The Institute for Molecular Science (IMS) is Japan’s common research facility into molecular science. Containing the “Research Center for Computational Science” it provides high-scale and super –high-speed arithmetic processing and library functions to researchers world-wide. Operation of a new “Super high-speed molecular simulator” has commenced consisting of a tightly coupled “computational server sub-system” and “high- speed I/O computing server sub-system”. The computing server sub-system has achieved a theoretical peak performance of 4 TFLOPS which is the largest scale HPC system so far configured using PRIMEQUEST.
IMS was founded as a collaborative research institution for researchers belonging to public research institutions and universities across Japan. Since its founding 25 years ago it has been actively used as the center for molecular science research and has contributed to the development of new technologies based on fundamental materials science. This includes research into the formation and changes to molecules, interaction of molecules and light, plus energy conversion at a molecular level.
The high scale and super high-speed arithmetic processing functions and library of the “Research Center for Computational Science (Pic.1),” operated by IMS provide a testing ground for molecular science researchers and bio science researchers across Japan. Major servers within center are connected by gigabit Ethernet (max. 8 circuits) and in some cases 10 gigabit Ethernet and can be directly accessed by researcher from their own research laboratories using the academic information network “Super SINET.” As at April, 2006, over 500 users in 120 research organizations make use of this network. Users are selected by public submission with “504 researchers using the network and 380 articles published in 2005.”(Mr.Okazaki, Professor, Research Center for Computational Science, Pic.2)
One concept of the IMS is, “Not only to offer the computing environment but also play a central role in the academic development of computational science related to materials.”(Mr.Okazaki). In 2003, a national “Super high-speed computer network formation project” was initiated by the Ministry of Education, Culture, Sports, Science and Technology. The IMS was made responsible for “Nano-science verification research” based on grid computing, theoretical molecular science research, and computational science research. From 2006, they commenced the grand challenge of research into Frontier/High performance super computer usage/development in the Nano-technology field. IMS is operated as the core research center with a major role in the organizing of university and institutional research groups across Japan.
The Research Center for Computational Science basically replaces its computer systems every 5 years and was considering replacement of their legacy system. In 2004, they started planning the introduction of their next-generation super computer system (a super high-speed molecular simulator). The following year, they issued the specification and commended evaluating bids. On Dec.27, 2005 as the result of this evaluation Fujitsu was accepted as the system vendor.
Mr.Okazaki says, “We adopted Fujitsu’s proposal as a result of the total score of its performance, versatility and reliability. We assumed it enabled the highest performance”
“We introduced the new system to support and develop molecular science and bio science computing. Basically speaking, higher performance is better for us. Since high versatility is required in a super high-speed molecule simulator, we added the requirement of highest performance system, based on experience with and past record of the configuration and operation of high performance parallel computing machines. To achieve the highest performance within the budget was another important issue for us. During the bid evaluation process, given the from high performance and cost performance points of view, we decided that Fujitsu’s proposal, with its Intel™ Itanium™ 2 processor 2 server, had the best performance and cost efficiency.”(Mr.Okazaki)
The “Super high-speed molecular simulator” introduced by the Research Center for Computational Science for the new system consists of a tight coupling of a computing server sub-system configured using PRIMEQUEST and a high speed I/O calculation server sub-system configured using a super computer “Altix 4700”(Japan SGI). Linux is used as the OS on both of these sub-systems. The theoretical peak performance is 4TFLOPS respectively and 8TFLOPS overall. Memory capacity is 10.5TB and HDD capacity 160TB. The system commenced operation on July 1, 2006.
“Our legacy system which had operated since 2000 had a performance of 0.5TFLOPS. Compared to this, the new super high-speed molecular simulator increased the performance 16 times.” (Mr.Morita, associate professor, Research Center for Computational Science, Pic.3) Mr.Morita also says, “Performance was dramatically improved by the migration from super computer (vector style CPU) to Intel Itanium server (scalar style CPU). As the new system will be used by many different researchers, general versatility is also important. Research Center for Computational Science systems are widely used in molecule and bio science, so a wide range of calculation processes are required. These include quantum theory, molecular dynamics, electron theory, statistical mechanics, chemical kinetics etc. These applications are now in process of migration to scalar processor use. We think it is valuable for us to have scalable servers and a computing environment at the laboratory level.”
Mr.Mizutani (Technical Staff, Research Center for Computational Science, Pic.4) says, “It was lucky for us that dual core Itanium 2 was released when we were introducing the new system. Users were able to enjoy the doubled performance. We think improvements in CPU performance are the great advantage of scalar SMP (Symmetrical Multi-Processing). Special techniques were required to achieve full performance in vector style parallel systems but the new system doesn’t require such special techniques for computing performance. This is a great advantage of the new system.
The tight coupled computing server sub-system is configured using 10 PRIMEQUEST server nodes, each with 32CPU/64 core,, providing a full spectrum total of 640 cores. It is currently the largest scale HPC system configured using PRIMEQUEST. For each node a high speed crossbar is used to connect between system boards and enable high-speed SMP parallel processing. A 16GB/s network is used to connect between nodes and large scale hybrid high-speed parallel computing, using a combination of distributed alignment and SMP alignment, is available.
The high speed I/O computing server sub-system based on Altix 4700 also has dual core Itanium 2, 512 core+128 core(total 640 cores), plus 6TB of shared memory available. This system was also provided by Fujitsu together with the PRIMEQUEST.
In addition, a 24TB ETERNUS 3000 disk array and 3-node PRIMEQUEST were introduced for data archiving and communication operations.
On the separate uses of the system, Mr.Morita says, “For jobs which operate using many simultaneous calculations such as molecular dynamics (MD), the tight coupled computing server sub-system configured with PRIMQUEST, which is superior in high speed SMP alignment operation and distributed alignment, is applicable. For highly accurate molecular orbital (MO) calculation jobs, which require large memory space, the high speed I/O computing server sub-system configured with Altix, with its large common memory up to 6TB and I/O function equivalent to memory bandwidth, is applicable.”
Mr.Mizutani says, “In operational alignment of multiple processes using PRIMEQUEST, the high-speed crossbar solves the problem of performance degradation in operation between processes. Also, since PRIMEQUEST supports Infiniband, communication speeds of 16GB/s are assured. It has an advantage in high speed crossbar throughput. Fujitsu seems to have achieved high performance by utilizing their technology and skill in Infiniband clustering with PRIMEQUEST.”
In 2006, the Research Center for Computational Science collected issues from the public and 120 of them were allowed to use the center. “These issues are classified as Issue A, Issue B and Issue S, according to their computing scale. Issue S was founded when the super high-speed molecular simulator started operation. In 2006, 3 issues were adopted as Issue S. General issues use some portion of the super high-speed simulator CPU for a few days. Those of class Issue S fully utilize both PRIMEQUEST and Altix. We are thinking to use them for a month without suspension.” (Mr.Okazaki)
“It was impossible to do calculations which required high loads with the previous environment. The super high-speed molecular simulator enables us to solve both the higher level and a wider range of issues. We believe this will expand the possibilities in molecular science computing and will also contribute to the development of academic research.”(Mr.Okazaki)
Deel deze pagina via