The ASR9000 is a Virtual Output Queue (VOQ) design with memory on
the input side of the line cards. Packets are pulled across the cross-bar
fabric when output capacity is available and sent along their way. VOQ
switch design prevents head of line blocking.
The hard work is done by NPU - Network Processor Units. A
slide
from CiscoLive! gives buffer sizes for 10 Gb/s line cards in mS for several configurations. The
slide is the source of the data in this table:
|
NPU speed | per port buffer | |||
First Generation | Trident -L, -B | 17M PPS | 15 Gb/s | 50 mS |
First Generation | Trident -E | 17M PPS | 15 Gb/s | 150 mS |
Second Generation | Typhoon SE | 44M PPS | 60 Gb/s | 339 mS x 2 ports/NPU |
Second Generation | Typhoon TR | 44M PPS | 60 Gb/s | 170 mS x 2 ports/NPU |
Second Generation | Typhoon SE | 44M PPS | 60 Gb/s | 226 mS x 3 ports/NPU |
Second Generation | Typhoon TR | 44M PPS | 60 Gb/s | 113 mS x 3 ports/NPU |
There are three 100 Gb/s line cards with a different story.
Trident is not related to the Broadcom device of the same name. First Gen NPUs can handle 10 Gb/s ports comfortably, so: |
Second Gen NPUs can handle up to 6 x 10 Gb/s although other components could become a bottleneck e.g. with only a single Route Swith Processor. Three is safe in all configs. So, for example: |
Q. What is the default queue buffer size?
A. That depends on the QoS configuration. Each network processing unit (NPU) has frame memory attached to it and this frame memory is where packet buffering is seen. The Trident -L and -B cards send a 50ms burst of traffic.The Trident -E card sends a maximum 150ms burst of traffic. The Typhoon base line card is about three times as much as the Trident, so it sends 300ms per NPU and it is serve first-come first-served. So, if on a Trident card there is one interface on NPU, that interface can use either 50ms or 150ms of buffering. If you created two sub-interfaces, then 150ms will shared by those two sub-interfaces and then it depends on the queue limit configuration in the Q0S policy as to how much of that buffer you want to be assigned for the interface. You can allow oversubscription, but then you run into packet anarchy. Snatched from: https://supportforums.cisco.com/document/143011/understanding-cisco-asr-9000-series-aggregation-services-routers-platform |