Changes between Version 8 and Version 9 of Gec5ClusterBAgenda

07/22/09 15:34:14 (13 years ago)
Christopher Small



  • Gec5ClusterBAgenda

    v8 v9  
    3939Leveraging the backbone provided by ProtoGENI (in partnership with Internet2 and HP) and NLR FrameNet to provide physical connectivity between groups within Cluster B, or possibly to other clusters.  For groups which cannot easily access this backbone, connections will be tunneled over existing R&E IP network infrastructure.
     42''The following notes were taken by Christopher Small, GPO System Engineer''
     44= Monday 1:00-5:00 =
     46== To Do List ==
     48* Larry Peterson
     50Plan for this meeting is to bring each other up to speed on what we're doing.
     52Introductions around the room
     54"We are extremely heavily focused on federation."
     56Three ways to federate.
     58SFA only -- WSDL interface specification, you implement it. This is
     59what Enterprise GENI has done. Generated this based on the running
     60code, but plan to go forward using the WSDL spec as the base and we'll
     61use it. SFA-lite is a variant, without security (creation,
     62destruction, and job control on slices). "... called a slice manager
     63is our terminology; it's what the GPO calls a clearinghouse." Only
     64Enterprise GENI is using this.
     66SFA + geniwrapper -- includes security. Write your own back end to
     67geniwrapper. If you've got a preexisting testbed, this is probably the
     68fastest way to go; use the security machinery and interface spec, and
     69have it call into your system.
     71SFA + geniwrapper + MyPLC -- includes PlanetLab O&M machinery
     73RSpecs are the place where we've always put things that we don't know
     74how to deal with. Deined on a per-aggregate basis. Two examples:
     75PlanetLab (nodes only), Max and VINI (nodes + topology). Two
     76representations, XML and XSGR (easier to read).
     78We're dropping Eclipse types going forward, using Sugar instead going
     81"Aggregate of aggregates" is "slice manager".
     83SFI tool runs on the desktop, talks to aggregates directly. Larry sees this
     84as a feature, not a bug; each end user should be able to decide which
     85aggregates he/she want to talk to.
     87The SFI tool is handed a bunch of RSpecs, and it transmits them to the
     88aggregate manager(s) / slice manager(s) that do the work. The SFI tool
     89is not very smart (Larry's words).
     91Naming convention: plc.organization.slice, plc.organization.user,
     92plc.aggregate, plc.aggregate.user (plc.princeton.codeen,, plc.max, plc.max.christracy).
     95Did demo of PlanetLab with MAX nodes.
     97* James Sterbenz
     99GpENI demo, showed that Lancaster University UK, ETH-Zurich are
     100getting set up. Not using SFI to create topology. They are running
     101their own PLC at KSU. Separare PLC, separate VINI, not federated.
     103* Jeannie Albrecht
     105XML description of the experiment. Command line program reads the xml,
     106sets up and runs the experiment.
     108* Chris Tracy
     110Describes the recursive aggregate manager model they have running,
     111including a mention of OSCARS (although not by name).
     117= Tuesday 1:00-3:00
     119* Larry Peterson
     121We need to work on attracting real users
     122- 4300+ registered PlanetLab users
     123- GetResources() + CreateSlice()
     125Let's advertise our new cluster B aggregates to those users
     127* James Sterbenz
     129Are we just adding users to have users? Are you suggesting we
     130advertise our resources as standard PlanetLab nodes?
     132* Larry Peterson
     134Control Framework todo
     136- Complete transition to "new" RSpec model. WSDL file is there.
     137- Rework with standard SSL
     138- Round out SFI as a full-fledged per-user slice manager
     140Federation todo
     142- Settle on peering plan and naming conventions
     143- Settle on SFA-lite
     144- Define realistic policies
     146How are we going to put Enterprise GENI in? Suggest that we plug in
     147the Enterprise GENI control framework underneath SFI as a slice
     150* Guido Appenzeller
     152Eventually we'd like to have something that abstracts some of the
     153details away, a web interface, to make this all easier to deal with.
     155* Larry Peterson
     157Chris Tracy put up a recursive aggregate manager, this may be what you
     160Maybe we can teach GUSH about recursive aggregates, so it can recurse
     161and provide more detail when desired.
     163As long as this remains homogeneous, the slice manager is pretty easy
     164to manage, e.g. PLC, PLE, PLJ. When it's heterogeneous it gets harder.
     166If the user wants to use VINI, the user needs to know how to deal with
     167VINI. Can't abstract that away.
