Creating a low bandwidth channel within a high bandwidth packet stream

0001The present invention pertains to the art of packet-switched digital networks, in the preferred embodiments to injecting a low-bandwidth undirectional data stream into a high-bandwidth channel. In many applications involving digital networks, there is a need for channels for management, monitoring, and/or measurement functions. In these functions, it is common to have a device connected to a high-bandwidth channel. The device performs some function, producing a low-bandwidth data stream as a result. Handling that low-bandwidth data stream requires that the device be connected to another communications channel, such as a wireless link, or a port on a high-speed switch. In devices such as switches and routers which have built-in measurement and management capabilities, additional resources are dedicated to providing communications capability to these functions. In either case, additional resources are tied up in the process of placing the low-bandwidth data stream back into the network. WO 01/56195 discloses a digital status channel for a broadband communication system. The present invention seeks to provide an improved channel operation. According to an aspect of the present invention there is provided apparatus for creating a low-bandwidth channel in a high-bandwidth channel as specified in claim 1. According to another aspect of the present invention there is provided apparatus for creating a low-bandwidth channel in a high-bandwidth channel as specified in claim 8. In the preferred embodiments, a low-bandwidth channel is created in a high-bandwidth channel such that extra bandwidth is only used for the low-bandwidth channel when there is data to be sent, minimal latency is introduced in the high-bandwidth channel when there is no packet to be sent over the low-bandwidth channel, and the effects of sending a packet on the low-bandwidth channel are absorbed and distributed amongst other passing traffic.

0002Embodiments of the present invention are described below, by way of example only, with reference to the accompanying drawings, in which: <ul id="ul0001" list-style="none"><li>Figure 1 shows packet insertion,</li><li>Figure 2 shows a block diagram of a first embodiment of the invention,</li><li>Figure 3 shows FIFO pointers for the first embodiment of the invention,</li><li>Figure 4 shows a state machine suitable for the first embodiment of the invention,</li><li>Figure 5 shows state transitions in the first embodiment of the invention,</li><li>Figure 6 shows a block diagram of a second embodiment of the invention,</li><li>Figure 7 shows a state machine suitable for the second embodiment of the invention,</li><li>Figure 8 shows state transitions for the second embodiment of the invention,</li><li>Figure 9 shows a block diagram of a third embodiment of the invention,</li><li>Figure 10 shows state transitions for the third embodiment of the invention, and</li><li>Figure 11 shows a state machine suitable for the third embodiment of the invention.</li></ul>

0003The preferred embodiment creates a low bandwidth channel by inserting packets into the high bandwidth packet stream. A low-bandwidth channel is created within the high-bandwidth channel by inserting packets at a predetermined interval. This insertion introduces latency into the high-bandwidth channel. This latency is not constant, but is recovered by minimizing inter-packet gaps between packets in the incoming high-bandwidth channel following the inserted packet. While the inserted packet is being transmitted, forming the low-bandwidth channel, arriving high-bandwidth packets are stored in an elastic buffer.

0004Considering the analogy of entering highway traffic, incoming traffic joins highway traffic by "squeezing in" (merging). If there is enough space between cars on the highway, there is no problem. However, during rush hour, joining traffic becomes more difficult. When we enter heavy traffic on the highway, cars that will end up behind us often must slow down to make room. This may however, not be noticeable for cars further behind. Traffic will compensate for entry by slowing down and making gaps between cars smaller because of lower speeds. In the old days, when there were no metering lights, such entries very often created havoc. By introducing metering lights, the situation improved in the sense that cars entered the highway at a predetermine frequency. This eased the impact of "injecting" cars into existing traffic. The system relies on sufficient under-utilized link capacity as packets are injected at a reasonably low rate.

0005Various Ethernet network studies have demonstrated that Ethernet networks work well if the bandwidth is utilized at no more than 30% of capacity. Heavier usage may cause collisions that lead to congestion. Retransmission may create more traffic and more collisions to the point that there will be very little traffic that makes it through. For backbone networks or inter-domain links, where there are fewer transmission originators, the traffic is much smoother, and the utilization of such links may reach 60-70% of capacity. There are fewer collisions because there are fewer parties competing at the switches where the traffic converges. Higher utilization and occasional traffic bursting may lead to substantial packets loss at the routers. While maximum packet size for basic Ethernet is 1536 bytes, typical traffic consists of a range of packet sizes. This means that a router's packet switching capabilities must match traffic patterns as well as link bandwidth. In other words, driving a communication line close to 100% of a router's capability, either in packets per second or link capacity, will result in substantial packet loss. The implications of packet loss are very severe. Packet loss may have an avalanche effect and may create even more traffic due to retransmissions. Therefore, most network planners leave a little slack, extra capacity, rather than running communication lines at 100% capacity. Some of this extra capacity is available for us as a low bandwidth channel as this invention proposes. This low bandwidth channel may use just 1% of the communication line total bandwidth.

0006The described embodiment defines the upper bandwidth of the low bandwidth channel by specifying a hold timer. The hold timer defines an interval during which only one packet may be sent. When the hold timer expires, another packet may be sent over the low bandwidth channel. This may only occur under the condition where the high bandwidth packet stream absorbed any previous packet that was inserted. In other words the high bandwidth packet stream must return to moving unaltered before an extra packet may be inserted into the stream. Of course, the extra packet insertion would cause small, temporary delays in the high bandwidth packet stream. The hold timer delay in clock cycles (characters; each transmitted character takes up one clock cycle) is calculated as follows: <maths id="math0001" num=""><math display="block"><mi>hold_time</mi><mo>=</mo><mi>packet_in_bytes</mi><mo>×</mo><mfrac><mi>high_bandwidth_channel_speed</mi><mi>low_bandwidth_channel_speed</mi></mfrac></math><img file="EP1524807B1_D0001.tif" /></maths>

0007For example, if the inserted packet size is 1.5 Kbytes, the high bandwidth channel bandwidth is 1 Gbps, and the low bandwidth channel is 1% of the high bandwidth channel capacity, i.e., 10 Mbps, then the hold timer is approximately 150,000 clock cycles. The clock cycle per byte for 1Gbps and with 8B/10B encoding of the bit stream is 10 nsec. This means, in this example, that no other packet could be introduced to the high bandwidth packet stream in less than 1.5 msec. Smaller packet sizes will have smaller hold timer values and could be introduced more often. However, choosing smaller packets may increase the number of packets per second which may affect the router's processing capability if it is near its packet-per-second limit.

0008Instead of defining the hold timer as a fixed function of packet size and bandwidth, the hold timer could have a random factor built in. In this case, the above calculated hold timer could be an average value that randomly fluctuates between a lower and upper bound. For example, the hold timer could be within +-20% of an average. Randomization of the hold timer may prevent synchronization of traffic flow when the same size packets are inserted into the traffic. It is not proven that randomization of this hold timer may affect flow synchronization but various studies of traffic flows have shown that synchronization in general may have an adverse effect on the overall stability of the networks.

0009Another important parameter is the inter-packet gap. This is a gap between two consecutive packets or frames (e.g., Ethernet frames). In 1Gbps Ethernet, the minimum inter-packet gap must be 96 nsec. The only way to reduce the impact of insertion of an extra packet (actually a packet that is framed) into the high bandwidth stream is by reducing inter-packet gaps of following packets that are greater than the minimum. Of course, the minimum gap could be defined as larger than a minimum gap defined by a specific standard. This may be necessary if the line card has problems handling minimum gaps defined by the standard. Similarly to the hold timer, the minimum gap could also have a random factor built in to avoid possible harmful traffic synchronization. But a study has to be conducted to determine if that is necessary.

0010Figure 1 illustrates how extra packets (frames) 110 and 112 are inserted into a high bandwidth packet stream 100. It should be repeated here that a "packet" is not just an IP packet but it is a frame that may contain an IP packet. In the case of Ethernet, it is an actual Ethernet frame that is already encoded e.g., using 8B/10B encoding. The type of encoding will depend on what type of encoding is used by passing traffic. In Figure 1, unaffected packets are shown in clear, delayed packets in gray, and inserted extra packets in black.

0011At to extra packet 110 is ready to be sent. By this time, the predetermined hold timer has expired. This means that the next opportunity to insert an extra packet will be when the high bandwidth channel will be transmitting IDLE characters. At t₀, packet P0 was being transmitted. After completion of transmitting P0 and minimum gap at t₁, new packet EP1 could be inserted. At the same time, the invention must absorb incoming traffic in an elastic buffer for future retransmission. In the example in Figure 1., the incoming traffic between P0 and P1 has an inter-packet gap greater than the minimum. This embodiment will minimize this gap and drop extra IDLE characters until it sees packet P1. The P1 packet will enter the elastic buffer, as well as the minimum gap between P1 and P2 when the extra packet is in the process of being sent. Also, the gap between P2 and P3 will be reduced so that the P3 packet will move without delay. Only P1 and P2 are affected (delayed) by the insertion of the EP1 packet but P3 is not delayed. At t_1, the embodiment also starts a hold timer as defined earlier, shown as packet stream 120. Packet stream 130 illustrates a situation when a random factor is added to the hold timer. At t_2, another extra packet EP2 112 is ready to be sent. However, at this time, we have not yet met the low bandwidth channel criteria and must wait until t_3, when the hold timer expires. After t_3, the process repeats, i.e., at t₄ an extra packet is injected into the stream and so on. It should be noted here that the speed at which the new extra packets are absorbed would depend on the inter-packet gap sizes. Larger gaps will "absorb" new packets more quickly. Also it does not matter whether incoming traffic contains normal (1.5 Kbytes) or jumbo (9 Kbytes) Ethernet frames. However, bursted Ethernet frames will not be broken apart.

