Not logged in » Login
Jul 04 2016

HPC at the Next Level: FUJITSU Server PRIMERGY CX600 M1

/data/www/ctec-live/application/public/media/images/blogimages/FF-2015-12-08_PRIMERGY_CC-09.jpg

Until not too long ago, high-performance computing (HPC) was considered to be a discipline that mainly benefits academia, research, and large governmental or public institutions. But with Big Data and the Internet of Things (IoT) on the rise, more and more traditional industries – from automotive through healthcare and retail to utilities – have developed, and are still developing, a desire for server systems that will successfully tackle the expected surge of data and workloads. Fujitsu's new PRIMERGY CX600 M1, equipped with PRIMERGY CX1640 M1 server nodes based on Intel's latest-generation Xeon Phi™ processors (aka Knights Landing), was designed to fulfill their desire.

Introduced on June 20 at this year's ISC High Performance event, an annual meeting organized by the International Supercomputing Conference, the FUJITSU Server PRIMERGY CX600 M1 not only provides the massive parallel computing power the above-mentioned sectors need, but also offers a very attractive price/performance ratio. To achieve this, it relies on the freshest line of Intel's supercomputing processors labeled Xeon Phi™ 7200, commonly known as Knights Landing or KNL.

Intel® Xeon Phi™ 7200
First announced in 2013, KNL builds upon experience garnered from prototypes and predecessors code-named Knights Ferry and Knights Corner (aka Xeon Phi™ x100) and at the same time expands the product range: until now, the Xeon Phi™ brand denoted a line of co-processors that had to be fitted into server systems in order to take over massive parallel computing tasks that the regular CPU couldn't handle effectively. In other words, Intel-based HPC systems almost always had to house a mix of processors and co-processors. On the one hand, this led to increased purchase prices. On the other hand, the 'cohabitation' mandated a split between portions of code written for standard Xeons and portions written for the Xeon Phi™, thus creating artificial gaps that could thwart the performance of the entire system.

By contrast, the Xeon Phi™ 7200 is a full-fledged, bootable Intel processor that was optimized for highly parallel workloads so as to overcome the separation between the two types of code. At its core is an Intel® Atom™-based SoC with up to 72 cores of the Silvermont microarchitecture that has been enriched with numerous HPC enhancements – in particular, the ability to do some kind of 'double hyper-threading,' i.e. to process four threads per core simultaneously. However, this alone would not facilitate the necessary improvements, and so Intel's engineers also added MCDRAM, a new type of internal memory that really makes systems fly. MCDRAM resides directly inside the processor and provides up to 16 GB capacity, allowing for internal transfer rates of 500 GB/s. A fully configured server will thus be able to deliver up to 3 teraflops (3 trillion operations per second) – just what it takes to process extremely demanding and critical workloads in fields like genome sequencing, flight and wind tunnel simulation, financial computing or (big) data analytics. And since the Xeon Phi™ 7200 is 'backwards-compatible' with software customized for existing processor/co-processor configurations, nobody will have to start recoding. Likewise, software developers won't have to acquire new skills; instead, they may use the same programming models, tools and binaries as before, regardless of whether they specialized in parallel or serial computing.

FUJITSU Server PRIMERGY CX600 M1
The FUJITSU Server PRIMERGY CX600 M1, which also debuted at the ISC High Performance event, is among the first systems to wholly build on KNL's expanded capabilities. Like all PRIMERGY Scale-out models, it consists of two main components – the eponymous PRIMERGY CX600 M1 chassis, which provides a 2U housing, and the matching FUJITSU Server PRIMERGY CX1640 M1server nodes, a line of mono-socket, 1 U and half-wide systems powered by the new Xeon Phis (cf. Fig. 1). While the small size of the server nodes may come as a surprise to some, it does allow us to pack 8 of them into the chassis, which results in an ultra-dense server cluster whose unprecedented parallel performance and highly competitive price/performance ratio will make it a must-have for organizations that need to branch out into HPC.

