Phantom Flow Control Method and Apparatus

This invention relates generally to asynchronous transfer mode (ATM) communication and data networks and, more particularly, to flow control in ATM communication networks.

A network is generally defined as a collection of nodes interconnected by means of one or more communication links. End systems connected to the network both send data to, and receive data from, other end systems connected to the network. When an end system is a sender of data, it is referred to as a "source" for that data. When an end system is a receiver of data, it is referred to as a "destination" for that data. Typically, end systems act as both sources and destinations, depending on whether they are sending or receiving data. When acting as a source, an end system typically sends data in the form of messages over a communication link. Messages can flow between communication links and end systems within the network through bridges or routers. Multiple communication links generally are interconnected through a series of bridges and routers in a network.

Each message comprises a sequence of bits. Typically, messages sent over a network are divided into smaller blocks of information called "packets". The flow of packets in the network is usually referred to as "traffic." The packets are sent over the network in accordance with a data transfer protocol which specifies the form of the data as it is being transferred from switch to switch.

Asynchronous transfer mode (ATM) is one of the general classes of digital switching protocols that are used to relay and route traffic through a network. Unlike other known data transfer protocols, such as X.25 or frame relay, ATM transfers data by means of short, fixed length units of information called "cells." In applications utilizing ATM, data packets at a source are first broken up into such fixed length cells, transmitted through the network, and then reassembled at a destination. ATM cells are 53 bytes long consisting of a 5-byte header and a 48-byte information field. The header contains information used by the network to relay the cell from one node to the next over a pre-established route through the network. User data is contained in the remaining 48 bytes of the cell.

The ATM protocol uses a concept of "virtual networking" (also called "virtual channeling") to pass traffic between two end systems in a network. In accordance with this concept, ATM cells always traverse the same path from source to destination, but that path is not reserved for one end system. The path for a particular cell is only set up through the network as that cell is traversing the network. Any time a given end system is not actually occupying a link, another end system is free to use that link.

Accordingly, an ATM connection exists only as sets of routing tables stored in each network node, switch, or other intermediate system along the virtual channel. The routing tables are based on a virtual circuit identifier (VCI) and virtual path identifier (VPI) contained in the cell header. When a virtual channel is established, each node is provided with a set of look-up tables. By accessing the look-up tables, the node first identifies an incoming cell by header address, then routes it through the node to the proper output port.

The cell is thus passed from node to node over the preselected channel. Two cells that are ultimately bound for different destinations may be carried, one after the other, over the same physical medium (e.g., wire) for a common portion of their travel through the network.

With ATM, the successful transmission of packets from source to destination requires successful delivery of all the ATM cells formed from the original information packet. If any one cell is lost, the entire packet must be re-transmitted since the ATM network has no facilities for error correction. If cells are lost along the way, a sender must repeatedly resend the entire original information packet, resulting in an increasing load on the network, thereby creating more network congestion and more cell loss. As a result, ATM networks can operate efficiently only when there is minimal cell loss.

One way to avoid cell loss is to design each communication link and switch to handle the peak load expected through the network. This principle is the basis for Constant Bit Rate service (CBR) in ATM. This results in highly inefficient use of network resources, however, during the non-peak times. To ensure more efficient utilization, the ATM Forum has defined Available Bit Rate service (ABR), which is designed to maximize the utilization (e.g., bandwidth) of network resources while sharing the resources in a "fair" way among network users. The ABR Service is typically used for data traffic.

In order to fairly share network resources between users, a flow control mechanism is used to control the flow of traffic through the network. A common method of flow control (called "rate based" flow control) informs each data source of the maximum data rate at which it can transmit. Rate based flow control in ATM, as defined by the ATM Forum, requires that each data source periodically transmit a resource management (RM) cell. The RM cell loops over the virtual circuit and eventually returns to the data source. As the RM cell flows away from the source, it carries a value called the current cell rate of the data transfer session which informs the switches along the virtual circuit of the allowed cell rate of the session. Each switch along the route can modify another field in the RM cell called the "explicit rate" (ER), reducing it if congestion develops in that switch. When the RM cell returns to the source, the ER value is the maximum rate at which the source may transmit.

Rate based flow control is designed to provide "fairness" in sharing network resources. Specifically, the network bandwidth must be fairly shared among different users of the network. A widely accepted measure of fairness is called "max-min fairness." A bandwidth allocation by definition is max-min fair if, for every session, one session cannot increase its bandwidth without decreasing the bandwidth of sessions of equal or lower bandwidth.

