Skip to content

Summary Jim Waldo’s keynote at the 10th Jini Community Meeting

October 5, 2006

Summary of Jim Waldo’s Keynote at the 10th JCM





[Disclaimer: This is my summary of
Jim’s keynote from the last JCM. It is not a transcript; Any errors
are probably mine! Thanks to Jerome Bernard for video taping Jim’s
presentation! The presentations and videos from the JCM can be found
here]





Jim is thinking about things that cause
a quantum jump in complexity. You are OK, and then something changes
in a big way.





The basic jumps in complexity occur at
these points (I add abbreviations here to refer to them later on)






  • Sequential (SEQ)– life is
    good, life is easy.


  • Multi-threaded (MT) – takes
    retooling and a competent programmer to think about MT


  • Multi-Process (MP) – For
    everyone other than kernel developers, this came before MT


  • Multiple Machines (MM) on the same
    network. Not the same as multi-process, but some people think it is


  • Multiple Untrusted Machines (MMU)
    – Essentially the web






All of the above cause discontinuities
in the programming model. As you move through each stage, you lose
something:






  • seq -.-> MT – you lose
    ordering (multiple things can happen at once) This is hard –
    as we naturally think sequentially.







  • Move to multiple processes
    – lose single context (i.e. A shared context that we can rely
    on). Global state is used all the time in development (think
    anything static).







  • Multiple Processes to
    Multiple Machines – state gets lost. Global State of your
    “system” is a fiction. There is no consistent state in
    an interesting distributed system (Jim references Lamport’s work on
    this). Distributed OS projects attempt to introduce global state –
    they have largely failed.


  • Move to untrusted machines. You
    lose trust. In the difficult position of not knowing who you can
    trust.



But you also gain some things as you
move through the discontinuities (otherwise why would you do it?)






  • Seq-> MT – you gain
    parallelism


  • MT – MP – you gain
    isolation (gives you safety)


  • MP to MM gives you independent
    failure (parts of your system can survive if things fail)


  • MM to MMU – gives you scale
    (web scale, Internet scale). Use someone else’s resources (or allow
    someone to use ours).



The Platform





The platform – allows us to do
the work. The model the developer sees:






  • SEQ – A batch OS is fine.


  • MT -language extensions needed for
    correctness (ensure system wont re-order things underneath you)


  • MP – need communication
    mechanism between the processes.


  • MM – Not clear what the
    platform is? Attempts include RPC (invented at PARC), CORBA, Jini,
    XML/SOAP. Not clear we have figured this out. We know what it is
    not. Grids trying to be the platform.






Most grids are an attempt to do batch
on a large scale (scientific). Scheduling jobs. Use individual OS in
any way you wish – not a platform, but a way of aggregating
platforms. How do you give a programming abstraction. Jini is a good
attempt – but only the beginning. What is outside the grid, vs
inside. Outside – untrusted, inside is trusted – mutual
trust





Two discontinuities –
inside/outside – we are conflating them – trying to solve
all the problems at once won’t work. We need two solutions.





Jini 1.0 – built for the
1st discontinuities – MM but assumes full trust.
Built on mobile code.


Ad-hoc organization – changes
over time. When a service enters into the system, you could trust it.


Failure handled – but not a
failure of trust.


Types – the way which you
identity things





Jini 2.0 – Multiple
Untrusted Machines. Adds security. This is hard to do with mobile
code. Adds a lot of complexity (e.g. proxy verification). Perhaps too
hard.


Configuration – deployment
control. Deployment errors cause a lot of failures in distributed
systems. A lot of the complexity is making services reliable for all
possible deployments. Maybe we could communicate the reliability
needs in the language. Needs to be part of the platform.





Program vs. deploy – we are
trained to keep them separate – but this probably makes life
harder. Famous disclaimers “We will put management in later”
(just like “we will put security in later”).





Services – the things we assume
are always there (but sometime they might not be!) Example:
persistence. A file system may not always be there in all
environments. What would a good persistence service look like?





Containers – is a function
converting a type to another type. Must be able to import things,
give bindings, yields another container, Containers are type
functions.





Virtual machines. JVM is an example. VM
is the important thing – not necessarily the JVM. Universal
binary allows us to move code and data. Jini is Java VM thing. Other
Vms might be interesting (e.g. ones that fix classloader problems?)
The abstraction is the important thing.





Feeling of Deja Vu for Jim. Back when
all kinds of different O/S existed there were lots of debates “Why
mine is better than yours”. What arose was UNIX and the “other”
(i.e. There was a lot of convergence about what an O/S looks like).
Model of the platform happened after lots of experimentation and
discussion. Are we at the same stage as 30 years ago with O/S –
a breakthrough may be close.




Advertisements

Comments are closed.

%d bloggers like this: