IDF Prague: Terascale Computing Review 4

IDF Prague: Terascale Computing Review

(4 Comments) »

What do we need such immense processing power for? Imagine a search engine like Google but for all music online. You give it a sample clip of the music you like and it will find all similar music that exists. Being able to go beyond text-only indicing make things a lot more slower. To get a rough idea compare the CPU time your system uses while scrolling through a text only document and then how much it needs to play back an audio file. And even one step further than just audio would be video, don't forget about HDTV multiplying the amount of data to process even more.


On massively parallel systems traditional algorithms in use today would see only very limited scaling, so new algorithms have to be found. There is a lot of research going on in that field even today - massive parallel clusters are used by researchers today for their computational problems. But on the x86 architecture such applications are not very common. A big obstacle are the programmers themselves. They have been writing sequential code all their life and now they have to throw over this paradigm and learn something completely new. The holy grail of parallel programming is a compiler that automatically identifies code that can be run in parallel and changes it to threaded code automagically without any user intervention. Intel is working on that too, their C++ compiler can do it to a certain extent already today as was demonstrated in several other training sessions.


This slide shows the expected speedup when increasing the number of parallel cores. As you can see the optimum without any clever optimizations is about 16 parallel threads. Beyond that the overhead for management and synchronization of the individual cores will make the whole system go slower. That's why certain optimizations are needed to allow proper scaling.


In summary one can say that the field of Terascale computing is a highly active area of research which covers much more than just the computational side of things and you often have to think outside of the box to achieve solutions that will work.
Discuss(4 Comments)
View as single page
Dec 27th, 2024 23:36 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts