Michael Resch started out by explaining that the Cloud is not the continuation of the Grid: the Cloud is something completely different. The Grid is focused on moving out of HPC and reaching out whereas the Cloud is more coming from the community that has a much better understanding of how you handle hundreds of thousands of users of providing services that are packaged and encapsulated.
The last 20 years, we have moved HPC to small and medium sized enterprises and also to larger industries, Michael Resch stated. This has worked out as long as the prize for hardware was relatively high. Software was getting more and more important. One hour of a CPU core is 2 eurocent which is cheap. In 1995, one CPU core was 50 euro. This is a reduction with a factor of 2500.
The software cost is still high. People are no longer interested in a high performance system, they are more interested in a solution, Michael Resch went on. The Cloud technologies have developed solutions in a completely different environment of very big number of small jobs. These small jobs are irrelevant enough to not require too much security nor reliability. Therefore, we have pretty stable solutions in the Cloud.
Rather than try and convince a huge market to adapt our solution, we have a huge market and we can adapt their solutions to our requirements. That is a much easier task because the bulk of work and financial effort is coming from a huge community for which this is extremely important. That's why at HLRS we figured that by Cloudification we can probably extend the reach of our supply, support and services to small and medium sized enterprises because their problem is that everything should be packaged, explained Michael Resch.
Still, they need some knowledge, so HLRS has to do a lot of educational training. The strategy is twofold: on the one hand to adapt and adopt Cloud technologies and on the other hand to do training for people who are in desparate need of solutions, although they are very skilled in providing solutions to their own customers. We have to leverage this solution space of Cloud with our training and education activities and with the hardware that we can provide. As such, we have to able to create solutions for small and medium sized enterprises.
I still think that they will need some support because an HPC system is a special system, Michael Resch believes. You will never be able to completely hide all the complexity but we will make a huge step forward by Cloudification.
What we typically have in our systems is a usage that is in the range of 85 percent. Going higher than 85 percent for a supercomputing system is extremely difficult because you have medium sized and large jobs and you always have small pockets of cores that are not used in the system, 64 here and 128 there, for example.
When we did the movie for the small company, Mark13, in Stuttgart, we were able to fill all these gaps, Michael Resch told. For three or four months, when we did all the pictures for the movie, our usage was going up to 98 or 99 percent, which I would have said two or three years ago, is absolutely impossible because you cannot operate the system. However, we did not increase the waiting time for our research users who did not even realize that we filled all the gaps with small jobs.
In June/July last year, they were complaining so we tailored the queuing system slightly to reduce waiting time and they did not even notice that there was this huge set of small jobs, Michael Resch concluded.
More information is available at http://www.hlrs.de .