0012A first embodiment of the present invention is shown in Figures 2 through 5; it introduces zero delay. A second embodiment of the present invention is shown in Figures 6 through 8 and uses a two-character delay. A third embodiment of the present invention is shown in Figures 9 through 11 and uses an arbitrary delay.

0013While the present invention may be implemented in various forms, the present invention is also suitable for implementation in a highly integrated form suitable for replacing industry-standard interface converter modules, for example those known as GBICs, or GigaBit Interface Converters. Current GBICs are basically transceivers translating one media type (optical, twisted pair, etc.) to another media type. By providing a replacement GBIC including the present invention, numerous applications requiring low-bandwidth channels are enabled. Such applications include many network monitoring applications.

0014The zero delay solution of Figures 2 through 5 operates normally at line speed (no latency introduced to traffic passing by except latency introduced by serializer, mux and deserializer). As shown in Figure 2, Extra packets are injected into the traffic stream from buffer 260.

0015The two-character delay shown in Figures 6 through 8 introduces a minimum, two characters delay for normal operation (i.e., traffic passing by is delayed by two characters). As shown in Figure 6, the extra packet is copied into FIFO buffer 650 before transmission.

0016The arbitrary delay solution of Figures 9 through 11 is important for those types of applications that require manipulation of data before forwarding, e.g., updating packet headers, removing parts of headers, etc.

0017While descriptions are included here to these representative embodiments, the main focus is on the zero delay solution that is described below in detail.

0018Any time a serial bit stream is de-serialized into parallel n-bit words, the result can only be forwarded for further processing (even just to re-serialize) after all n bits arrived. This means that de-serializing and then re-serializing a bit stream introduces a latency of at least one word cycle or n bits. In practice the serializer and the deserializer each can be expected to have more than one bit times of internal latency, for a total of about two word cycles (20ns for the Gigabit Ethernet example), even in the case of what we call the "zero-delay" solution. In addition, the invention must read and write to on-chip memory and detect idle characters at the word rate, e.g. 125MHz for Gigabit Ethernet. This should not be a problem in a modem CMOS process. However, faster interfaces such as 10Gb Ethernet, may require a wider, multi-word parallel stream which would incur additional SER/DES latency of 1 word cycle (n bits) per word. Of course, in absolute terms those cycles would be 10x faster.

0019<b>Zero delay</b>. Figures 2 through 5 show a zero delay (latency) embodiment of the current invention. Referring to Figure 2, the incoming stream of packets comes in on interface 211. Because it is a serial stream of bits (for 1 Gigabit Ethernet, it runs an actual line rate of 1.25 Gbps) it is de-serialized by de-serializer 210 into parallel streams of bits. If 8B/10B encoding is used (e.g., 1 Gigabit/sec Ethernet) the stream will be de-serialized into a 10 parallel bit stream 212. That stream will go three different routes, depending the system state.

0020De-serialization is a well-known technique to process high-speed data at lower speeds. For example in 1 Gigabit Ethernet data can be processed at 125 MHz instead of 1.25 GHz. If there is no other packet being sent or FIFO buffer 250 is empty (i.e., there is no effect of previously inserted packets on the main packet stream) the state of the control logic 230 is in the initial state - <b>S0</b> of Figure 4. The state machine described in Figure 4 shows how control logic 230 transitions from one state to another based on events. Figure 5 shows an example of a state transition. Figure 2 illustrates the basic building blocks and Figure 3 describes how FIFO buffer pointers transition.

0021It should be noted that while state machines are provided as an example and an aid to understanding the present invention, state machines are not necessary in practicing the invention. More compact embodiments may be obtained using fixed logic which implements the concepts described herein.

0022Similarly, while a FIFO is used as an example, any elastic buffer may be used. Hardware implementations may not require direct control of read and write pointers, for example, and single-port buffers may be used.

