Tip:
Highlight text to annotate it
X
And in the next couple of slides, I... wanted to just
give you an idea of what one can do with such a framework.
For example, here are some of the... user objectives
that we studied in a... paper a few years back.
And so we said, 'Okay, let's look at... if I'm looking at a...
hybrid cyber-infrastructure, what would be
interesting user objectives?' And one could be acceleration.
[pause]
Can I use Clouds as accelerators to somehow
improve the execution time of my application?
And what we used there was, at that point,
queueing times would become... on popular resources,
on things like TeraGrid were... significant,
and so we were saying that, 'If I have an
application workflow, and if there were certain tasks...
[pause]
that could be offloaded to Clouds, then can I...
offload them while I'm in the queue and try to spend
a few dollars to get as much work done as possible?'
And so especially if you had large ensemble runs,
you could start priming things so you could start
doing some sort of data preparation, you could
do some initial explorations while you're waiting
in the queue and using Cloud resources.
And so we have some interesting experiments
of the benefits of doing something like that
and what are the associated costs.
[pause]
Another thing we said was, which is not
very unrelated to the first objective, was one of conservation.
Clearly there are some parts of your application workflow
that have to run on this end resource that you have,
maybe because the configuration or capabilities
that... resource provides. But the other parts
of your application that may not need this kind of
capability, that as I said, data preparation,
some exploration of parameters or some simple analytics
that don't have to run on this high-end resource.
So can I... and since I have limited allocations,
can I use Clouds to run these... tasks that don't need
the high-end resource, and then conserve
the overall resources that I use on the... HPC system.
And then again, I have constraints in terms of
how much I'm willing to spend for Cloud resources,
or what percentage I want my... or resource that
I want to conserve. And we can play with
different tradeoffs between benefits, costs...
and execution time of the application, and there was...
the paper included some... detailed analysis in this space.
And the final one was resilience and the idea here
was that there was... unexpected downtime
on the system or something goes wrong,
and rather than having my week-long job terminate,
can I offload it to the Cloud while the system
comes up or partially upload it... offload it to the Cloud.
And so we played with some scenarios in terms of
the cost and benefits of doing something like that,
and even the feasibility of doing something like that
for different applications. And can I use the Cloud
as a way to improve the overall... resilience of my application,
given that infrastructures sometime tend to go down?
[pause]
And the different constraints that we... looked at were
constraints in terms of... deadlines, constraints in terms of...
[pause]
completion time of the application. For example,
to select the resources that were most appropriate
for that type of application based on certain
application deadlines. For example, I need this data
done because I need to use it to make some
decisions or to configure some experiment,
and there might be some stricter deadline
on... the execution of certain parts of it.
Budget, when you talk about public Cloud budget,
becomes important, and so you might want to
constrain how much you want to spend,
and how do you want to trade off time for cost.
And so we played... we included mechanisms
to be able to do that, and then we allowed...
we've also done some studies as to how
these tradeoffs for different workloads...
[pause]
come about. And then finally we also looked at...
how to trade off different classes of resources and...
to be able to keep higher level constraints such as
cost, budget and deadline, but then be able to
trade off resources from an economic point of view
so that we are able to meet these higher level constraints,
but then make things as cost-effective as possible.
And there are other such constraints you have looked at,
these are just a few sampling that I've put in here.