This Week in HPC News
It’s been another packed week in high performance computing, with a bevy of new partnership, supercomputing installations and news about coming systems hitting our radar.
Interestingly, despite some of the top news items summarized below, one of the most fascinating stories this week (at least in terms of the massive numbers of views/listens) is the podcast interview with a certain mysterious John Fitzpatrick, who claims to have $50 billion lined up for his exascale-class supercomputer that he’ll be opening for currency trading (and donated time for science) in Oregon. Hard to hide the skepticism during the interview, but this couldn’t be ignored in case it actually (somehow) happens. Oh, and by the way, he says the system will be up and running in 2014. So there’s that. (Yes, that’s what I thought too)…
If you’re in the habit of listening, we try to keep things a bit more grounded. However, one more on the (more realistic) speculative technology side to consider is our interview with Dr. Larry Smarr, who talks about everything from new materials, quantum computers and the exascale systems of the future. Other topics included adapting Cray machines for whole genome analysis at Argonne, networks for powering climate research, architectural considerations for astrophysics applications, and more. A fun week, for sure.
Several news items to cover in no particular order…
A new collaboration of Oak Ridge, Argonne and Livermore (CORAL) will seek to develop systems in the 2017-2018 timeframe to support the research missions at their respective institutions.
A joint Request for Proposals for the CORAL procurement was issued Jan. 6 and responses were submitted Feb. 18. These are now being evaluated. The intention is that CORAL partners will select two different vendors and procure a total of three systems, two from one vendor and one from the other. Livermore is leading the procurement process.
Livermore’s system, to be called Sierra, will be best suited to support the applications critical to stockpile stewardship. Oak Ridge and Argonne will employ systems that meet the needs of their DOE Office of Science missions under the Advanced Scientific Computing Research (ASCR) program. Vendors are submitting test clusters now.
A3Cube Comes out of Stealth
The company emerged this week with its ‘brain inspired’ data plane encapsulated in a Network Interface Card (NIC) aimed at transforming storage networking to eliminate the I/O performance gap between CPU power and data access performance for HPC, Big Data and data center applications.
The RONNIEE Express data plane profoundly elevates PCI Express from a simple interconnect to a new intelligent network fabric, leveraging the ubiquity and standardization of PCIe while solving its inherent performance bottlenecks. A3CUBE’s In-Memory Network technology, for the first time, allows direct shared non-coherent global memory across the entire network, enabling global communication based on shared memory segments and direct load/store operations between the nodes. The result is the lowest possible latency, massive scalability and disruptive performance that is orders of magnitude beyond the capabilities of today’s network technologies including, Ethernet, InfiniBand and Fibre Channel.
We spoke with the company’s Emilio Billi, CTO and founder of San Jose-based startup, A3Cube, who has picked up a thing or two over the last twenty years of addressing a range of performance bottlenecks at the storage and network levels. In addition to developing the HiDRA “personal supercomputing” and companion code, he helped develop the HyperTransport Consortium’s HyperShare scalable network technology and remains one of the leads behind that effort.
With his new company out of stealth and rushing headlong into a well-established storage and network ecosystem to serve the needs of both HPC and demanding big data environments, he admits that they’re up against some challenges. However, he makes the argument that A3Cube’s technology, which uses PCIe as the interconnect via an enhanced NIC card, can alter the price, performance, and programmability of modern HPC and data-intensive systems.
Billi says that five years ago, when he began work on A3Cube’s host of technology, he was looking for a way to combine storage, compute and fit this within the massively parallel analytics software that’s coming. As he explained, doing this demanded the creation of “a 3D torus network interconnection data plane (it’s more of a data plane than an interconnection network) that has all the characteristics of supercomputing fabric but was designed specifically to create a massively parallel storage architecture.”
The argument is that storage systems need to take the leap from a few standalone engines to thousands of individual storage devices running in parallel to address the needs of true scale-out storage. This is managed by the PCIe-based approach they call the RONNIEE in-memory network.
As he describes, this is a completely new paradigm for networks that provides a whole application with transparent memory-to-memory direct connections. This “in-memory network discards the protocol stack bottleneck involve in remote memory access, which cuts the latency down even for user-level software.” The key is that the TCP/UPD stack is snatched from view and replaced with their own memory-to-memory mapped TCP/UDP socket as the performance hinge. He says it’s still possible to use RDMA if desired, but they’re adding to ease of programmability by the abstraction.
In other news…
Maxeler and the Science and Technology Facilities Council (STFC) and are collaborating in a project funded by the UK Department of Business Innovation and Skills to install the next generation of supercomputing technology in a new facility at the Daresbury Laboratory focusing on energy efficient supercomputing and offering orders of magnitude improvement in performance and efficiency to enable UK industry to have the edge in using a technology designed for the move towards Exascale computing.
The dataflow supercomputer will feature Maxeler developed MPC-X nodes capable of an equivalent 8.52TFLOPs per 1U and 8.97 GFLOPs/Watt, a performance per Watt that tops the Green500 today. MPC-X nodes build on the previous generation technology from Maxeler deployed at JP Morgan where real-time risk computation equivalent to 12000 x86 cores was achieved in 40U of dataflow engines. The new MPC-X supercomputer will be available in Summer 2014 and will focus on medical imaging and healthcare data analytics, manufacturing, industrial microscopy, large scale simulations, security, real-time operations risk, and media/entertainment.
Nova Southeastern University’s (NSU) Graduate School of Computer and Information Sciences has received a multimillion dollar IBM Supercomputer that will place NSU’s research at the forefront of computational biology, data mining, graphic visualization and software engineering.
Each of the 32 nodes will sport 16 Power CPU’s with 256 GB of RAM. Each CPU has two processor units that can run two threads each. The machine is water-cooled using internal chilled plates and a rear cooling door on each rack. The software stack consists of AIX, General Parallel File System (GPFS), C++, Fortran, IBM Parallel Environment Runtime (PE), Engineering and Scientific Subroutine Library (ESSL), Parallel Engineering and Scientific Subroutine Library (PESSL), and Tivoli Workload Scheduler LoadLeveler. The university says it will use the system for scientific projects as well as to help train the next generation of HPC skilled graduates.
ScaleMP updates vSMP with Version 5.5 and uses the word “configurator” in a sentence, which is awesome. In addition, new pricing has been announced.
vSMP Foundation Version 5.5 is currently available for download or purchase. Highlights of the latest version include, in addition to the “an open, online configurator”–
- Enhanced hardware support
- Broader IO support options with AnyIO subsystem. With AnyIO, customers can enable aggregation with almost any device:
- Any network device, such as 10GigE cards as well as Infiniband cards
- PCI-flash devices such as Fusion-IO
- GPUs/accelerators such as Intel Xeon Phi or NVIDIA GPUs
- Enhanced private interconnect options
- Mellanox Connect-IB
- Improved performance for Intel TrueScale
- Support for recent Intel Xeon Processors: Ivy Bridge – E5-2600v2 and E7-x8xx v2
- Enhanced performance for IO-intensive and large-memory workloads
- New flexible pricing model allowing lower price points
For the Presenter in You…
Thanks again for tuning in this week—back again Monday with more podcasts, announcements and in-depth features.