0023In state <b>S0,</b> the incoming packet stream data uses the "fast path". It moves from interface 212 to multiplexor MUX 270 through 271 interface. MUX 270 is in state 0 which allows the incoming data 271 to be forwarded through interface 273 and 221 to SERializer 220 which converts the parallel streams back into one serial stream 222 at the end. The output of SERializer 220 typically connects to outside network equipment, and may present an electrical or optical interface.

0024Referring to the example in Figure 5, at time t₀ an incoming packet <b>P0</b> is passing through using the fast path described above. Control logic 230 is in state <b>S0.</b> At t_1, the hold time expires (Figure 5) and control logic 230 moves from <b>S0</b> to <b>S1</b> state (event: <i>hold timer expired;</i> Figure 4,5). At t_2, an extra packet (Figure 5) is ready to be injected (event: <i>extra packet ready</i>) into the stream. The state machine of control logic 230 switches to state <b>S3</b> (Figure 4 &amp; 5). At this time the control logic 230 starts sensing when to inject the extra packet. The sensing is done by checking each individual character 231 passing through control logic 230.

0025The data can take 3 different paths after deserializer 212 of Figure 2.. The first path was described above as the fast path. The second path 251 goes into FIFO buffer 250. The third path 231 enters control logic 230. Control logic 230 senses IDLE characters indicating the possibility of injecting the extra packet. Control logic 230 counts IDLE characters to meet the minimum gap. This occurs in state <b>S3</b> shown in Figures 4,5. Once the minimum gap is met, an event: <i>current gap == minimum gap,</i> is generated and the state machine switches to state <b>S4</b> (insert packet into stream and start absorbing incoming traffic into FIFO). It should be noted here that in Gigabit Ethernet IDLE characters come from two different coding groups. In other word there are two different IDLEs. Which of the IDLEs to use will depend on DC signal balance of entire stream of sent characters. If for example an injected packet is followed by one type of IDLE and there are IDLEs in the FIFO that should follow immediately IDLEs of the inserted packet then those IDLEs should be replaced by the IDLE followed the inserted packet. This way proper DC signal balance will be preserved. In addition an IDLE is not a single character but two characters. At t₃ the extra packet 260 is being inserted. At the same time a new hold timer starts, the current gap counter is set to 0, FIFO 250 R (read) pointer is saved and a new R pointer points to the beginning of the extra packet (Figure 3).

0026At the same time, the incoming packet data is also forwarded 251 to FIFO 250 but is not forwarded through MUX 280 because the state of MUX 280 is 0 and interface 282 only allows forwarding data if the state of MUX 280 is 1. Because the R pointer points to the extra packet the W (write) pointer will not advance at this time. The data will be written into FIFO 250 to the same location but reading will occur from the extra packet buffer.

0027It should be noted here that control logic 230 would change state <b>S0</b> -> <b>S2</b> -> <b>S3</b> if event: <i>extra packet ready</i> occurs before event: <i>hold timer expired</i> (Figure 4). Once those two events occur, control logic 230 will switch to state <b>S4</b>.

0028Once control logic 230 enters state <b>S4</b>, it sends a signal through interface 232-274 to change MUX 270 state to 1 and signal 236-284 to set MUX 280 to state 0. At the same time, through signal 235-262 it starts a process of sending the extra packet 260 character by character. At t_3, the R pointer is pointing to the extra packet data. Extra packet characters will move through interface 261-281 to MUX 280 and then they will be forwarded through interface 283-272 to MUX 270. Because MUX 270 is in state 1, data goes through interface 273-221 to SERializer 220 and leaves thorough interface 222. At t₄ packet P1 arrives. An event: <i>START frame</i> will be generated and control logic 230 will transition to state <b>S5</b>. FIFO 250 starts accumulating P1 characters, i.e. its W pointer starts advancing. This means also that a gap t₃-t₄ was eliminated from the stream of packets to compensate for the effect of the extra packet insertion on the incoming packet stream delay. In the example shown in Figure 5, this gap (t₃-t₄) is not sufficient to completely absorb the effect of the newly inserted extra packet and more inter-packet gaps must be used to compensate the effect.

0029It should be noted here that the transition to the next state, e.g. from <b>S3</b> -> <b>S4</b> or <b>S5</b> -> <b>S4</b> will only occur if the configured minimum gap is met. In other words, for example a packet could not be inserted after P0 if the actual minimum gap (as defined by the standard) between P0 and P1 is less than configured minimum gap. The extra packet will have to wait until a large enough inter-packet gap will show up. This option of having a configurable minimum gap larger than those defined by a specific standard is left for those deployments where standard defined minimum gap puts too much stress on the receiving end. It is assumed that under normal circumstances the configurable minimum gap will be equal to that defined by the standard. For brevity in this document any time this invention is referring to minimum gap is referring to configurable minimum gap.

