RIKEN's New Supercomputer Up and Running
Fujitsu today announced that it has recently completed the deployment of a new supercomputer for Japan's Institute of Physical and Chemical Research, known as RIKEN. The system began operations in August.
The new supercomputer system achieved a LINPACK(1) benchmark performance of 97.94 teraflops. In comparison to the PC cluster systems included in the June 2009 world TOP500 list(3), the system's performance ranks first in Japan.
The new system, called the RIKEN Integrated Cluster of Clusters (RICC), is a complex system comprised of three separate computing sub-systems, each with a specific purpose: a massively parallel PC cluster, a large capacity memory server, and multipurpose parallel PC cluster. In addition, each computing sub-system is connected to a common front-end system, and disk and tape systems.
The massively parallel PC cluster is comprised of 1,024 of Fujitsu's PRIMERGY RX200 S5 PC servers, with 2,048 CPUs (8,192 cores).
The supercomputer's operating environment includes Parallelnavi, Fujitsu's HPC middleware which integrates the system management functions, high-speed file system, and program-development environment. The system is also equipped with an HPC job-management tool, the "Meta-Job Scheduler," which enables integrated management of massively parallel jobs on multiple clusters as a whole.
LINPACK benchmark measurements produced the following result, based primarily on the 1,024-node massively parallel cluster of PRIMERGY RX200 S5 servers. This result places it first in performance among PC cluster systems currently in operation in Japan.
Performance: 97.94 teraflops
Rank on TOP500 list (June 2009): 34th worldwide
Rank among PC clusters on TOP500 list (June 2009): 1st in Japan
The benchmark test result can be attributed to a combination of factors: the high performance of hardware and HPC middleware Parallelnavi, as well as advanced system integration technologies such as program tuning.
Because Parallelnavi's math library dictates LINPACK's processing performance to a large extent, it played a significant role in contributing to the latest result. The math library, realizing world-class performance, leverages R&D achievements of Fujitsu Laboratories.
In addition to high-performance hardware, Fujitsu's strength also lies in its ability to offer middleware which incorporates its leading-edge R&D achievements, and provide comprehensive system integration support that flexibly responds to customers needs.
About "Meta-Job Scheduler", the HPC job-management tool
The "Meta-Job Scheduler" is a high-performance scheduler developed through collaboration between RIKEN, Fujitsu, and Fujitsu Laboratories in the course of developing and operating the previous RIKEN Super Combined Cluster (RSCC) system.
The following functions make it possible to eliminate the waste from scheduling of computation center users' multiple jobs as well as minimize job latency in order to maximize the system's utilization.
• Multi-layer scheduling: Scheduling operates at a highly elaborated level, managing computing resources at the core, CPU, server, and cluster layers.
• Fair share: The system refers to each user's assigned rank and past execution log to ensure that all jobs are scheduled in an equitable manner.
• Back file: This function reorders the sequence of standby jobs to make effective use of empty resources.
Source: Fujitsu
