Cloud computing offers an exciting alternative to conventional compute-platform paradigms
in the field of scientific computing. CMI's cloud computing project (involving
Chris Hill,
Constantinos Evangelinos,
John Marshall,
Glenn Flierl and
Lodovica Illari) is exploring ways in which
MITgcm can be run as
an internet-based application providing an alternative to running on a specific machine or system.
CMI's current thrusts in this area are towards:
Advancing the computer science associated with
getting MITgcm to run on a generic cloud provider
data center.
Developing and packaging an integrated suite of
MITgcm-driven, cloud-based educational applications
for use K-12 and beyond.
High Performance Computing and the Cloud.
While many scientific problems lend themselves to
ensemble techniques involving large numbers of
relatively small calculations run in parallel, most
computing centers give priority to massively parallel
simulations and are not geared towards handling
many-task-computing application types: It is often
easier to run a single 4096 processor job than 2048, 2
processor jobs.
In addition, shared super computing centers, and even
more moderately sized clusters, tend to be geared
towards batch computing (jobs must be queued, notice
given and permission sought) and are therefore less well
suited to handling large peak demand with the hard
deadlines than can occur in, for example, a field
experiment where data must be analyzed in a timely
fashion (of order hours) to optimize operational
decision making in real-time.
Constantinos Evangelinos
(in collaboration with Pierre Lermusiaux, Jinshan
Xu and Patrick Haley in the
Mechanical Engineering
Department at MIT), has been looking at an oceanic error
subspace statistical estimation (ESSE) problem (an
uncertainty prediction and data assimilation methodology
employed for real-time ocean forecasts), involving a
large number of ensemble calculations, as a vehicle for
exploring the execution characteristics and challenges
of a distributed workflow on a large dedicated cluster
and the usability of enhancing this with runs on
Amazon
EC2 and the Teragrid with its
concomitant I/O
challenges.
In removing the need for platform specific special
access and expertise, this technology is also seen as
opening the door to powerful new educational
possibilities. CITE (Cloud-computing Infrastructure and
Technology for Education) – is an
NSF STCI funded
project aimed at supporting the development of
middleware that will enable numerical models to be run
on commercial compute farms (like EC2) via cloud
computing and which can be exploited in ongoing and
future classroom educational activities.
The two linked goals of the CITE project are:
The development of technology that will be
suitable for many educational scenarios, including
providing access to parallel computing resources in
classrooms (K-12 on to university). Students and teachers will be able to run and
interact with numerical models developed by leading
researchers without the overhead of supporting
software distributed to desktops in a school or the
logistical headache of maintaining a cluster
resource. Commercial compute farms will be
exploited in which the technical `nitty-gritty' is
outsourced to these specialized providers.
Showcasing the technology in a `virtual fluid
laboratory' that will be used in ongoing
undergraduate and graduate courses being offered at
MIT (eg in 12.804,
12.307) and collaborating universities.
CITE (Cloud-computing
Infrastructure and
Technology for Education) -
A taste of things to come?
Here one of the team's
children demonstrates how
this technology makes fluid
modeling "child's play"...
Because the software technology that is being developed is very general,
it can be
applied to make computer models part of almost any
course available in, for example,
MIT's Open Course Ware initiative. It can release the power of compute
clusters to anyone with a low cost laptop or `netbook'
computer.
Read
an MITgcm news story about work to develop earth
science visualization tools intended for use in a
cloud-computing environment.
Publications
Evangelinos, C., P.F.J. Lermusiaux,
J. Xu, P.J. Haley, and C.N. Hill, 2010.
Many Task Computing for Real-Time Uncertainty
Prediction and Data Assimilation
in the Ocean, IEEE Transactions on Parallel
and Distributed Systems,
Special Issue on Many-Task Computing, I. Foster, I.
Raicu and Y. Zhao (Guest
Eds.), Submitted.
Evangelinos, C., P.F.J. Lermusiaux, J. Xu, P.J.Haley,
and C.N. Hill, 2009, Many Task Computing for
Multidisciplinary Ocean Sciences: Real-Time
Uncertainty Prediction and Data Assimilation, 2nd
Workshop on Many-Task Computing on Grids and
Supercomputers (MTAGS'09) at SC09.
Evangelinos, C. and C.N. Hill, 2008,
Cloud computing
for parallel scientific HPC application: feasibility
of running coupled atmosphere-ocean climate model on
Amazon's EC2, Cloud-Computing and Its Applications
conference, CCA-08, extended abstract.
Read
an MITgcm news story about work to develop earth
science visualization tools intended for use in a
cloud-computing environment.