0030The extra packet must be trailed by at least the minimum gap. The gap between packets P0 and P1 is too small to accommodate the trailing gap for the inserted extra packet and therefore a minimum gap has to be included as part of the extra packet. The absorption of the effect of the insertion of extra packet (t₃-t₄) is done simply by FIFO 250 not advancing its own W pointer.

0031Once packet P1 is absorbed by FIFO 250, the FIFO will accept IDLE characters following P1 only until the minimum gap trailing P1 is met at t₅. At t₅ an event: <i>current gap == min gap</i> is generated and control logic (230) will transition back to state <b>S4</b> to eliminated extra IDLE characters. FIFO (250) will skip IDLE characters until the arrival of a new packet P2 (t₆). Again, here the gap t₅-t₆ is used to compensate the effect of the extra packet insertion. At t₆ a new packet P2 arrives and an event: <i>START frame</i> is generated. This event transitions control logic (230) state machine to state <b>S5</b> in which the FIFO is going to save incoming packets for future transmission.

0032At time t₇ the extra packet with minimum gap is finally sent. Control logic 230 switches through 236-284 interface MUX 280 to state 1 and sets R pointer to the beginning of FIFO 250 (Figure 3, State S6, S7). An event: <i>end of insert</i> is generated and the state machine of control logic 230 transition to state <b>S6.</b> At this time, data is being sent from FIFO 250. The FIFO's R pointer is set to a first character and this will be the first character of packet P1 saved in FIFO. In the mean time, FIFO 250 is accepting 251 and storing packet P2. At tg minimum gap trailing packet P2 is reached. An event: <i>current gap == minimum gap</i> is generated and state machine transition to state <b>S7</b> (skip IDLE characters while emptying FIFO).

0033At time t₉ the last character from FIFO 250 is sent. It should be noted this is not the actual last character saved in the FIFO. The last character means a character that could be read from FIFO. The R pointer is, at this time, two characters behind the W pointer. Because we were receiving IDLE characters after the minimum gap was reached, the W pointer has not been advancing. In other words, IDLE characters between t₈ and t₉ were intentionally dropped from the incoming packet stream. If at t_9, the last two received characters were representing IDLE then, in the next clock cycle (next character) it will be safe to switch to the fast path. The only thing, which could be lost from the incoming packet stream, is this IDLE (i.e., two characters representing one IDLE). And that is OK, it is actually what we want. In this case, events: <i>FIFO == 2</i> and <i>last char rcvd == IDLE</i> are generated and the control logic state machine goes back to state <b>S0</b> (fast path). Control logic 230 via interface 232-274 switches MUX 270 to state 0 and allows the packet stream to follow the fast path (212-271-273-221-222).

0034As shown packet P3 arrives after t₉. However, if by coincidence, the last received character was the start frame character of packet P3, then we have no choice but to stay in the FIFO path until the inter-packet gap is long enough (minimum gap plus two extra characters representing one IDLE) to allow us to switch to the fast path. This is a very crucial element of the invention. Switching from FIFO path to fast path can only occur when two characters representing an IDLE that were received but not yet sent can be dropped. And the only characters that can be dropped are IDLEs (two characters each), assuming of course, that the minimum gap was already sent. By its nature, the FIFO path is always at least two characters behind the fast path, i.e., R pointer follows W pointer by no less than two characters.

0035<b>Two-character delay</b>. A embodiment of the present invention using a two-character delay (size of an IDLE) is shown in figures 6 through 8. As shown in Figure 6, this embodiment uses register 640 in the path whose output 642 is fed either to FIFO 650 or leaves the logic through MUX 680 and SERializer 620. The difference between this embodiment and the zero-delay embodiment is twofold. In the fast path of this embodiment, register 640 introduces a permanent two-character delay and the extra packet is not injected into the packet stream from a separate memory but instead is copied to FIFO 650 first. This may simplify implementation of the memory controller and management of FIFO 650. Figures 6 through 8 illustrate the block diagram, suitable state machine, and example state transitions for this embodiment.

0036<b>Arbitrary delay solution</b>. An embodiment of the present invention using an arbitrary delay solution is a simplified version of zero delay without an option of zero delay. Figures 9 through 11 illustrate this embodiment, which is applicable for applications that require delays. For example, manipulation of IP headers or when entire packet will require withhold before releasing and modifying information in the packet header. It is, in concept, similar to store and forward techniques used by packet switches. Both the extra packet 240 and the packet stream 211 are stored in FIFO 250 and selectively routed to SERializer220 under control of a state machine according to Figure 11.