Image

Fig. 1: 1U, half-wide, ultra-dense – the PRIMERGY CX1640 M 1 server node

The FUJITSU Server PRIMERGY CX1640 M1 offers a plethora of configuration options, enabling customers to order tailor-made systems that exactly fit their needs. At launch, they can choose between the following components:

  • Two types of Intel® Xeon Phi™ processors – the Xeon Phi™ 7210 is equipped with 64 cores that will process up to 256 threads and operates at a clock speed of 1.3 GHz. The slightly more powerful Xeon Phi™ 7250 comes with 68 cores, which allow for 272 threads to be processed at 1.4 GHz. Both models use MCDRAM as a third-level cache that can be configured to accommodate a vast range of usage scenarios: when set to "cache mode," it serves as a regular L3 cache between the CPU and the server's DDR4 main memory, which is especially recommendable if a server has to handle large applications or data sets that won't fit into MCDRAM alone, or in case its memory access behavior is unknown or unstructured. The "flat mode" is for use cases that require peak performance; here, the administrator can basically define which parts of an application or data set must be kept in the MCDRAM for fast reuse. In "hybrid mode" one portion of the MCDRAM runs in cache and the other in flat mode, which is ideal for mixed scenarios. (72-core Xeon Phi™ processors will be available as soon as Intel ships them, presumably in fall.)
  • DDR4 memory modules with 8, 16 or 32 GB capacity – aside from MCDRAM, the PRIMERGY CX1640 M1 also needs regular memory bars to be able to serve the outside world. At launch, we are offering registered DIMMs that support ECC (namely, single-device data correction/Chipkill™) and connect to a 2,400 MHz memory bus; LR-DIMMs are to follow soon. Regarding memory capacity, the new server scales from 16 to 384 GB.
  • 2.5 inch HDDs or SSDs, alternatively SATA DOM for internal storage – like most HPC systems, the PRIMERGY CX1640 M1 relies on external storage systems, so the internal capacity is fairly limited and amounts to one HDD or SATA DOM or two SSDs. Moreover, the choice of storage media depends on whether customers want an air-cooled or liquid-cooled edition of the new system; the former supports all types and form factors, the latter only SATA DOM.

In accordance with these specifications, a fully-configured FUJITSU Server PRIMERGY CX600 M1 Scale-out system currently packs up to 408 cores, 128 GB of MCDRAM and 3 TB of main memory into a 2U chassis, thus achieving a level of compute density that has no precedent in the industry.

Image

Fig. 2: Provides shared infrastructure for eight servers – the PRIMERGY CX600 M 1 chassis

As noted above, Fujitsu packs eight of these systems into the PRIMERGY CX600 M1 Server Enclosure, a 2U, multi-node, 19-inch chassis with four front and four rear bays that give admins easy access to the hot-pluggable server nodes (cf. Fig. 2). Like all our server enclosures, it provides the 'shared infrastructure' components for the nodes that come with it. The most important among these are four 1200W hot-plug PSUs that offer 94% power efficiency (80 PLUS Platinum rating) and are configured in a 3 + 1 redundancy setup – meaning that if one PSU fails, the reserve unit will take over until the regular one has been replaced. Also on board are five fans that come in a 4 + 1 redundancy configuration and either do all the cooling by themselves or support the optional direct-to-chip liquid cooling modules that reside on the server nodes and capture up to 75% of the server heat. Thanks to this advanced cooling technology, the FUJITSU Server PRIMERGY CX600 M1 Scale-out system will operate within an extended ambient temperature range of 5–40 °C (41–104 °F).

Target Markets and Use Cases
So what's the use of ultra-dense server systems equipped with Intel® Xeon Phi™ 7200 processors? Or more to the point, in which scenarios would customers deploy the FUJITSU Server PRIMERGY CX600 M1? Supporting both massive parallelization and high throughput, these systems were built to shoulder the more demanding variety of HPC tasks. As one would expect, this includes the area of scientific research (e.g. weather modeling, fluid dynamics, VR etc.), which previously often had to rely on customized (or rather: souped-up) versions of general-purpose server CPUs that were only available in limited quantities and at a hefty price. These users may turn to attractively priced off-the-shelf hardware. At the same time, the new processors and systems will bring a massive performance boost to areas that so far had to resort to classic systems, such as business intelligence and analytics (BI/BA), 3D computer graphics, or product development/CAx. For example, applications like LS-DYNA, Nastran or ANSYS Fluent that already delivered good results on 'normal' Xeons will now come up with excellent ones without the need to rewrite one line of code.

Conclusion
The FUJITSU Server PRIMERGY CX600 M1 is available for order immediately under project release. Pricing varies by configuration and region. To learn more, please see the product pages for the system and the server node. To get a feel for what the PRIMERGY CX600 is capable of, check out the video below.

 

Florian Frimmel

 

About the Author:

Florian Frimmel

Global Product Marketing Manager, responsible for PRIMERGY tower and scale-out servers, at Fujitsu 

SHARE

Comments on this article

No comments yet.

Please Login to leave a comment.

X

Please login

Please log in with your Fujitsu Partner Account.

Login


» Forgot password

Register now

If you do not have a Fujitsu Partner Account, please register for a new account.

» Register now