re: Caspian Scores in Korea4 10-GigE's isn't a lot of media diversity, but as they said, Asian countries don't have a lot of the same infrastructure considerations as western countries. It's like a "greenfield" deployment for newer network systems. Doubt it would fly anywhere else, but still a nice win in a climate where every win counts.
re: Caspian Scores in KoreaOne of the main mandatory requirements in Asia and Korea is MPLS. All leading Korean carriers will deploy only MPLS and IPv6 capable routers.
To get the ETRI deal they probably had to open source code, hence Cisco and Juniper "loss".
re: Caspian Scores in KoreaWhile the flow-based architecture is indeed good classify packets into flow and then treat it with the relevant QoS, but it has a serious scaling problems.
I generally like the flow-based architecture, where you can have very detailed statistics, and they have very flexible and granular QoS mechanism, but issue of maintaining a high and large level of flows is very CPU and memory intensive. Every packets that comes into the router need to be register by the central CPU. Given this, in a Core router position, it will need to handle thousands if not million different packets coming from various edge. The first hit will be Central CPU that needs to register so many flows. So the question of flows per second, memory size, hashing tables etc comes into play.
I really wonder how Caspian overcome or scale this? You really cannot distribute the flow register and rate into the linecards.
re: Caspian Scores in Koreamaxng said " Every packets that comes into the router need to be register by the central CPU."
- I am not familar with the details of how Caspian are maintaning their flows but to be able to scale I would doubt a 'CPU' is involved, in maintaining the flow list. You would want to use some dedicated hardware to do this, probably some functionality incorporated into the Network Processor.
re: Caspian Scores in Koreadont need to do that. A NP or FPGA or ASIC can do it. The question is why bother. as i recall, ipsilon had flow-based routing (using a tag switching precurser to MPLS, bought by Nokia). both mpls and ipv6 offer flow-based routing. whoopie. (weird that caspian doesn't actually have mpls - makes it a tough sell)
----------- Every packets that comes into the router need to be register by the central CPU.
re: Caspian Scores in KoreaMore importantly memory and access to it becomes an issue. From the edge assuming 20 flows average per user a flow based router could easily see 10 million active flows. With P2P an end machine can easily launch as many as 5000 UDP sessions with 200 active TCP sessions. This doesn't even consider port scans which can easily consume tens of thousands of flow entries per host at up to 20 flows per second.
re: Caspian Scores in KoreaAre the caspian flows really down to the ip src/dst + port src/dst level? that is nuts. what value does it bring? who can manage that many flows really, at a OAM level? (and what would they manage at that level?)
----------------- More importantly memory and access to it becomes an issue. From the edge assuming 20 flows average per user a flow based router could easily see 10 million active flows. With P2P an end machine can easily launch as many as 5000 UDP sessions with 200 active TCP sessions. This doesn't even consider port scans which can easily consume tens of thousands of flow entries per host at up to 20 flows per second.
IronCCIE