[Twg] LNET channel bonding

[Doug Oucharek]

Hi John,

I feel that the HLD does give enough guidance that a vendor should be able to estimate what is required to implement as well as design the feature.  However, it would be important to point out Liang's last comment regarding the possible interaction between SMP and Channel Bonding so that can be taken into consideration for the RFP.

I also agree with Liang that splitting the project would be helpful in making it more manageable.  There seems to be two important parts:

1- Coming up with a way to have a NID refer to a logical grouping of channels.
2- Implementing channel bonding over this logical grouping.

Number 1 can be tricky as it is very similar to how the internet had to come up with DNS to logically refer to IP addresses.  The proposed design calls for a "bonding server" which, in some ways, is working like a DNS server.  That is no small task.  Also, if done properly, it has the potential of addressing other LNet issues (i.e. wanting to support DHCP). As such, this is a project unto itself.

Doug

On 2012-09-26, at 9:02 AM, John Carrier wrote:

> Liang,
>  
> Do you think the HLD provides enough guidance that a vendor could respond to an RFP to design _and_ implement the feature?  Or do we still need to design and implement separately, as has been proposed at our last meeting?
>  
> Thanks,
>  
> --jc
>  
> From: Liang Zhen [mailto:liang at whamcloud.com] 
> Sent: Wednesday, September 26, 2012 12:12 AM
> To: John Carrier
> Cc: twg at lists.opensfs.org; discuss at lists.opensfs.org
> Subject: Re: [Twg] LNET channel bonding
>  
> Hi John, 
>  
> I think most of the design is still valid, but I wouldn't suggest to skip the design phase for a few reasons:
> - it's a HLD, not a DLD. OpenSFS project process has both HLD and DLD in design phase.
> - the HLD described this feature as a single project, I'm wondering if it's possible to have multiple phases for it (or sub-projects), which will make it more manageable.
> - it's worked out three years ago and core LNet has been changed a lot, it could be worth taking time to review it and consider if there's any new complexity or issue based on the latest LNet, i.e: Any potential SMP performance issue if adding this feature.
>  
> Regards
> Liang
>  
> On Sep 21, 2012, at 4:38 AM, John Carrier wrote:
> 
> 
> Hi all,
>  
> Since LNET channel bonding has made it to the TWG's short list of requirements for 2013, I would like to remind the community that Cray already funded the design of such a requirement in 2009.  The high level design document is available here:
> http://wiki.lustre.org/images/e/ee/Channel_Bonding_06_15_09.pdf
>  
> Is this design still valid?  What else would need to be done?  If the design still looks good, then I suggest that the RFP for this requirement should move directly to implementation and avoid the cost of another design phase.
>  
> Please post comments to the reflectors.
>  
> Thanks,
>  
> --jc
> _______________________________________________
> twg mailing list
> twg at lists.opensfs.org
> http://lists.opensfs.org/listinfo.cgi/twg-opensfs.org
>  
> _______________________________________________
> twg mailing list
> twg at lists.opensfs.org
> http://lists.opensfs.org/listinfo.cgi/twg-opensfs.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opensfs.org/pipermail/twg_lists.opensfs.org/attachments/20120926/794c6b8e/attachment.html>