很抱歉,此 PDF 仅供下载

Shared Memory Standards Scale Up

Executive Summary
Shared-memory systems gave way to clusters in high performance computing (HPC) because of the perceived price/performance benefits of industry-standard processors and operating environment, leaving behind proprietary architecture that scaled well but which were based on more expensive RISC processors (or later, Itanium) and customized UNIX implementations. This cost savings came at a cost: the loss of shared-memory programming capabilities and performance benefits in the wholesale movement to MPI on distributed-memory clusters. SGI, a long-time architecture differentiator, has kept true to that path through significant upheavals in the company’s fortune, and now the vision is coming full circle with Altix UV, a scalable shared-memory model based on Intel® Xeon® processors and Linux. Some of the features of Altix UV are:
• Scalability to 2,048 cores and 16TB of globally shared memory, the architectural limit of Intel® Xeon® processor E7 family.
• NUMAlink 5 interconnect and hub, Intel QPI, and MPI Offload Engine provide an environment for scaling MPI codes in a shared-memory environment.
• Support for numerous topologies and programming models, allowing end users to suit the computer to the application.
Under new management, SGI has focused its roadmaps and remains committed to HPC, and the company is seeing an increase in its adoption rates. Ultimately SGI’s fortunes with Altix UV will be tied to how far people are willing to push their codes and their imaginations. MARKET DYNAMICS The Emergence of Clusters – Lower Cost, at a Cost In the relentless pursuit of increased performance for a given budget, the HPC industry has undergone periodic technology disruptions. Over a ten-year period from the mid-1990’s to the mid-2000’s, HPC buyers made a large-scale shift away from symmetric multi-processing (SMP) systems based on RISC microprocessors and proprietary UNIX systems, to more loosely coupled clustered architectures made up of independent nodes of commercial-off-the-shelf (COTS) servers running x86 microprocessors and Linux. The industry-standard x86 processors were produced at high volume for lower cost than RISC, and most Linux distributions were free, resulting in superior price/performance.
Read the full Shared Memory Standards White Paper.

Shared Memory Standards Scale Up

Executive Summary
Shared-memory systems gave way to clusters in high performance computing (HPC) because of the perceived price/performance benefits of industry-standard processors and operating environment, leaving behind proprietary architecture that scaled well but which were based on more expensive RISC processors (or later, Itanium) and customized UNIX implementations. This cost savings came at a cost: the loss of shared-memory programming capabilities and performance benefits in the wholesale movement to MPI on distributed-memory clusters. SGI, a long-time architecture differentiator, has kept true to that path through significant upheavals in the company’s fortune, and now the vision is coming full circle with Altix UV, a scalable shared-memory model based on Intel® Xeon® processors and Linux. Some of the features of Altix UV are:
• Scalability to 2,048 cores and 16TB of globally shared memory, the architectural limit of Intel® Xeon® processor E7 family.
• NUMAlink 5 interconnect and hub, Intel QPI, and MPI Offload Engine provide an environment for scaling MPI codes in a shared-memory environment.
• Support for numerous topologies and programming models, allowing end users to suit the computer to the application.
Under new management, SGI has focused its roadmaps and remains committed to HPC, and the company is seeing an increase in its adoption rates. Ultimately SGI’s fortunes with Altix UV will be tied to how far people are willing to push their codes and their imaginations. MARKET DYNAMICS The Emergence of Clusters – Lower Cost, at a Cost In the relentless pursuit of increased performance for a given budget, the HPC industry has undergone periodic technology disruptions. Over a ten-year period from the mid-1990’s to the mid-2000’s, HPC buyers made a large-scale shift away from symmetric multi-processing (SMP) systems based on RISC microprocessors and proprietary UNIX systems, to more loosely coupled clustered architectures made up of independent nodes of commercial-off-the-shelf (COTS) servers running x86 microprocessors and Linux. The industry-standard x86 processors were produced at high volume for lower cost than RISC, and most Linux distributions were free, resulting in superior price/performance.
Read the full Shared Memory Standards White Paper.

相关视频