The Formation Of Hybrid Heterogeneous Clusters And Its Applications

Since the advent of Computer Science which is called the third-class science, thepeople’s pace of understanding and changing the world continues to accelerate, alsounceasingly to promote the further development of Computer Science: serialtechnology was slowly replaced by parallel technology; high-performance computinghas been gaining popularity; in the consequent high-performance computers andsupercomputers. With the further development of computer hardware and software, aelectronic system called cluster is able to use the ordinary PC to build ahigh-performance calculation and high communication efficiency computing systems,and as it has more advantages than HPC and supercomputers such as the more cheaperprice, more simple for building managing and maintain. It is welcomed by more andmore users, especially small and medium-sized research groups whose computingresources have some common characteristics, such as the performance of astand-alone compute is not high, the architecture of compute nodes have largedifference.In this paper, we use three kind of compute nodes which architectures aredifferent as experimental hardware, and use Rocks as the basic cluster system to buildthis kind of electronic system, mainly through design rational node attributes andre-config the torque to achieve the control of the different schema nodes andmaximum utilization of resources, and add GPU module in ganglia to achieve thesurveillance of GPU nodes in cluster. In addition, this paper elaborates the entireprocess of how to design, implement and maintain this electronic system, andsummarizes experience, and provides a reference for these people who have theintention to form a cluster.We analyze the needs of this electronic system, then design and achieve thecorresponding functions. After the completion of the cluster, a test for this electronicsystem is also carried out. In HPL test, let the problem size and partitioned scale to31000and256as a subsequent test parameters. Use these parameters in8-nodesparallel test, the measured performance only30%of the theoretical value, but whenincrease the problem size to5W, it ascend to42%, but the utilization rate of CPU stillhave big promotion space. In addition we also use NAMD to test GPU, using the benchmark of ApoA1, we get a4times speedup. This paper further explored otherapplications in this electronic system out of high performance computing. Thiselectronic system simplifies difficulties for the internal and external website buildingand maintenance, save the service charges; internal collaboration platform strengthenthe communication within research group, and achieve a variety of ways to sharemutual aid; project management system achieve the real-time tracking the projectschedule, viewing the project log in multi-platform, assigning tasks to projectmembers dynamically and so on.