The Road to TSUBAME and Beyond
The Road to TSUBAME and Beyond
复制标题
通往 TSUBAME 及超越之路
DOI：
10.1007/978-3-540-74384-2_19
复制
发表时间：
2008
期刊：
影响因子：
--
通讯作者：
S. Matsuoka
中科院分区：
文献类型：
--
作者：
S. Matsuoka
自1994年Tom Sterling和Don Becker提出“Beowulf”式的商用PC集群以来，已经过去了12年[12]。第一个集群(称为“Wigraf”)的性能只有几十兆浮点运算(MFLOPS)，但已经取得了巨大的进步，最快的系统达到了10万亿浮点运算(TFLOPS)或更高。然而，与专门和专用的超级计算机如地球模拟器[3]相比，当时还不知道是否有可能通过商用PC集群扩展到10s到100s的数千个处理器，以便不仅在峰值性能方面，而且在操作的实际稳定性和广泛的适用性方面，达到100s的TFLOPS甚至Petaflop(PFLOPS)。Tsubame(东京工业超级计算机和无处不在的大容量存储环境)是2006年4月在日本东京的东京工业大学安装的新的超级计算机集群，拥有超过85TFLOPS的峰值计算能力和加速能力，21.7TB内存，1.6PB的在线磁盘存储空间，以及基于传统超级计算机的Fat Node和快速并行互连架构原理。Tsubame于2006年4月投入使用，在峰值性能、内存和存储容量等方面成为亚洲速度最快、规模最大的超级计算机。同时，TSUBAME基于PC架构，也可以看作是一个大型的PC集群服务器，能够
It has been 12 years since Tom Sterling and Don Becker proposed the “Beowulf”-style commodity PC cluster in 1994 [12]. The performance of the first cluster ever built (called “Wigraf”) was a mere few tens of megaflops (MFLOPS), but enormous progress has been made, with the fastest systems reaching 10 teraflops (TFLOPS) or beyond. However, compared to specialized and dedicated supercomputers such as the Earth Simulator [3], it was unknown then whether scaling of 10s to 100s of thousands of processors with commodity PC clusters would be possible in order to attain 100s of TFLOPS and even petaflops (PFLOPS), not just in terms of peak performance but actual stability in operation as well as wide applicability. TSUBAME (Tokyo-Tech Supercomputer and Ubiquitously Accessible Massstorage Environment) is a new supercomputer cluster installed at Tokyo Institute of Technology in Tokyo, Japan in April 2006, boasting over 85 TFLOPS of peak compute power with acceleration, 21.7 terabytes of memory, 1.6 petabytes of online disk storage, and “Fat Node” as well as fast parallel interconnect—architectural principles based on traditional supercomputers. TSUBAME became the fastest and largest supercomputer in Asia in terms of peak performance, memory and storage capacity, etc., when it became operational in April 2006. At the same time, being PC-architecture-based, TSUBAME can also be regarded as a large PC cluster server, being able to