7.3 Process-level Parallelism

Parallel systems are tricky beasts, whether you bought the full package or put it together from a kit, and applications programming is always time consuming. The difficult issue then is the relative additional costs of a Beowulf with respect to the commercial offering. This often depends on the environment in which the system is to be used and the culture/talent base available. Sites that already have the necessary talent usually know it, and hence have been more likely to adopt the Beowulf approach. While efforts are being made to reduce the required level of expertise, it is unlikely that Beowulf systems will ever be any more turn-key than, say, an equivalently powerful commercial system. Nobody thinks twice about investing in the talents of an experienced sysadmin to manage or program a million dollar commercial MPP. Although a great deal cheaper to purchase, an equally powerful Beowulf system will probably need equally talented staff.
10.4 Total Work versus Peak Performance
A perception exists that large and compute-intensive applications need supercomputers to carry them out. This is the peak-performance argument. But such machines are rarely made available to a single dedicated problem for long periods because they must be shared among many contending users. Beowulf-class systems may be more readily dedicated to individual tasks for a longer period of time than their mainframe/supercomputer/MPP counterparts because of their low cost. While the normal operational mode of a large expensive tightly-coupled system is as a shared resource, a moderately sized Beowulf may be devoted to the needs of a single computational experiment for a significant duration. The important issue is the total aggregate work performed on the problem, not the peak performance applied to the problem. With large systems in shared environments users get a small portion of the total machine or the entire machine for a short period of time. With a Beowulf, the user may get the entire machine for an indefinite period. Over an interval of time, say a week, the total work accomplished by the Beowulf on the problem is many times that available to a specific user of a major system shared among many users. Thus, from the standpoint of a computational scientist a moderate sized Beowulf can actually outperform a large shared system from the standpoint of the time between program submission and program completion.

 



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