19 November 2013

RSC PetaStream solution sets a world record of computing density of 1 PFLOPS per rack and shows a practical path to create Exascale level supercomputers

Denver (USA), SC’13, November 19th, 2013 — RSC Group, the leading Russian developer and integrator of innovative HPC and data center solutions, demonstrated revolutionary RSC PetaStream ultra-high density solution with direct liquid cooling at SC’13 conference in Denver, USA (booth #404). RSC PetaStream sets a world record of computing density of 1 PFLOPS peak performance per rack with just 1 m2, or 11 ft2, footprint, which is 5x times greater than the previous record of 211 TFLOPS per rack that was set by RSC Tornado supercomputer architecture earlier this year. One RSC PetaStream rack contains 1024 computing nodes and has capability to efficiently remove over 400 kW of power using RSC’s direct liquid cooling technology. This value is 4x times greater than the previous RSC’s achievement and sets a new world record of power density.

“We managed to achieve such a breakthrough due to unique extensive experience of RSC specialists in development of extra efficient direct liquid cooling technologies and ultra-high density integration of supercomputer solutions based on standard server components and technologies,” – said Alexey Shmelev, COO of RSC Group.

RSC PetaStream solution is an innovative implementation of massively parallel architecture using the best available industrial components, highly dense integration and unique system architecture delivered by the RSC engineers. Every system node is based on Intel® Xeon Phi5120D coprocessor featuring 60 cores and 8GB of high-bandwidth GDDR5 memory. All 1024 nodes are interconnected using high speed Infiniband FDR network, and altogether providing revolutionary high-density HPC solution capable to process over 250,000 execution threads in a single rack based on x86 architecture. This enabled a significant step ahead to create Exascale level supercomputers that will have millions of executable threads in a single application. With RSC PetaStream systems the users and developers already today can develop and optimize applications for future massively parallel systems with many-core processors.

“Efficient access to computing is the foundation of innovation and discovery,” –said Raj Hazra, Vice President Data Center Group and General Manager Technical Computing at Intel Corporation. “The coming era of computing will require efficiently deploying parallel software on energy-efficient, parallel hardware based on Intel Xeon processors and Intel Xeon Phi coprocessors. RSC’s PetaStream solution meets those requirements and allows customers to protect their software and hardware investments as they march towards the Exascale era.” 

RSC PetaStream solution is based on x86 architecture that is a de-facto a global industrial standard and guarantees investment protection in software development and optimization. Developers of new software for future computing systems will be able to use existing programming models and run existing applications on Exascale level supercomputers. At the same time, the flexibility of RSC PetaStream architecture allows testing of new innovative programming models that may become more applicable in the Exascale era.

Several HPC applications were showcased running on the new system during presentation at the SC13 exhibition. In addition to great application performance delivered by RSC PetaStream, it demonstrated 2x lower amount of energy required to complete highly-parallel simulation jobs comparing to traditional cluster systems. It creates clear value in reducing TCO, while delivering leading performance in small footprint.

To ensure maximum density, high I/O throughput, reliability and manageability multiple nodes inside RSC PetaStream system are grouped into a module that provides liquid cooling, highly efficient power delivery and conversion, and mechanical assembly of the nodes into a computing rack. RSC PetaStream module  consists of  up to eight nodes based on Intel® Xeon Phiproducts, I/O subsystem including Intel® Server Board S1600JP with Intel® Xeon® E5-2600 v2 processor, up to five solid-state drives Intel® SSD of S3500 or S3700 series, along with tightly integrated network adapters, such as Mellanox Connect-IB adapter cards or Intel® True Scale host adapters.

In addition to 1024 Intel® Xeon Phi based compute nodes, there are up to 640 solid state drives can be installed in every rack. These drives managed by scalable parallel file system, and guarantee high number of IO operations delivered per second to feed many cores in the system, as well as provide scalable capacity of local storage: when using 800 GB Intel® SSD DC S3700 or DC S3500 series drives every rack has total local file storage capacity of 0.5 PB for the system with performance of 1 PFLOPS. Intel® SSD DC S3500 and DC S3700 series drives meet the highest requirement to storage system redundancy and capacity. These drives are de-facto industry standard providing high performance and guaranteed 5 year life cycle on 10-times daily rewriting. They also provide fast response time and high QoS for query processing.

New RSC solution supports a broad choice of interconnect types and enables expansion using various network card types, as well as allows choice of network topologies and broad options for building innovative data storage system within a supercomputer. RSC PetaStream module has two PCIe Gen3 x16 slots for standard and non-standard interface cards, that, for example, would allow up to 4 Infiniband FDR or QDR ports, in addition to integrated for Gigabit-Ethernet adapter ports. Each module provides high throughput into external network that exceeds 200 Gbps when Mellanox Connect-IB dual-port adapter is used.

RSC’s direct liquid cooling technology has been used by RSC customers in many projects since 2009. It has proven energy and power efficiency that is inherited by the new RSC PetaStream solution. As it was measured in several HPC sites running RSC systems, the average PUE value reaches 1.06, which means that only 6% of total power is used for system cooling. PUE value is becoming critical for Exascale level supercomputers that use many megawatts of power.

Record computational density of 1 PFLOPS in ca. 2.2 m3 or 80 ft3 volume demands not only high performance cooling system, but also energy and cost effective power delivery and distribution subsystem. This is why RSC PetaStream uses high efficiency power distribution subsystem based on leading industrial 400V DC power supply standard and products delivered by Emerson Electric. The energy efficiency exceeds 90% guaranteeing improvement of overall system energy efficiency even further reducing TCO. Also, 400V DC solution eliminates excessive AC-DC conversion stages and requires fewer components, resulting into energy savings and improvement of entire reliability of the solution.

For user convenience RSC PetaStream provides extended hardware and software monitoring and management capabilities, based on RSC BasIS - an integrated software stack. The system uses Intel® Node Manager Technology to control and monitor node power consumption of every node, and to implement flexible power energy and optimization strategies to help HPC sites save power and reduce operational costs.

RSC PetaStream solution is supplied with Intel® Cluster Studio XE 2013 software kit for Linux operating system, as well as a set of optimized libraries for the RSC PetaStream system.