[dcrg-interest] 答复: IP over IP solution for data center interconnect

[Xuxiaohu <[e-mail]>]

> -----邮件原件-----
> 发件人: Benson Schliesser [mailto:[e-mail]]
> 发送时间: 2011年12月24日 1:45
> 收件人: Xuxiaohu
> 抄送: Eggert, Lars; [e-mail]; [e-mail]; [e-mail]
> 主题: Re: [dcrg-interest] IP over IP solution for data center interconnect
> 
> 
> On Dec 23, 2011, at 12:38 AM, Xuxiaohu wrote:
> 
> > By the way, the ARP table scaling issue on DC gateways, which is deemed by
> the ARMD WG as the only worthy ARP problem in data center networks, could
> also be solved with the IP over IP solution.
> 
> Two comments:
> 
> First, if anybody is interested in exploring additional address resolution issues in
> the datacenter, please speak up now on the ARMD mailing list. We are
> attempting to finish our Problem Statement soon. The Problem Statement will
> only include those issues that are described by WG contributors. So if you think
> we should deem other issues to be worthy problems, please contribute to the
> discussion.
> 
> Second, at a high-level I think that a number of ARMD participants will agree
> with your comment about ARP scale issues being solved by "IP over IP"
> solutions. In fact, it's not just IP-over-IP that can help. Any map-and-encap
> scheme, including L2-over-IP, can help by shifting the burden to a mapping
> mechanism. Local ARP/ND proxy functions (e.g. in the VM v-switch, TOR, etc)
> could take advantage of the mapping mechanism to facilitate address
> resolution.


As for L2 over IP solutions, I agree that the ARP broadcast flood impact on the servers/routers and networks could be alleviated by using the ARP cache mechanism (tools.ietf.org/pdf/draft-shah-armd-arp-reduction-02.txt) or some ARP mapping system.

However, since the DC gateways usually need to forward traffic to so many hosts at most time, they need a very large ARP cache table to contain so many ARP entries. Otherwise, the ARP cache aging-out timer should be set short enough. In the way, the DC gateways would have to send ARP requests more frequently, in addition, the gateways are required to have a much large packet cache to cache the received packets destined for the hosts before the ARP responses for them are returned, otherwise the received packets would have to be discarded when the cache is overflowed. In a word, I don't know how the mapping system could help on this ARP table scaling issue.

Best regards,
Xiaohu

> In my personal opinion: The real problem is figuring out how to do the mapping,
> what the right mapping approach is for a datacenter, etc.  This is the topic of
> the NVO3 effort that was discussed during L2VPN in Taipei.
> 
> Cheers,
> -Benson
>