One known method of implementing max-min fairness, known as "Phantom flow control" (Phantom), is a variant of rate based flow control. The Phantom method is an attempt at bounding the rate of each session (i.e., virtual channel) that share a link by the amount of unused bandwidth on that link. Accordingly, each link includes an additional imaginary session, referred to as the "Phantom" session. The Phantom method is described completely in an article (incorporated herein in its entirety, by reference) written by Afek et al., entitled "Phantom, A Simple and Effective Flow Control Scheme," Proc. SIGCOMM '96, Palo Alto, Calif., August 1996.

The Phantom method disclosed in the Afek article is used for rate based flow control of ATM cells passing through an output-buffered ATM switch. The Phantom method measures free bandwidth available on a link over a given time interval and computes a smoothed average by applying an exponential filter to the measured free bandwidth value to produce a value called the "Maximum Allowed Cell Rate" (MACR). The MACR is used to adjust the ER parameter in the RM cells, thereby controlling the maximum source cell rates. A number of variables are required to determine the MACR. In particular, MACR is a function of weighting parameters (alpha values) which are a function of the total number of cells in the output buffer queues ("queue length") relative to a threshold queue value ("queue-threshold"). Alpha values also vary depending on whether the MACR is increasing or decreasing.

FIG. 1 illustrates, in very generalized form, a data packet local area network involving four stations 100, 102, 106, and 108 (hereinafter 100-108), respectively, in which the Phantom flow control method can be utilized. In such a network, each station 100-108 contains software which breaks the packets up into ATM cells and affixes an appropriate header indicating how each cell should be routed through the network. Each station also contains software which receives ATM cells and reassembles the cells into the completed data stream.

Stations 100 and 102 are connected to an ATM switch 110 and stations 106 and 108 are connected to another ATM switch 114. Switches 110 and 114 route cells between the stations 100-108 and control access to the network. For example, station 100 may send cells over media line 116 to media input port 118 of switch 110. Switch 110 may route data cells to a destination station, for example station 106, based on the information in each cell header. To do this, cells are first switched by switch 110 onto output line 112 to switch 114. Switch 114 then routes the cells to station 106. Although only one output line 112 is shown, switch 110 may have several output lines.

Each switch, for example switch 100, acts as a cell relay. Cells that arrive from one of the incoming communication links are examined by the switching process which determines (from the routing codes in the cell header) which output line should be used. The cell is then added to a FIFO buffer queue for that output line which holds cells waiting to be transmitted on the associated outgoing communication link 112.

The Phantom method of computing and stamping the ER value of the RM cell for such a network is shown below.

One problem with the Phantom method is that it commonly drains queues very slowly during certain conditions such as, for example, during a transient overload. This often causes buffer overflow and cell loss, thus requiring retransmission of the signals from the source to the destination.

Accordingly, it would be desirable to have an apparatus and method of using the Phantom flow control method that limits such cell loss. It is among the general objects of the invention to provide such an apparatus and method.

In accordance with one aspect of the invention, the Phantom flow control method is modified to improve performance by facilitating queue draining. To that end, a portion of the available bandwidth is reserved to facilitate a faster queue buildup. This modification is necessary because when the number of virtual channels is large, the amount of free bandwidth available in the Phantom method for queue draining often is too small to effectively drain transient queues. This frequently causes buffer overflow and resulting data loss. The improvement effectively limits such problems.

Accordingly, it is among the objects of the invention that improves the performance of the Phantom flow control method;

It also is an object of the invention to improve queue draining in an ATM network; and

It is another object of the invention to provide an apparatus and method that reduces cell loss caused by cell buildups at a queue.

The foregoing and other objects and advantages of the invention will be appreciated more fully from the following further description thereof with reference to the accompanying drawings wherein:

FIG. 1 is block schematic diagram of a data packet switching network on which the inventive method and apparatus may be implemented;

FIG. 2 is a schematic illustration of a shared memory ATM switch that may be used with the invention;

FIG. 3 is a flow chart describing the improved Phantom flow control method;

FIG. 4 is a flow chart showing an exemplary embodiment of estimating free bandwidth, alpha, and the total of Q.sub.-- LENGTH;

FIG. 5 is a flow chart of an exemplary process for estimating a number of active flows;

FIG. 6 is a flow chart of an exemplary embodiment of the process of computing MACR;

FIG. 7 is a flow chart of a process for computing the smoothing parameters AINC and ADEC using Q.sub.-- TOTAL and NUM.sub.-- ACTIVE; and

FIG. 8 is a block diagram of a process for computing MACR using the modified alpha values.

In accordance with the invention, the Phantom flow control method is modified to improve its performance and to allow its use with less expensive ATM switches such as shared memory switches.