     169Went to nested RSpecs and it got messy, for now pulling back to flat
     170(or set of flat)
     172* Guido Appenzeller
     174Don't we need some kind of pluggable user interface to manage
     175heterogeneous aggregates?
     177* Larry Peterson
     179That's what Chris Tracy demonstrated, this is what we need. The SFI
     180tool is a very primitive version of this.
     182There is no single point that authoratatively knows where all the
     183slices are.
     185We've just broken the clearinghouse model.
     187* Peter O'Neil
     189From a usability standpoint this will be chaos. Need a single portal.
     191* Larry Peterson
     193There's not just one portal to the internet.
     195There is a philosophical issue here. Is GENI this single NSF-funded
     196thing, or is it a loose federation of entities?
     198* Guido Appenzeller
     200Seems to fit into GENI goals to have one centralized clearinghouse.
     202* Larry Peterson
     204What do you mean by GENI?
     206* Guido Appenzeller
     208The internet has DNS, it tells me where I can look for things
     210* Larry Peterson
     212we can have a root registry and if we want to kill a slice we can
     213iterate through it and ask each to kill its part of the slice.
     215I don't want the GMOC to kill slices on PlanetLab. My definition of
     216what's bad behavior isn't necessarily the same as yours. I'd hate to
     217have the British in charge of when slices are killed because I don't
     218agree with their policy.
     220I think we have a three month plan here, but I think we're only
     221partway done.
     223We need to define some realistic policies. Goal ought to be modest;
     224"these are the slices I'm willing to host." E.g. "Codeen I will
     225support, XYZ I won't support." Some sort of peering agreement.
     227Policies will only be as sophisticated as the mechanisms.
     229* Guido Appenzeller
     231We only talk with known entities. We talk to an SM / clearinghouse,
     232not a user. A researcher hand-editing XML sounds ambitious.
     234* Larry Peterson
     236Writing an SM that spoke to a hererogeneous collection of aggregates
     237bogged us down, so we have one that talks to homogeneous aggregates.
     239Alternative to concatenating VLANs is to talk IP. We'll say to the GPO
     240that we'll do both, push concatenated VLANs as far as we can push it
     241and we always have IP as an alternative.
     243If you're creating non-IP protocols, you can run them through tunnels
     244or over layer 2 links.
     246I can create a tunnel that acts within epsilon of a true layer 2 link.
     248Building internets out of concatenated VLANs is a path that was not
     249chosen; we went with the Internet instead. Why are we doing this now?
     250Can somebody explain this to me?
     252* James Sterbenz
     254Because the GPO told us to.
     256* James Kempf (?)
     258You might want dedicated bandwidth
     260* Larry Peterson
     262IP encapsulation is only 28 bytes, it's running over layer 2.
     264* Guido Appenzeller
     266We want to give people hop-by-hop layer 2 so they can put a router in
     267between. We want it to look like there is a dedicated link between
     268them, not have an IP header.
     270* Larry Peterson
     272My argument is cost effectiveness. This will be a large investment of
     273time and effort with minimal value.
     275* Jon Turner
     277Both are legitimate; you can get to more places with IP, layer 2 VLANs
     278will let you get closer to the bare network.
     280* Guido Appenzeller
     282Every slice must support IP fallback
     284* Larry Peterson
     286We haven't solved the problem because we still need to demux between
     287different slices both tunneling between a pair of hosts. GRE tags? UDP
     290== User level tools and user level services  ==
     292My view is that we're not going to make a lot of progress on
     293per-aggregate RSpec viewing / editing tools. I'm setting the bar
     294fairly low on user-level tools.
     296* Jeannie Albrecht
     298Once you have a stable version of SFI, I can try to make things
     299prettier. But it doesn't make sense until SFI stabilizes.
     301* Larry Peterson
     303Look at the VINI rspec.
     305If we're going to attract real users, we've got to give them something
     308Goal for SFI: readable/editable RSpecs e.g. using XSGR
     310* Chris Tracy
     312We have some tools that might be usable
     314* Larry Peterson
     316Show of hands -- who has time for a GUI?
     318* Guido Appenzeller
     320What we see as the hard part of the GUI is how to help the user if the
     321request fails, what changes they can make to get their request to be
     324* Larry Peterson
     326Will try for pretty text plus policy feedback.