[GHPN-WG] Scalability

[Michael Welzl]

Hi all,

I couldn't agree more with all of this.	Being a networking person,
but not really from the "optical world", who got into Grids, I
was heavily disappointed that 90% or more of the work goes into the
direction of letting end systems participate in GMPLS-style
signaling. This (as you rightly state) cannot scale, at least not
to sizes that some Grid people are dreaming of. To me, it's a key
aspect of the Grid that it should run over the Internet, not some
dedicated (or half-dedicated, possibly IP based) network.
However, it seems to me that "Grid Networking" is usually not
"Grid InterNetworking", and this is why we have that in the name
of our European project EC-GIN: http://www.ec-gin.eu
where doing network research for truly Internet-based
(i.e. scalable!) Grids is the focus.

With a few exceptions (e.g. Grid-specific network measurements),
developing Internet mechanisms for Grids has two major
practical hurdles, though:

1. a lot of this work involves changing the TCP/IP stack, which
   seems impractical in many Grids today, where administrators
   often dismiss code if it even requires root access

2. developing truly new and suitable network mechanisms
   is a long-term effort, involving modeling etc. (we still
   don't know how to properly simulate a Grid, including a
   full flavored network simulation - e.g. ns-2 with Grid
   traffic), and no 100% guarantee of deployable code even
   after a Ph.D. thesis - in my opinion there's a funding
   gap here, because no Grid project could enable such work.
   Well again, this is why we have our project, but I truly
   believe that we're quite an exception here, and even we
   were asked to deliver working code earlier than we wanted
   and planned.

Some work should go into fixing these two problems.

My 2 cents...

Cheers,
Michael

PS: risking to get on your nerves by mentioning that again -
but really, this focus on scalable, Internet-centric network
things for Grids is what I've been propagating for the
GridNets conference since I became involved, so this is the
ideal venue for such results - if you have something, please
submit a paper to us!



On Wed, 2008-06-04 at 18:19 +0200, Freek Dijkstra wrote:
> Simeonidou, Dimitra wrote:
> 
> > -Dimitra: ghpn should open its consideration to wider user
> > communities (not only lambda service users) and look at scaling
> > capabilities and interfacing issues of different technologies. i.e.
> > the heterogenity considerations should also extend to mobile/wireless
> > domains.
> 
> Dimitra, thanks! Quite a few suggestions actually; I'm sorry to have 
> left them out (at least, now the best is saved till last). I think your 
> comment on scalability is particularly dead on.
> 
> Today, I've seen an impressive list of control planes, including 
> multi-domain control planes. Unfortunately, I do not think any will 
> scale to the scale of the Internet. AutoBAHN and Phosphorous for example 
> have a database of all domains and their interrelation. The same applies 
> for interdomain GMPLS.
> 
> To quote RFC 4655 (section 4.9.1), about the scalability of GMPLS: "[the 
> path computation architecture] is not considered to be a solution that 
> is applicable to the entire Internet.  That is, the applicability of 
> [this architecture] is limited to a set of domains with known 
> relationships."
> 
> I think this statement applies for all control planes at the moment.
> 
> With NDL, we have created a distributed topology database. However, that 
> is only part of the scaling problem. Other solutions may include 
> abstraction of multiple domains, hierarchy of addressing (such as IP 
> addresses). Remember that the number of constraints is a lot larger 
> though: topology constraints, technology constraints, scheduling 
> constrains, policy constraints, etc. And that's just for network 
> resources, ignoring computing and storage resources that a 
> meta-scheduler needs to take into account.
> 
> I think this is a worthwhile research topic, and the timing is right. It 
> has been shown that multi-domain control planes are possible. Now is the 
> time to investigate if this can scale up to 1000s of domains.
> 
> Regards,
> Freek Dijkstra
> --
>   ghpn-wg mailing list
>   ghpn-wg at ogf.org
>   http://www.ogf.org/mailman/listinfo/ghpn-wg
>