classifier output, or the smoothed WiSAT function, possibly for multiple values of the parameter alpha depending on state of system αs. State is intended to capture the temporal dimension of the evolution of the system. The formula below indicates that the computation will be different at different times (if the state changes, as is described below:
{tilde over (s)}(t+W)=(1−αs) {tilde over (s)}(t)+αs·s(t+W)
d. End.
where s, {tilde over (s)} ure the classifier and smoothed classifier outputs, respectively, which are measures of packet congestion and the parameter αs is the smoothing rate for computation of {tilde over (s)} depending on state s.
More particularly, referring now to
Here, the degree of packet congestion on the channel is saturation level of the channel where saturation level is a function of packet arrival rate at a receiver on the channel and total packet throughput on the channel. The function is that if there is a set of decreases in the packet arrival rates at each receiver that produces an increase in the total throughput, the channel is at the saturation level of the channel.
Here, for example, the measurable features of the network include at least one of: time delay between transmission starts of a station on the channel and terminations of the previously transmitted packet from such station; the fraction of time the channel is busy with transmissions, regardless of the origin of the transmission, or whether packets were correctly transmitted and received; and, average number of packet transmission retries on the channel.
Below, we consider a number of features essential in detecting saturation. Other features could also be used in our approach; however we consider that this is a sufficient set of features to result in a good saturation detector and illustrate our concept and an instantiation of the concept:
1. Medium Access Delay (MAD)
Different chipsets offer various statistics about the operation of the hardware. In our formulas we will consider instantiations corresponding to the parameters reported by the Atheros AR5212 chipset hardware (manufactured by Atheros Communications, Inc. 5480 Great American Parkway Santa Clara, Calif. 95054) that can be used to calculate channel usage. In other implementations/hardware, similar statistics can be found or computed. These are:
Medium Access Delay (MAD) is defined based on the following times:
Generally we know only t0 and t2, however the following formula can be used to compute t1:
t1(n)=max(t0(n),t2(n−1))
which reflects the fact that, transmission starts as soon as the transmission of the previous packet ended, provided the current packet arrived.
MAD is defined as the average of t241 over N packets:
2. Time Utilization (TU)
Time Utilization (TU) measures the fraction of time the medium is busy with wireless transmissions, regardless of the origin, or whether packets were correctly transmitted/received. Thus, if T1 and T2 denote the starting and the ending time, respectively, for an observation period (e.g. 5 seconds), and that the RRC clear channel counter is incremented when the medium is idle, then:
Alternatively, the time utilization can be computed from the TFC and RFC register counters, which indicate the number of cycles transmission and reception flags are active over time:
For practical implementations, we have to explore the most accurate alternative. For different chipsets, alternative flags may enable us to compute TU.
3. Average Number of Retries (R)
Average Number of Retries (R) represents the average of DataFailCnt over a number of N packets:
Here, ns-2 code and perl scripts were written to simulate behavior of the network topologies of interest, or to measure the parameters of interest from real experimentation. We considered one AP and several stations with non-perfect channel conditions and interference.
Simulation Outline
In simulation, we have controlled these factors as follows:
Simulation Results: Balanced Network with Several Levels of Interference and Frame Error Rates (FER)
Following parameters are varied in simulation to cover various possible conditions:
In simulation, interference is defined as in
The simulation aims at computing the following measures: Goodput GP, MAD, TU, R. For a fixed FER, interference level, and S/traffic type, we derived GP(N), TU(N).
Overall, the results below cover 420 simulations, for various values of the discussed parameters: N, FER, InfAr, InfDu. Below we give a complete description of all these parameters, and some of the quantities tracked in simulation.
The simulation topology used is:
Infrastructure mode
Considering now how operation points (before or after saturation) are positioned in the joint space TU vs. MAD vs. Rt, it is first noted that simulation obtains undefined values for MAD and Rt when the interference is too bad and no frame starts to transmit within the simulation time. This induces some outliers in the plots, which should be discarded. Their positions, however, are obvious.
The saturation detector is based on the following rule:
More particularly,
More particularly,
Congestion can be predicted based on the position of the operation point in the space of features: as load increases, the operation point moves in the direction of the saturation boundary. Consider for instance a saturation boundary given by the first crosses on the parameterized curves with increasing load. The separation surfaces in these projection subspaces are virtually invariant lines for the various experimental cases (noise and interference).
As indicated, the boundaries can be precisely computed using a classification approach: the training data for the classifier that separates the x and o regions can be created from a large number of simulations.
The next step here is to compute a simple, efficiently computable, formula using the features as inputs in order to predict saturation and test the classification power of the formula for various “unseen” cases. Or, as noted above, the next step is to generate a mathematical relationship between the degree of packet congestion on the channel and a plurality of measurable features of the network over a plurality of network conditions.
