DU ER HER: news

Hot topics at MADALGO Summer School

Due to an exciting program and popular topics at the MADALGO Summer School, the organizers had to close for sign-up this year.

22.08.2012 | Marlene Nybro Thomsen

- This year we are discussing algorithms for modern parallel and distributed models. Hot topics. The number of registrations was huge, so we had to close enrollment, says Center Director Lars Arge at Center for Massive Data Algorithmics (MADALGOs) annual Summer School held this week at the Computer Science Department. MADALGO is a center of the Danish National Research Foundation (“Grundforskningsfonden”).

Computer Science is about trade-offs: The study of trade-offs and design choices. Suresh Venkatasubramanian, University of Utah explains in short why.

from cs.au.dk on Vimeo.

Effective use

The main hot topic is the development of algorithms that can handle large amounts of data. It is a cross-community area, which meets major challenges in use of the parallel computing resources in e.g. graphics cards, multi-core CPU and large data centers (map-reduce algorithms). To get the computers to run faster and use the right resource at the right time. The outcome has consequences for example cloud computing or energy-saving.

One of the four summer school lecturers Phillip Gibbons from Intel Lab Pittburgh explains the challenge,

- Good performance is about fast running times but you can also consider questions of energy efficiency and less use of energy, he says.

The amount of energy in computation use is a combination of powering up and down different parts of the multicore processor computer. And then how long the computation runs.

Saving energy

The goal is not to use more energy than needed. One way to save energy is to make effective use of the cache. Individual processors on a single multicore will turn themselves to a lower power state if they are not being used and that will save energy.

- We orchestrate an algorithm so that most of the time you are only accessing the closest memory – the nearest cache. That’ll speed up the computation and thereby save energy, Phillip Gibbons explains.

From an algorithm designer’s perspective he is challenged by the machine architecture, when trying to make algorithms more efficient.

In the last couple of years there has been a particular approach based on a so-called work stealing scheduler to decide which part of the computation to run at which processor at which time. The particular scheduler leads to computation that is fast and takes less energy.

- In order to get good performance you have to make effective use of the cache hierarchy.

Problems unsolved

A number of fundamental research questions remain unsolved. The research can help algorithm designers and programmers to get better performance on multicores without having to work so hard.

- So the question is how we can think about algorithms in this setting, and perhaps in a way that can abstract from the specific hierarchy. So we can design the algorithm once and it will be good for different hierarchies, automatically, Phillip Gibbons outlines.

Lars Arge from MADALGO adds;

- Our goal is two-fold. Of course to design efficient algorithms for current parallel architecture, but also to understand their limitations in order to suggest new better architecture, he says.

Exciting time to look at algorithms

The goal of the summer school is to provide an in-depth introduction to some of the models of modern parallel and distributed technology, and the key techniques used to design efficient algorithms in these models.

- It is a very exciting time to look at multicore algorithms. Because it’s a relatively new aspect. We’ve seen an explosion over the last 5-6 years in the extent to which multicore processors are available and in heavy use. The number of processors continues to increase. The features that go into the multicore computer are getting more interesting for algorithm designers. The frameworks are interesting to make it more easy to use, Phillip Gibbons states.

For more information:

www.madalgo.au.dk/pardist2012

alle
Henvendelse om denne sides indhold: 
Revideret 19.12.2012