FIG. 2 shows an exemplary shared memory switch 200 having eight input ports 202, a shared memory 204, a bus 205 and eight output links 206. There are many possible implementations of a shared memory switch. Typically, when a cell is received at an input port 202, it is forwarded by means of the bus 205 to the shared memory 208, where it is queued until it becomes possible to forward this cell to the corresponding output port. Typically, each flow is allocated a separate queue 208 in the shared memory 208. A scheduling mechanism then determines the service order of the queues in the shared memory 208.

The computation of the allocation of MACR in the switch shown in FIG. 2 is performed for each output port 206 and for the bus 205. FIG. 3 describes the processes for computing MACR. The computation of MACR (process 302) is performed periodically. Process 302 requires knowledge of the amount of free bandwidth (a/k/a link capacity) for an output port (denoted "D") and the total number of cells queued (denoted "Q.sub.-- TOTAL"). The determination of these values is performed by process 300. In addition, process 302 requires knowledge of the number of active flows (a/k/a virtual channels) sharing an output port (denoted "NUM.sub.-- ACTIVE"), which is estimated by process 304. In general, processes 300, 302 and 304 can operate in parallel. No synchronization is required between these processes. However, process 302 is assumed to have access to the current estimates of the free bandwidth D, total queue length, Q.sub.-- TOTAL, and the number of active connections (NUM.sub.-- ACTIVE) computed by processes 300 and 304. This is symbolically shown by dofted lines in FIG. 3.

There are many ways of determining the free bandwidth of an output link and the total number of cells queued for each output port in the shared memory. An exemplary embodiment of process 300 is shown in FIG. 4. The computation of FIG. 4 is performed for each output port 206, and for the bus 205. At the initialization time, a time interval (Tfb) is chosen (step 400). The process 300 computes the number of cells received in time intervals of length Tfb. It does this by initializing the CELL.sub.-- COUNT variable to zero in block 404. The CELL.sub.-- COUNT variable is incremented any time a cell destined to the output has been added to the shared memory 204 in blocks 408 and 412. When the timer expires (in block 418), the free bandwidth D is computed in block 420 by subtracting the ration CELL.sub.-- COUNT /Tfb from the total bandwidth of the output link corresponding to the output port 206. After the new estimate of the variable D has been computed, CELL.sub.-- COUNT is reset to zero in block 404, the timer is reset in block 406 and the process repeats to obtain the new estimate of D. In addition to estimating the value of D, the process 300 maintains the total number of cells in all queues in the shared memory Q.sub.-- TOTAL corresponding to a given output port. At initialization, Q.sub.-- TOTAL is set to zero in block 402. Any time a cell destined to this output port is added to the shared memory, Q.sub.-- TOTAL is decremented by 1 (blocks 414,416). The values D and Q.sub.-- TOTAL computed inside block 300 of FIG. 3 are available at all times for the computation of block 302 in FIG. 3.

FIG. 5 shows an exemplary embodiment of process 304 of FIG. 3. This is just one of many embodiments for estimation of the number of active flows (i.e. virtual channels). This embodiment assumes that a bit "NA(I)" be maintained for each flow "I". At initialization (block 500) an interval Tna for estimating the number of active flows (NUM.sub.-- ACTIVE) is chosen, and NUM.sub.-- ACTIVE is set to zero. In block 502 the variable CELL.sub.-- COUNT.sub.-- NA is initialized to zero, and in block 504 all bits NA(I) are set to zero as well. The timer is started in block 506. When a cell of flow I is added to the shared memory for the first time after the timer has been started, the process sets NA(I)=1 for this flow (block 512), and increments the variable CELL.sub.-- COUNT.sub.-- NA by one in block 514. For subsequent arrivals of the cells of flow I in the current estimation interval Tna the check in block 510 fails, and the variable CELL.sub.-- COUNT.sub.-- NA is not incremented. This ensures that each flow is counted only once during each estimation interval. At the expiration of the timer in block 516 the current value of CELL.sub.-- COUNT.sub.-- NA is saved in the variable NUM.sub.-- ACTIVE (BLOCK 518.sub.--, after which the CELL.sub.-- COUNT and all bits NA(I) are reset in blocks 502, 504, and the new estimation interval starts in block 506. The value NUM.sub.-- ACTIVE computed inside block 304 of FIG. 3 is available at all times for the computation of block 302 in FIG. 3.

