Network equipment hard core technology insider router Chapter 10 Cisco asr9900 disassembly (III)

It was disassembled for everyone yesterday ASR 9K Of NP Realization , that , When a packet enters ASR99 When it comes to routers , What kind of life journey will it go through ？

Please look at the chart below. .

In the figure , Into the direction PHY Can handle MACSEC、Retimer And so on . stay NP Find the second floor / Three layer forwarding table entry , perform ACL/QoS/PBR Strategy .

Specially , Port aggregation and ECMP It involves equivalent multiple exports , It needs to be optimized Hash Algorithm , Share data packets evenly across multiple streams .

The packet passed through the line card Switch Fabric Or backplane Switch Fabric, We reached the exit direction . Out of direction FIA And in the direction FIA Will cooperate to complete VoQ Credit control .

Pictured ,FIA The chip is moving Switch Fabric Before the chip sends data , You need to make a request first . obtain Switch Fabric After the arbitration is approved , The upstream FIA Downstream FIA Send data and consume the corresponding credit value (credit). It will be returned after downstream processing Credit.

that ,FIA What strategies are used to share the load ？

actually ,FIA It's through something like Round-Robin The way of load sharing .

Pictured , stay FIA Connect multiple Fabric ASIC Under the circumstances ,FIA To all Fabric The arbitration unit initiates a request , And to the first to respond Fabric Send packet .

FIA When sending packets , Use something similar BCM Of HiGig The way , Insert in the header metadata, This includes the sequence number of the packet . In the downstream FIA Reorder by serial number , To avoid disorder caused by parallel processing of data packets .

The sending method of multicast is not the same .

As shown in the figure above , Multicast load sharing is based on (S,G) Of the group , Every (S,G) The group will only be shunted to one Fabric ASIC On . This is to ensure the integrity of each multicast stream .（ Why? ？） Because the forwarding of multicast does not involve one flow going through multiple Switch Fabric, There is no need to send a request to the arbitration unit of the switching network .

Let's go back to the first figure again ：

We see , From the same NP Data packets in and out of the chip , You also need to bypass the line card Switch Fabric And backplane Switch Fabric, go back to NP Finished processing . Why is that ？

In fact, the problem is very simple .

If the same NP Incoming and outgoing packets , Only in NP Internal processing , that , There is no way to achieve credit based VoQ Flow control mechanism , When NP Upper interface A、 Interface B At the same time, the interface C It's a fight , And cause the interface C Congestion , Interface A/B Packets sent to other non congested interfaces , It will also be discarded in a certain proportion .

therefore , Whether it's NP Inside or across NP Chip forwarding , All need to go through FIA And two levels Switch Fabric Handle .

When the packet reaches the outgoing direction NP after , Exit direction NP Go out to the second floor / Three level search 、 Exit direction ACL/PBR Processing, etc .

For layer 3 forwarding , You also need to find the destination in the adjacency table IP The corresponding layer-2 address and send . because Tomahawk and LightSpeed Integrated TM,QoS And other functions are also completed in the same chip .

ASR9900 Of QoS Very powerful . With a 3 level H-QoS Configuration as an example ：

L0 and L1 For port groups and ports . from L2 Start , Configure traffic shaping policy ;L3 Work on the three-tier sub interface ( be based on VLAN), Traffic shaping with different priorities can be configured , The discard strategy is WRED.

L4 Configurable based on CBQ Bandwidth guarantee / Bandwidth control /WRED Discard and other strategies .

Tomorrow, , We will explain in detail ASR9900 Distributed control plane mechanism , and QoS Detailed explanation of , We will share in the following topics .（ In fact, I haven't studied it thoroughly ）

Today's remaining problems ：

since ASR9900 Yes N+1 Redundant switching network board , And between the same line card NP It also requires interworking on the switching network board , Why design on online card Switch Fabric？

The answer to the question left over yesterday ：

Multiple packet processing pipelines process packets at the same time , It may cause packet disorder , What are the consequences of packet disorder ？

answer ： stay TCP Streaming , If the packets are out of order , The receiver receives the following packet first , When the previous packet is not received within the set time , Will trigger a fast retransmission ( Continuous reply 3 individual ACK), Accordingly , The sender congestion window increases , Reduce transmission speed , Finally, transmission efficiency will be affected .