Hi Qiong,

On 2011/08/04, at 18:40, Qiong wrote:

> Hi  Remi, Tetsuya,
> 
> Thank you for your reply. Please see inline.
> 
> Thousands of rules seems to me a lot. 
> (I keep doubts that, if CE's support statically shared addresses, keeping 
> thousands of IPv4 prefixes would be needed to support IPv4 via IPv6.)
> In any case, this can be among factors that differentiate which solution 
> applies best to which network.
> 
> [Qiong]: Agree. 
> 
> Adapting a CE to up to 1000 rules doesn't seem difficult to me, and with 
> O(log n) matching this can be done with satisfactory performance. 
> (More details would be private consultancy ;-)). 
> 
>     [Tetsuya] Thousands of rules sounds a lot to me as well. But I don't 
> think it is difficult for CE to support such a    number of rules. In fact, 
> our implementation was tested under 10k rules in our lab.
> 
>     [Tetsuya] Also, the rule consists of Domain IPv6 prefix/length, Domain 
> IPv4 prefix/length, CE IPv6 prefix length and Domain IPv6 suffix/length. So, 
> the total data size of these tuples is 40bytes roughly. So, if using 1000 
> rules, the total data size is roughly 40kbytes. So, I don't think the big 
> chunk of memory is not required for storing all of rules in CE.
> 
> 
> [Qiong]: I can fully understand that rule matching algorithm would not be 
> difficult to handle 1000 rules. But we should still keep in mind that 
> traditional CPEs are usually embedded systems with restricted resources. The 
> amount of NAT sessions it can handle is approximately about 2000~3000. When 
> we add another 1000 rules to do prefix matching in CPE, I think it would 
> still cost more than before.  

Agreed. Some of traditional CPEs might not be able to handle a bunch of NAT 
sessions due to the resource limitation such as memory, cpu, etc. From the 
implementation point of view, I think the required resources for 4rd mapping 
rule is smaller than the resources for NAT session because usually NAT session 
contains many information.

> BTW, what's the output performance of your algorithm apart from memory 
> occupation and what's the configuration of your platform in your test ? 

Did you asked about the forwarding performance? We used netbsd running on arm 
core cpu (400MHz). The accrual ratio of the latency is just 2 to 5 % under 4rd 
environment.

Thanks,
Tetsuya Murakami

> At the same time as other parameters (once in awhile to avoid lifetime 
> expiry).
> 
> [Qiong]: I think in this case, we would need another specification to 
> distribute this rules automatically. 
> 
> 
> Thanks
> 
> Best wishes
> 
> Qiong Sun 
> 

_______________________________________________
Softwires mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/softwires

Reply via email to