FIG. 6 shows an exemplary embodiment of process 302 of FIG. 3. In block 600 the interval for MACR computation is chosen, and the variables margin and floor are initialized. The timer is started in block 602. The computation of MACR in blocks 606-614 is performed upon the expiration of the timer, and the timer is restarted again. In block 606 the values of the necessary variables computed independently by processes 300 and 304 are obtained. In block 608, the values of the smoothing parameters AINC and ADEC (the alpha values) are computed using the values Q.sub.-- TOTAL and NUM.sub.-- ACTIVE from block 606. The details of process 608 are shown in FIG. 7. In block 610 the estimate of the free bandwidth (D) is adjusted by subtracting (i.e., reserving) a fixed ratio of the total bandwidth of the output link (margin) from the estimate of free bandwidth obtained in block 606. Typically the value of margin is chosen to be 0.1, which corresponds to the target reservation of at least 10% of the total bandwidth for queue draining. This adjustment is necessary to provide sufficient free bandwidth on the link for draining transient queues and thus, is an improvement of the Phantom method by limiting extremely slow queue draining, buffer overflow and cell loss commonly produced by Phantom. In block 612, MACR is computed using the equations of the Phantom method, except the values of AINC, ADEC are chosen in blocks 606-608 differently from the Phantom method, and the adjusted free bandwidth DADJ computed in block 608 is used instead of the variable D. FIG. 8 shows the details of block 610. The value of MACR is further adjusted in block 614 to disallow the value of MACR to fall below a certain value given by variable floor. Typically it is chosen to be at least as large as a certain fraction of total output link bandwidth divided by the maximum number of flows supported by the output port. The purpose of the adjustment of MACR in block 614 is to prevent very low MACR values which may result from transient overload. Since the rate-based flow control standardized for the ABR service in the ATM networks requires that the rate of sending RM cells be proportional to the flow's rate, very low MACR values result in very low flow rates for long periods of time. In turn, this results in the increase of the switch allocation, triggering the onset of severe oscillations.

FIG. 7 describes the details of block 608 of FIG. 6. In block 700 the value of the queue threshold, used by the Phantom method for determination of the alpha values in Table 1, is adjusted by adding a small number of cells ("A") per each active flow to this threshold. The value of the parameter "A" is typically chosen to be one or two. This adjustment is necessary because the value of the threshold is used to determine the degree of congestion. Essentially, the degree of congestion is quantified by the ratio of the threshold and the actual number of cells queued for the output. If the number of active connections is large, arrival of one cell per flow in a short interval of time (which can happen frequently due to synchronization frequently occurring in real networks), may result in the number of cells queued to be significantly larger than the fixed threshold value of the Phantom method. This may trigger the method to behave as if the link were severely congested even if in reality each flow is transmitting data at its correct fair share, onsetting potentially severe oscillations and instability. In block 702, the alpha values AINC and ADEC are computed according to Table 1, except the adjusted queue threshold is used instead of the fixed threshold.

In block 704 the values of AINC and ADEC are further adjusted by dividing the values obtained in block 702 by the number of active connections plus one. This adjustment is needed for the following reason. The smoothing parameters AINC and ADEC in the Phantom scheme defines the amount that the value of MACR is allowed to change in a single computation. A single change in the value of MACR for some port results in the change of the rates of all flows bottlenecked at the link corresponding to that port. If the number of flows is large, even a small increase in the value of MACR may result in a very large cumulative change of the input rate of all flows. In turn, this may cause a large transient queue growth, and the resulting drop in MACR. This frequently results in severe oscillations and instability of the Phantom method. This invention addresses this problem by making the values of the smoothing parameters inversely proportional to the number of active flows. As a result, the magnitude of the cumulative rate changes due to a single change in MACR is no longer proportional to the number of flows sharing the bottleneck link.

FIG. 8 shows the details of block 612 for calculating MACR with the modified alpha values. As mentioned earlier, the computation of MACR described in this figure is the same as in the Phantom method, except the values of parameters AINC and ADEC are chosen in accordance with this invention, and the adjusted value DADJ is used instead of the value d used in the Phantom method.

The bus 205 in a shared memory switch may be an additional bottleneck. To account for this bottleneck, the same improved flow control methods as described in FIGS. 3-8 may be utilized in accordance with the invention. Specifically, the improved method considers the bus 205 to be an output link shared by all flows traversing the switch and calculates an MACR value for the bus 205. When a backward RM cell is changed, the ER field is reduced to be the smaller of the MACR for the bus 205 and the MACR for the output link.

Although various exemplary embodiments of the invention have been disclosed, it will be apparent to those skill in the art that various changes and modifications can be made which will achieve some of the advantages of the invention without departing from the spirit and scope of the invention. These and other obvious modifications are intended to be covered by the appended claims.