July 21, 2013
This week’s handpicked assortment of HPC cloud research initiatives include a special spotlight on MapReduce, where more HPC research on data intensive applications is seemingly published every week. Further, we take a look at how EC2 systems perform against benchmarks, the always important issue of security, and implementing GPUs in the cloud.
High Performance Computing applications are scientific applications that require significant CPU capabilities. They are also, according to researchers from the University of North Florida, data-intensive applications requiring large data storage.
While many researchers have examined the performance of Amazon’s EC2 platform across some HPC benchmarks, an extensive study and their comparison between Amazon’s EC2 and Microsoft’s Windows Azure is largely missing with metrics such as memory bandwidth, I/O performance, and communication and computational performance.
The purpose of their paper was to fill that missing gap and implement existing benchmarks to evaluate and analyze these metrics for EC2 and Windows Azure that span both Infrastructure-as-a-Service and Platform-as-a-Service types.
They accomplished this by running MPI versions of STREAM, Interleaved or Random (IOR) and NAS Parallel (NPB) benchmarks on small and medium instance types. In addition, they also included a new EC2 medium instance type (m1.medium) in the analysis. These benchmarks measured the memory bandwidth, I/O performance, communication and computational performance.
Distributed and parallel computing have emerged as a well developed field in computer science, argues research out of Indiana University.
Tak-Lon Wu of the School of Informatics and Computing wrote a paper on how cloud computing offers new approaches for commercial and scientific applications because it incorporates different perspectives for the use of both hardware and software.
According to Wu, MapReduce distributed data processing architecture, instantiating the new paradigm of “bring the computation to data” has become a popular solution for large scale scientific applications, such as data/text-mining, bioinformatics sequence alignment, Matrix Multiplication, etc.
To understand whether the MapReduce computing model applies to these data-intensive analytic problems, he explored several problems by analyzing their usage for different MapReduce platforms in HPC Cloud environments. He mainly reviewed the state-of-the-art MapReduce systems for scientific applications, along with summarizing research issues found in prior studies.
Modern operating systems continue to be the victims of attacks and information leaks. Emerging architectures such as cloud computing or HPC are complex to set up and face many kinds of security threats.
However, according to research out of DAM Île-de-France and Laboratoire d'Informatique Fondamentale d'Orléans in France, they still rely on traditional access control mechanisms to protect the system and users' data, whereas these mechanisms can be misconfigured and easily defeated.
They presented a full architecture to enhance the protection of HPC clusters. Per their research, it provides three levels of access control in order to allow the users control over their files while enforcing advanced security properties.
More specifically, the integration of mandatory access control enables to control direct information flows, and a new and specific reference monitor deals with indirect information flows. In order to keep a low impact on operating system performances, they proposed to centralize this second reference monitor on a dedicated node, controlling the flows on all other nodes through the low latency network.
They presented the whole architecture and the results of several benchmarks that indicate a low impact on performances. Then they exposed how they made this architecture fault-tolerant. This study took advantage of previous works dealing with access control on workstations or virtualisation technologies, and extended the concepts to the HPC environment.
A growing need of cloud computing services makes it a more challenging field for cloud providers and researchers, according to a study produced by a team at Gujarat Technological University in India.
After a brief introduction of Xen hypervisor, their paper discussed the problem of inefficient access to graphics accelerators in the cloud. Some of the reasons behind such a problem include proprietary models and heterogeneity in architecture of graphics processors. Their paper focused on such problems of GPU (Graphical Processing Unit) acceleration by providing a sample GPU passthrough model.
Graphics passthrough is a delicate task as a result of secretive architecture of graphics adapters, particularly those of NVDIA. Accomplishment of graphics virtualization via GPU passthrough enabled cloud for providing quality graphics services at very low cost.
Graphics card vendors are secretive in disclosing the architecture of graphics devices, making graphics acceleration specific to the graphics vendor.
Developing platforms for large scale data processing has been of great interest to scientists, noted Dr. Shuju Bai in a recent dissertation. Hadoop is a widely used computational platform which is a fault-tolerant distributed system for data storage due to HDFS (Hadoop Distributed File System) and performs fault-tolerant distributed data processing in parallel due to MapReduce framework.
It is quite often that actual computations require multiple MapReduce cycles, Bai argued, which needs chained MapReduce jobs. However, Design by Hadoop is poor in addressing problems with iterative structures.
In many iterative problems, Bai said, some invariant data is required by every MapReduce cycle. The same data is uploaded to the Hadoop file system in every MapReduce cycle, causing repeated data delivering and unnecessary time cost in transferring this data.
In addition, although Hadoop can process data in parallel, it does not support MPI in computing. In any Map/Reduce task, the computation must be serial, according to Bai. This results in inefficient scientific computations wrapped in Map/Reduce tasks because the computation cannot be distributed over a Hadoop cluster, especially a Hadoop cluster on a traditional high performance computing cluster.
Computational technologies have been extensively investigated to be applied into many application domains. Since the presence of Hadoop, scientists have applied the MapReduce framework to biological sciences, chemistry, medical sciences, and other areas to efficiently process huge data sets.
In their research, they proposed a hybrid framework of iterative MapReduce and MPI for molecular dynamics applications. They carried out molecular dynamics simulations with the implemented hybrid framework.
Further, they improved the capability and performance of Hadoop by adding a MPI module to Hadoop. The MPI module enabled Hadoop to monitor and manage the resources of Hadoop cluster so that computations incurred in Map/Reduce tasks can be performed in a parallel manner. They also applied the local caching mechanism to avoid data delivery redundancy to make the computing more efficient.
Their hybrid framework inherits features of Hadoop and improved computing efficiency of Hadoop. The targeting application domain of their research is molecular dynamics simulation. However, the potential use of their iterative MapReduce framework with MPI is broad. It can be used by any applications that contain single or multiple MapReduce iterations, invoke serial or parallel (MPI) computations in Map phase or Reduce phase of Hadoop.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?