Actually, current codes are implemented like a BSP interface, and its client/server programs for only graph package. As you said, and I also think, it should be able to perform any BSP (Bulk Synchronous Parallel) programs, and it needs to be re-factored.
Would be nice if you can find the problems of independent BSP cluster. On Sun, Mar 28, 2010 at 8:28 PM, Hyunsik Choi <[email protected]> wrote: > Hi, > > I would like to discuss the plan that I think. > > First, I participated in hama in order to contribute graph package. I > plan to develop graph package (called Angrapa) for large scale graph > data running on shared-nothing architecture. Many of people think > about that Angrapa may be appropriate in BSP model, like pregel, and I > also agree with that thinking. In addition, the pregel seems to be > proven by ACM SIGMOD Conf. 2010. > > The BSP package is the base of Angrapa. Thus, we first should complete > BSP package. Besides, we should be concerned with two design > considerations caused by large-scale data and shared-nothing > architecture, recently called cloud computing environments. They are > heterogeneity and fault tolerance respectively. In addition, it is > important to design BSP to be general purpose since it will be used > for matrix package and other problems. > > After pregel paper is published, we should embrace the techniques > discussed in the pregel paper. > > In sum, > * We should complete BSP package, general purpose as possible. > * BSP package has to include considerations of both heterogeneity and > fault tolerance. > > I need advices. > > Best wishes, > -- > Hyunsik Choi > Database & Information Systems Group, Korea Univ. > http://diveintodata.org > -- Best Regards, Edward J. Yoon Co-founder of Twieat.Us [email protected] http://blog.udanax.org
