This application incorporates by reference a patent application filed on the same day as the present application, Ser. No. 10/955,893 and entitled Method and Apparatus for Preventing Head of Line Blocking among Ethernet Switches.
The present invention relates to networking, and more specifically to traffic flow control in Ethernet networks.
In the line card 104, CPUs 1042 and 1043 communicate with each other via a network interface 1045, the switch 1041, and a network interface 1044. In the line card 105, CPUs 1052 and 1053 communicate with each other via a network interface 1055, the switch 1051, and a network interface 1054. In the line card 106, CPUs 1062 and 1063 communicate with each other via a network interface 1065, the switch 1061, and a network interface 1064. A CPU and a network interface may be connected over a bus (e.g. a PCI Express bus), while other lines in the system are Ethernet connections.
It should be noted that the network interface functionality within blocks 1044, 1045, 1054, 1055, 1064 and 1065 may be implemented in any number of ways, whether as a chip, a portion of a chip, a card, or a portion of a card.
An Ethernet switch has information about its own ports, so that the switch can receive a packet and switch it over to the right port by examining the content of the packet and component information inside the switch.
A traffic flow may, for example, proceed from the CPU 1063 in the line card 106 to the CPU 1053 in the line card 105 via the switch 1061, the backplane switches 101 and 102, and the switch 1051. Other traffic flow may proceed from the CPU 1052 in the line card 105 to the CPU 1053 in the same line card via the switch 1051. If these two traffic flows try to exit the same egress port of the switch 1051, congestion can occur.
In the conventional Ethernet system, information passed between the network interface 1054 and the switch 1051 is traffic flow only. There is no information exchanged between the conventional switches indicating that there is congestion on a port or a specific receive queue of the network interface, and that certain packets are going to be dropped by the network interface because of the congestion. If there is congestion, a switch usually would just drop the packets. The problem of recovering the packet drops is then handled by higher level software running on both sides of the network, i.e., the transmitter and receiver, which detect dropped frames and request retransmission. The protocol that is usually used for this purpose is TCP/IP. The only standard way of avoiding drops would be to employ IEEE 802.3x flow control. However, that flow control causes blocking in the network. As a result, the slowest link would degrade the performance of the entire network.
Usually, a switch uses several priority queues on the ingress side of a network interface, employing a classification mechanism to decide how to classify packets on the link and which priority queue a packet should go to. The packet is then received by the network interface, which employs an independent classification mechanism in assigning the packets to a certain queue inside the CPU memory. The CPU provides the network interface with resources in the CPU memory. The network interface usually supports several DMA queues that take the packets received from the network, classify them into receiving DMA queues and put them in the CPU memory. Each DMA queue is serviced by the CPU with a certain amount of buffer memory which is managed dynamically by the CPU and the DMA as packets are being received and consumed by the CPU. The CPU allocates CPU time between the queues according to a predetermined policy. For example, queues of control data may have high priority, and thus other priority queues may get congested and their receiving (RX) DMAs will run out of buffer capacity, and will be forced to drop packets that keep coming from the network (i.e. from the switch). The switch does not know what the network interface and the CPU are going to do with the traffic flow from the switch.
For example, the switch 1051 has two input traffic flows: the first one is the one from the CPU 1063, and the second one is the one from the CPU 1052. As an example, the switch 1051 may send to a destination, e.g., the CPU 1053, a flow of data comprising 50% of the first traffic flow, and 50% of the second traffic flow under certain circumstances.
The destination of packets has an internal queuing mechanism. For example, there are two queues from the network interface 1054 to the CPU 1053: the first queue for the first traffic flow and the second queue for the second traffic flow. If the network interface 1054 then detects that the first queue is already filled up, the CPU 1053 cannot serve the first queue. The network interface 1054 then drops the next packet to the first queue.
In this case, the link between the switch 1051 and the network interface 1054 is used inefficiently because the switch does not know the status of the network interface queue. The switch 1051 continues to send 50% of the first traffic flow, although the network interface 1054 will just drop the packets anyway. At the same time, although the CPU 1053 can serve the second queue, the switch 1051 only sends 50% of the second traffic flow.
However, if the switch 1051 had known about the congestion, it could have sent more packets from the CPU 1052, and fewer packets from the CPU 1063. In addition, if the switch 1051 had informed the switch 1061 about the congestion, the switch 1061 could have employed a packet discard mechanism to remove the packets from the CPU 1063 at the outset, thus reducing the load on the entire switching system, and allowing traffic flow from the CPU 1052 to pass through with higher bandwidth.
However, conventional network interfaces do not communicate with their attached switches about queue status of the network interfaces. In addition, conventional Ethernet switches that are connected via standard Ethernet ports to each other do not communicate congestion information over the Ethernet link. The only such known mechanism is the disadvantageous 802.3x flow control mechanism. The prior solution has been to use a separate link 110 to communicate congestion information. However, that information had no relation to priority queues.
Therefore, it would be desirable to provide a method and apparatus for communicating the queue status of a network interface to its attached switch, and for communicating the queue status between switches.
One object of the present invention is to provide a method for preventing head of line blocking in an Ethernet system, using Ethernet protocol and media to communicate congestion information. In one embodiment, a network interface detects whether there is traffic flow congestion between the network interface and a data processing unit to which the network interface is connected, such as a CPU on the line card, a CPU connected to the line card, or a peripheral. If yes, the network interface communicates the congestion status to its attached Ethernet switch. The Ethernet switch then stops serving the congested port or queue.
