7.3.1 Example: Ray Tracing Animation

10.5 Big Memory versus Out-of-core
Large shared systems have one crucial advantage over dedicated Beowulfs. Even if access to the machine is limited to only a few hours per week, the user with a 500GByte data set will accomplish more on the large shared system than on the dedicated Beowulf. The reason is simple. The problem simply does not fit in the Beowulf's memory so the Beowulf can make no progress at all. Purchasing enough memory would be prohibitively expensive for a dedicated system. On the other hand, the large capital investment in the large system's memory can be shared by all its users.
One way for Beowulf systems to sustain their price/performance advantage, even for large memory problems, is out-of-core techniques, i.e., exchanging utilizing disk space as dynamic memory. In a sense, this is doing explicitly from the application program what virtual memory is supposed to do. The advantage of controlling secondary storage from the application is that the access patterns may be well understood and therefore can be exploited. As noted in other sections of this book, almost nothing is cheaper than a spinning bit (a bit of disk storage). When data can be partitioned, operated upon, and replaced, disk storage can be used in place of main memory. The programming methodology for application management of disk storage is similar to that of programming loosely coupled clusters; they both require that latency tolerant algorithms be employed. Thus, once an application has been crafted to perform well in the presence of latency, it may be able to exploit disk storage to hold large data sets. For such applications, a moderately scaled Beowulf dedicated to a single problem for a significant period of time can perform large applications in less time, and for far less money than a heavily shared large high end computer. Returning to the scenario above, if the application can be recoded to effectively use out-of-core storage, then there is no reason at all why a modest Beowulf could not address a 500GByte data set.
10.6 Bounding Influences on the Beowulf Domain
While Beowulf-class computing affords great opportunity for many applications and contexts, its innate properties limit its ultimate scope. The success of Beowulf is due primarily to the wide availability of its components. But Beowulf's limitations can be found in the same market forces that drive the development and distribution of those components. The market, while friendly to the needs of spreadsheets, word processors, and video games, is indifferent to the integrity of 3-D plasma

 



How to Build a Beowulf
How to Build a Beowulf: A Guide to the Implementation and Application of PC Clusters (Scientific and Engineering Computation)
ISBN: 026269218X
EAN: 2147483647
Year: 1999
Pages: 134

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net