Dell Power Solutions

HIGH-PERFORMANCE COMPUTINGDesigning High-Performance Clusterswith the Dell PowerEdge SC1425 ServerHardware selection for high-performance computing clusters is often driven by thecharacteristics of the parallel applications that will be deployed. This article discussesdifferent classes of parallel applications and presents the Dell PowerEdge SC1425server as a viable, low-cost platform on which to build clusters for different classesof applications.BY RON PEPPER AND RINKU GUPTAAn important consideration when designing a highperformancecomputing (HPC) cluster is the characteristicsof the parallel application that will run on thecluster. Application characteristics go a long way in determiningthe components needed for the cluster. For example,a long-running parallel application that exchangesmany small messages between nodes in the cluster mayrequire a special network communications infrastructure.In this case, the cluster design should specify that thecompute servers be connected to each other by a fastnetwork interconnect that can send and receive manysmall messages very quickly.Defining the types of HPC applicationsA parallel application runs on a distributed collection of nodes.Hence, all applications consist of communication steps(to communicate between themselves) and computationsteps (to perform independent computation). Based on thedegree of communication and computation, parallel applicationsin the HPC field fall into three broad categories.Coarse-grained parallel applications. Beowulfclusters—that is, parallel-processing HPC clusters comprisingindustry-standard components—have traditionallybeen built to exploit coarse-grained parallel applications.Parallel applications for which the overall time spentin computation is much higher than the time spent incommunication fall into this category. A coarse-grainedapplication is an ideal candidate for running Beowulfclusters, because there is a great probability of obtaininghigher performance at a lower price when additionalservers or CPUs are added to the cluster. Because of theirhighly parallel nature, these applications are also calledembarrassingly parallel applications. A classic exampleis the Monte Carlo simulation problem.Medium-grained parallel applications. Applicationsfor which the computation time is greater (but notmuch greater) than the communication time are consideredmedium-grained parallel applications. As comparedto coarse-grained parallel applications, medium-grainedapplications have a much lower computation-to-communicationratio. The communication overhead is apparentfor medium-grained parallel applications.Fine-grained parallel applications. In a fine-grainedapplication, the time intervals spent on communicationtasks are greater than or equal to the time spent on computationtasks. Such an application is not considered thebest candidate for Beowulf clusters. Even if the applicationis parallelized and run on a cluster of nodes, the scalabilityof the application may be limited.Choosing the right cluster componentsComponents in a cluster are generally selected based onthe generic applications that will be run on them. Everypart of the cluster plays an important role in the clusterenvironment, depending upon the specific application’sability to exploit it. Two of the most important componentswww.dell.com/powersolutions Reprinted from Dell Power Solutions, February 2005. Copyright © 2005 Dell Inc. All rights reserved. POWER SOLUTIONS 115