Another object of the present invention is to provide a method for reducing a load on an Ethernet system. When a network interface communicates its congestion status to its attached Ethernet switch, the attached Ethernet switch informs a switch from which the traffic flow causing the congestion originates. The originating switch then reduces bandwidth for the traffic flow causing the congestion.
Another object of the present invention is to provide an Ethernet network interface, which can detect traffic flow congestion between the network interface and a data processing unit to which the network interface is connected, such as a CPU on the line card, a CPU connected to the line card, or a peripheral, and communicates the congestion status to its attached switch.
A further object of the present invention is to provide an Ethernet switch in which, when an Ethernet network interface communicates its congestion status to the Ethernet switch, the Ethernet switch can reduce bandwidth for traffic flow causing the congestion.
A further object of the present invention is to provide an Ethernet system, comprising a network interface which can detect traffic flow congestion between the/network interface and a data processing unit to which the network interface is connected, such as a CPU on the line card, a CPU connected to the line card, or other peripheral, and communicates the congestion status to its attached switch; and an Ethernet switch which can reduce bandwidth for traffic flow causing the congestion when receiving the congestion status from a network interface.
Embodiments of the present invention are described herein with reference to the accompanying drawings, similar reference numbers being used to indicate functionally similar elements.
Embodiments of the present invention are described herein with reference to the accompanying drawings, similar reference numbers being used to indicate functionally similar elements.
At step 201 in
In one embodiment, the network interface 1054 communicates to its attached switch the network interface queue status. The network interface 1054 uses IEEE 802.3x compliant packets to reflect its queue status. A frame compliant with the IEEE 802.3x standard is shown in
IEEE 802.3x defines reserved bits in the filler. In one embodiment of the present invention, the filler is used to send congestion information about priority queue or priority port. For example, the network interface 1054 sends 8 bits to a port of the switch 1051, where the first bit corresponds to a priority queue 0 and the last bit corresponds to a priority queue 7. Each bit represents status of one queue of the network interface at a moment. A value 0 means “do not send any more” and a value 1 means “continue to send”. In another example, the bits correspond to priority port status.
Usually, an 802.3x compliant packet only defines two states, on and off, to control data flow. In one implementation, a timer with a value that indicates how much more data can be sent by a link partner is employed. A typical usage is binary indication of transmission allowance. However, as discussed above, in one embodiment of the present invention, the filler of a 802.3x compliant packet is filled with a bitmap, indicating status of priority queue or priority port. This can be extended to hold one timer per class as well. For example, the filler could contain multiple timers, one per class of traffic.
The network interface 1054 sends to the switch 1051 its queue status periodically, so that the traffic flows to these queues continue or stop for a certain period of time, until a new 802.3x packet with an updated bitmap is received. Thus, the 802.3x compliant packet of the present invention does not just inform the switch to send packets or not. Instead, it provides the switch with status of the network interface's priority queues on receiving, so that the switch can distribute its resource accordingly: reducing packets to congested queues, and increasing packets to other queues.
It should be understood that the network interface could inform its attached switch about the congestion in other ways, which the ordinarily skilled artisan will understand.
At step 203, the switch 1051 stops serving the congested queue. The congested queues start to fill up in the switch, and then depending on the switch, could eventually be dropped.
In one embodiment, the present invention ends with step 203. In another embodiment, to further reduce the load on the Ethernet system, the originating switch of the congested queue could be informed to stop sending traffic flows. To do so, at step 204, the switch 1051 informs the originating switch of the congested queues, e.g., the switch 1061, about the congestion. The switch to switch notification could be done over a proprietary line, as shown in
At step 205, the originating switch reduces the bandwidth of the traffic flow going to the congested queue.
According to yet another embodiment of the invention, at step 206, the originating switch could also increase the bandwidth of other traffic flows, to make better use of the switch's available bandwidth.
It should be understood that obtaining the queue status information from the multi-channel transmitting DMA 403 has a similar effect. In addition, a skilled artisan would appreciate that instead of queue status, the network interface could monitor its port status and communicate the port congestion status to the switch for the traffic flow control via a bitmap in an IEEE 802.3x compliant frame.
The management module 504 could program mapping of the switch queues to the network interface queues. From the 802.3x frame from the network interface, the management module 504 decides the affected switch queues, and reduces bandwidths for these affected queues accordingly. In one embodiment, the management module 504 also could increase bandwidth for other queues.
While the invention has been described in detail above with reference to some embodiments, variations within the scope and spirit of the invention will be apparent to those of ordinary skill in the art. Thus, the invention should be considered as limited only by the scope of the appended claims.
Number | Name | Date | Kind |
---|---|---|---|
5726976 | Thompson et al. | Mar 1998 | A |
5790891 | Solt et al. | Aug 1998 | A |
5809557 | Shemla et al. | Sep 1998 | A |
5841722 | Willenz | Nov 1998 | A |
5913042 | Shemla et al. | Jun 1999 | A |
5923660 | Shemla et al. | Jul 1999 | A |
5930261 | Shemla et al. | Jul 1999 | A |
5999981 | Willenz et al. | Dec 1999 | A |
6240065 | Medina et al. | May 2001 | B1 |
6327615 | Kasper | Dec 2001 | B1 |
6601116 | Shemla et al. | Jul 2003 | B1 |
6628613 | Joung et al. | Sep 2003 | B1 |
6678278 | Medina et al. | Jan 2004 | B2 |
6738838 | Shemla et al. | May 2004 | B2 |
6980520 | Erimli | Dec 2005 | B1 |
7002911 | Linville et al. | Feb 2006 | B1 |
7212534 | Kadambi et al. | May 2007 | B2 |
7593328 | Yang et al. | Sep 2009 | B2 |