I just finished a preliminary data modeling session with the results I gathered from the ASTI cluster running the parallel number finder, and am currently seeking the possibility of getting the results into CDROM's. But from the looks for the control tests, something's going to have to give, especially in the theories regarding performance scaling, and the effect of having more nodes working on a huge problem. It's high time I read more on current studies regarding the effect of a load balancing method on the performance of a certain application.
I'm now more convinced that some sort of load balancing and delegation would need to be put in especially on a system which needs to work under stress, and stay up almost all the time for a reasonable level of performance enhancement. Seeing the results, and checking what I already know about parallel computing, the communication network and distinct properties of networking hardware affects heavily the performance of a parallel application being run even on a dedicated loosely coupled parallel computing architecture.
And if the control tests are something to abide by, I would dare say that the theoretically sound concept would definitely have trouble with empirical results. And with just the control tests, I can even try to show that the proper choice of chunk sizes for computation would indeed have an effect on the scalability and the performance of the application as a whole. And knowing that the chunk sizes do affect the performance of the tests, then certainly a means of determining the proper chunk sizes would definitely have an effect on the performance of the tests. It sounds redundant, but I have to repeat that over and over in my head -- and after a while I realize that I'm getting to something here.
The paper I was working on a while ago would stay untouched for yet another hour or so, just so that I catch up with the things I have to do other than the thesis (there's eating, personal hygiene, and family matters). But seeing the pattern and the trends in the data, I would definitely like to think that certain phenomenon would still need to be identified (not by me necessarily) but by other people also interested in the field. As for now, I am left bewildered, surprised and quite lost actually on how the data is letting me see things I never thought I would see. I would need to let the ideas incubate a little more, so as to get a more firm grip on the concept being thrown at me.
Until next time...
I'm now more convinced that some sort of load balancing and delegation would need to be put in especially on a system which needs to work under stress, and stay up almost all the time for a reasonable level of performance enhancement. Seeing the results, and checking what I already know about parallel computing, the communication network and distinct properties of networking hardware affects heavily the performance of a parallel application being run even on a dedicated loosely coupled parallel computing architecture.
And if the control tests are something to abide by, I would dare say that the theoretically sound concept would definitely have trouble with empirical results. And with just the control tests, I can even try to show that the proper choice of chunk sizes for computation would indeed have an effect on the scalability and the performance of the application as a whole. And knowing that the chunk sizes do affect the performance of the tests, then certainly a means of determining the proper chunk sizes would definitely have an effect on the performance of the tests. It sounds redundant, but I have to repeat that over and over in my head -- and after a while I realize that I'm getting to something here.
The paper I was working on a while ago would stay untouched for yet another hour or so, just so that I catch up with the things I have to do other than the thesis (there's eating, personal hygiene, and family matters). But seeing the pattern and the trends in the data, I would definitely like to think that certain phenomenon would still need to be identified (not by me necessarily) but by other people also interested in the field. As for now, I am left bewildered, surprised and quite lost actually on how the data is letting me see things I never thought I would see. I would need to let the ideas incubate a little more, so as to get a more firm grip on the concept being thrown at me.
Until next time...
Comments
Post a Comment