Note that the operating points are positioned on level curves with respect to load, interference or frame error rates, and the more these conditions worsen (i.e. increased FER, or increased load), the smaller the distance between the position of the operating point and the boundary of decision regarding saturation. The consequences are very important in determining that the system approaches saturation, and determining the cause of saturation: bad channel conditions or congestion or both.
Here, from the data, weights W and p are generated and presented as matrices, more precisely vertical vectors to be described below in Step 300.
As noted above, during a subsequent normal operating mode, the method periodically measures the plurality of measurable features and applies the generated mathematical relationship to such periodically measured plurality of measurable features to determine actual degree of congestion on the channel; and comparing the actual degree of congestion on the channel with a predetermined channel congestion threshold level.
Here, for example, a 2nd order Support Vector Machine based Classifier for the database obtained by the simulation described above is used. The database can be enlarged using a variety of experimental data to obtain an accurate classifier in general or under particular conditions.
The method next compares the actual degree of congestion on the channel with a predetermined channel congestion threshold level.
A relatively simple saturation detector has the following form, where the saturation state depends only on the instantaneous measurements at the present time; a more complex detector will be described later, where saturation is modeled using a finite state machine; in this case the state of saturation and the actions to be taken depend not only on the instantaneous measurements but also on the previous state, or previous measurements. The state machine will be called the WiSAT state machine:
s is a saturation parameter providing an indication of the degree of packet congestion; and
an instantiation of the coefficients W, E and p is, in this example, given by:
W=[−0.1528−0.96310.4933−0.2066−0.98021.3510−1.38150.1 5802.593.5]T, ρ=−7.2384ε=0
In practice, the offset parameter 8 can also be experimentally determined. This classifier was here obtained by applying cross-validation on our database of up to 4×15×7=420 examples (in fact there are fewer examples, due to the fact that some experiments have not defined or not-a-number results in the computation of features).
A random subset of approximately 70% was used for training. The following testing results were obtained:
On training database: 99.6% accuracy (241/242 correct classification)
On testing database: 98.97% accuracy (96/97 correct classification)
On entire database: 99.41% accuracy (337/339 correct classification)
Thus, from the above, sufficient relevant computable features for saturation detection have been defined and show feasibility of building a robust classifier based on these features. Under a large set of conditions (FER, interference, there exist invariant separation surfaces in the space of the features of interest, which can be used to robustly detect congestion conditions or proximity to such conditions. The approach presented above gives a heuristic but theoretically informed way of building an effective detector.
The method thus computes a simple, efficiently computable, formula based on the discriminating features as inputs. The formula represents the classification boundary of saturation vs. non-saturation. The saturation boundary is approached under either increasing frame errors or congestion conditions. Furthermore, one can discriminate between the two causes of saturation: bad channel conditions or congestion or both.
A typical instantiation is an implementation of the congestion detection formula in an access point.
It should be understood that the term periodically herein means either at regular or irregular intervals.
Having determined the degree of packet congestion, such degree of congestion being characterized by the saturation parameter, s, a method is used to control or shape packet traffic (herein referred to as a mitigation policy). If the degree of congestion is greater than the determined threshold in Step 400 (
As described in more detail below, once the degree of congestion is determined, herein above such degree of congestion being characterized by the saturation parameter, s, the mitigation policy represents action to be taken in each state of the WiSAT state machine, shown in
Without loss of generality, one can use other classifiers, such as a different order SVM or other linear or nonlinear machine learning algorithms. In practical instantiations, we used a first order SVM classifier due to the low computational requirements, as follows:
s(TU,MAD,RT)=[TU MAD RT]WT−ρ
where parameters W=[4.391995-0.005264-0.233946], ε=0.5 and ρ=0.288917 were learned from real-world experimental data.
A typical instantiation is an implementation of the congestion detection formula in an access point.
It should be understood that the term periodically herein means either at regular or irregular intervals.
Having determined the degree of packet congestion, such degree of congestion being characterized by the saturation parameter, s, a method is used to control or shape packet traffic (herein referred to as a mitigation policy).
If the degree of congestion is greater than the determined threshold in Step 400 (
A more complex saturation detector can be modeled using a finite state machine. In contrast to using instantaneous feature values as before to decide on saturation, the decision and actions to be taken depend also on the previous state, or previous measurements. The state machine is called the WiSAT state machine:
Referring now to
Particular statistics/thresholds/intervals for the classifier and smoothed classifier outputs
More precisely, the following components are necessary to define the state machine and the logic of the state machine:
Present state STATE (STATE=0 representing NonSAT, or non saturation; 1 representing PreSAT or pre-saturation, and 2 representing SAT or saturation)
Conditions Cij (i,j=0,1,2) define transitions between states of the following format:
Cij=(s(t)oijδij)rij({tilde over (s)}(t)õij{tilde over (δ)}ij)
Where:
s(t), {tilde over (s)}(t) are the WiSAT classifier and smoothed classifier outputs
oij, 5ij are relational parameters ≦ and > for s, {tilde over (s)} respectively
δij, {tilde over (δ)}ij are threshold parameters for s, {tilde over (s)} respectively
rif is one of the logical relational operators AND, OR
Examples (Note that unspecified conditions are defined such that all outgoing transition probabilities from one state add up to 1 and are mutually exclusive. Conditions Cu (i,j=0,1,2) are implemented as follows: (with 5 parameters renamed for simplicity of notation {tilde over (δ)}o={tilde over (δ)}01, δo=δ01, {tilde over (δ)}1={tilde over (δ)}12, {tilde over (δ)}02, ={tilde over (δ)}21 corresponding to above general names)
Mitigation policy represents action to be taken in each state of the WiSAT state machine.
Algorithm: Act on Drop Rate N
Parameters:
Max and Min Drop Rate N0, NSAT
Critical Threshold {tilde over (γ)}
Output: N
where N=−1 means no packet is dropped; otherwise, for N>=0, it means one packet is dropped out of every N consecutive packets; NSAT=2 for example.
Note: If saturation does occur, then the last action taken while in PreSAT state could be continuously taken onwards, until a change of state is dictated by the conditions C21/C20.
Other Algorithms. Act on MAC Queue Length L
Other algorithm schemes, in addition to scheme 1 above, can be used to achieve similar effects:
Scheme 2. Design virtual queue length, L. Any incoming packet when the virtual queue is full will be dropped
Scheme 3. For each client, drop its every (m+1) packet;
Scheme 4. For each client drop the incoming packet if R packets for same client already are in the queue;
For example, scheme 2 proposes to focus on controlling the MAC queue length:
Parameters:
Maximum/Minimum MAC Queue Length L0, LMIN
Critical Threshold γ
Output: L
3. Compendium of WiSAT Parameters
We recommend a parameterized implementation of WiSAT in order to be able to tune parameters of two types, for implementation of the state machine:
Implementation of States
State memory and management s
State smoothing rates α0, α1, α2,
s(t)
{tilde over (s)}(t)
N(t) and parameters for computing N: N0, NSAT)β2, γ
Implementation of transitions for each state transition (i,j)
Oij, Õij relational parameters ≦ and >for s(t), {tilde over (s)}(t) respectively
δij, {tilde over (δ)}ijij threshold parameters for s(t), {tilde over (s)}(t) respectively
rij logical relational operator AND, OR
For example, we have performed simulation of the WiSAT and mitigation policies (however with only with simulated decision of mitigation, without really intervening in the control loop to drop any packets) on the example state machine before, with the following parameters:
WiSAT classifier learned using a linear SVM, with features aggregated over periods of 10 readings of 100 msec, i.e. over windows of 1 sec;
WiSAT classifier is able to take a decision every 100 msec: therefore the window (time step) W for WiSTATS and decision making is 100 msec;
WiSAT classifier function: α=α0=α1=α2=0.1;
State transition conditions (logic) Cij (i,j=0,1,2) given in the example
Classifier distance (WiSAT function) parameters for the state transitions:
{tilde over (δ)}0, δ0, {tilde over (δ)}1, {tilde over (δ)}2, ε given in the example have the following values:
{tilde over (δ)}0=2.1, δ0=2.1, {tilde over (δ)}1, =2.0, {tilde over (δ)}2, =1.0, ε=0.5
N algorithm parameters N0=45, NSAT=2, β2=43/({tilde over (δ)}1,−2.8), γ={tilde over (δ)}1
A number of embodiments of the invention have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. Accordingly, other embodiments are within the scope of the following claims.
This application claims priority from U.S. Provisional application No. 61/076,742 filed Jun. 30, 2008, the entire subject matter thereof being incorporated herein by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6646985 | Park et al. | Nov 2003 | B1 |
| 20020186657 | Jain et al. | Dec 2002 | A1 |
| 20040071086 | Haumont et al. | Apr 2004 | A1 |
| 20040090917 | Ruutu et al. | May 2004 | A1 |
| 20050175024 | Ho et al. | Aug 2005 | A1 |
| 20050237929 | Leith et al. | Oct 2005 | A1 |
| 20090141650 | Balan et al. | Jun 2009 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20090323528 A1 | Dec 2009 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61076742 | Jun 2008 | US |
