Maciej Drozdowski:
Presentation - questions on slides.

Wilfried Jakob:
Grid model vs existing clusters, planning vs queueing.
Need for multi-criteria scheduling.

Maciej Drozdowski:
Using many criteria is too difficult for standard users. A user wants
to simply press the button and obtain the result.

Wilfried Jakob:
It's not the users but the administration who should tell the system
about the criteria (maximum acceptable time and budget).

Peter Krusche:
Maybe the users should manipulate the system, cooperate in some way?

Wilfried Jakob:
This is too expensive: the users could do this, but only to some extent.

Maciej Drozdowski:
There are papers about fairness and cooperation in sharing resources,
so some people already tried this approach.
Another issue: good prediction of time needed for task completion is
often impossible, but the existing algorithms rely on such data.

Krzysztof Kurowski:
Another issue: different systems have completely different
characteristics and user behavior. 
It makes no sense to create common strategies and algorithms for
different real systems.
The algorithms should be dedicated to existing real systems and be
tested on real workloads.

Maciej Drozdowski:
What should be added to theoretical models to make them more suitable
for real systems?

Ariel Oleksiak:
Some parameters should not be used, e.g. estimated time of executing
the task.
Often it is not known, or even if it is, people don't want to give
this information to the system.
Probably we need to take into account some other system parameters.
In practice it is possible to apply only the simplest algorithms based
on task queueing.

Maciej Drozdowski:
Maybe we don't need more sophisticated algorithms?

Łukasz Maśko:
Execution time should be given in many variants, for different
machines, because there are different computers in a grid.
Often it's impossible to give the estimated running time of parts into
which a program is split.
The complexity is not only in software, but also in hardware.
The current approach is to search for scheduling algorithms on a given
platform for many different applications.
Maybe we should change it: based on the information about applications
we want to run, we should try to build a system in which they can be
executed very fast.
Maybe simplify the hardware, so that only the things we really need remain.

Maciej Drozdowski:
But this is another level of abstraction, scheduling application
(inside), we were talking about scheduling execution of many
applications.

Łukasz Maśko:
Simplifying each layer (from hardware to software) should help.

Krzysztof Kurowski:
We need scheduling algorithms for each level, and good interfaces
between the levels -> scheduling stack.

Maciej Drozdowski:
There is strong pressure on utilizing all resources, but does it
really matter?
Usually a user can wait a few seconds longer to get the results.
We don't need maximum possible efficiency, 80% is enough.

Ariel Oleksiak:
There is no need for maximum efficiency on your laptop, but it may be
very important when sharing resources.

Maciej Drozdowski:
The number of cores is growing, so maybe we will have clusters on our
laptops in a short time.
And then everyone will need scheduling stack and effective algorithms.

Łukasz Maśko:
This will not be important, because most people use only simple
applications like text editor etc.

Maciej Drozdowski, Peter Krusche:
There will be new applications using paralelism and this scheduling stack.

Krzysztof Kurowski:
E.g. for GPU programming, system level schedulers will be needed.

Ondrej Zajicek:
Energy usage is different on laptop and supercomputer.
1% energy is not much for a laptop, but a lot for a supercomputer.

Maciej Drozdowski:
If not users or developers, then there will be some politicians who
want 100% efficiency just because it looks better than 80%.
Summary: what are the most important challenges in creating a
scheduling stack?

Krzysztof Kurowski:
New ways of application of the old models.
New criteria, not only based on energy consumption.
Maybe several scheduling stacks, using diferent criteria, instead of a
single stack?

Tomas Ebenlendr:
Parallel scheduling algorithms. Maybe they can use the same platform
as the tasks which are scheduled by them.

Wilfried Jakob:
No, because probably we will not have enough resources to use some of
them for scheduling in addition to processing the "real" tasks.

Maciej Drozdowski:
Summary of all what has been said (...), the end.