NCKU Develops Switchless HPC Cluster
The National Cheng Kung University (NCKU) Supercomputing Research Center (RSC), based in Tainan City, Taiwan, has teamed up with Acer to build what it claims is the first ever switchless HPC cluster. The “CK-Star” computer, as it’s known, connects eight computational nodes without employing a switch control, exceeding previous efforts by Intel, according to the university.
NCKU RSC officials unveiled the record-breaking CK-Star flexible topology switchless cluster along with a high-density node GS-R22PHL supercomputer, developed in tandem with GIGABYTE Technology, on November 18, 2013.
CK-Star touts eight Acer Altos R380 F2 servers with Intel Xeon E5-2630 V2 processors and Intel Xeon Phi 7120P coprocessors to achieve 15 teraflops. The cluster’s unique setup enabled it to achieve 80.2 percent efficiency for four nodes and 77.3 percent for eight nodes, surpassing Intel’s previous record of 79.6 percent for four nodes and 76.1 percent for eight nodes.
The NCKU staff also helped develop Gigabyte’s GS-R22PHL supercomputing module, which packs 8 double slot GPU/MIC cards in a 2U rack and supports the Intel Xeon processor E5-2600 & 2600 V2 product families, a dual 10G BASE-T and a dual 10G SFP+ LAN controllers, up to 512GB of ECC memory, and 8 x 2.5” hot-swappable HDD bays. The result, according to GIGABYTE, is unprecedented computing density.
When equipped with four Intel Xeon Phi coprocessors, the GS-R22PHL can achieve 3.7 teraflops, the highest computing power for a single-node server and comparable to 10 standard servers, according to NCKU. With one Phi coprocessor, the system achieved 83.3 percent computer performance, beating Intel’s previous record of 79.8 percent. When two coprocessors were used, performance was around 78.2 percent.
NCKU RSC Director Dr. Chi-Chuan Hwang explains that where traditional cluster supercomputing requires switches to control interaction between nodes, the CK-Star was built without such switches. There are a number of advantages to the innovative approach, which employs “daisy-chain design and intelligent networking.” First off, the CK-Star does not have to contend with performance bottlenecks caused by switches at high node counts. The design is also more power-efficient since switches can eat into power consumption.*
Mainland China’s National Supercomputing Center in Jinan (NSCCJN) “Taishan Scholar” Dr. Yuefan Deng, who along with Dr. Hwang, helped develop both of these systems, explained that the network topology of CK-Star can be tuned for specific application requirements to achieve the better performance.
Dr. Hwang believes that with a developed software ecosystem, there will be high demand for high-density computer systems from commercial interests. Going forward NCKU and Acer will focus on software integration and effective use of resources, while the NCKU-GIGABYTE team will emphasize hardware design and development.
*Editor’s Note: The original version of this article stated incorrectly that switches can consume 50 percent of the total power draw, a claim that comes from the official press release. Thanks to @HPC_Guru for pointing this out.