By Joe Clabby, Clabby Analytics
Last October, I wrote a report on the progress of the OpenPOWER Foundation, an organization dedicated to expanding the ecosystem that surrounds IBM’s POWER8 and future generations of POWER processors. In that report I described IBM’s plan to make processor specifications, firmware and related software available to the evolving OpenPOWER community – essentially making POWER an open standard microprocessor. I also described the rapid growth taking place in the OpenPOWER ecosystem. Since its founding in late 2013 by IBM, Google, NVIDIA, Tyan and Mellanox, OpenPOWER has grown to over 110 member organizations.
Last week I attended the first annual OpenPOWER Summit (in San Jose, California) which was organized as part of NVIDIA’s GTC 2015 conference and learned that OpenPOWER Foundation members are delivering several new POWER8-based platforms, and supporting IBM’s CAPI and field programmable gate array (FPGA) adapters.
Notable new POWER8 system designs include:
- TYAN’s TN71-BP012 POWER8 server;
- Cirrascale 1S Power8/NVIDIA GPU Server;
- ChuangHE 1U server; and,
The IBM S824L POWER8/NVIDIA GPU Server.Early prototypes of future POWER8 servers included:
- A RackSpace POWER8 OpenCompute form factor planar (board); and,
- A Google Power8 planar.Additional, other OpenPOWER Foundation members showed new coherence attached processor interface (CAPI) and FPGA adapters.
A Closer Look at the New POWER8-based Systems Designs
There’s a lot more going on at the OpenPOWER Foundation than enabling ecosystem vendors to make alternative-to-IBM-Power-Systems POWER-based servers. There is a tremendous amount of innovation taking place in terms of server design and acceleration, as well as a lot of activity taking place with alternative business models.
At the Summit, several OpenPOWER Foundation members took to the stage to describe their product offerings and their go-to-market plans. Here are some insights that I garnered from their presentations:
- I was most impressed with a server discussion by Teamsun presented by Mao Qiu Yin and Zhiqiang Tian on the topic of Trusted Computing Applied in OpenPOWER Linux. Both presenters candidly and unabashedly stated that the reason that Teamsun has adopted POWER8 microprocessors is to create a “trusted” server environment for Chinese markets. As I listened more closely, I inferred that some Chinese technologists and politicians are suspicious of United States computer hardware and software products – and want a more open view into how these products are designed. The OpenPOWER Foundation provides that view and that helps assuage fears that servers have been “tampered with”.
- Another impressive description of product and market plans for POWER8-based servers came from Rackspace – a major player in the x86 server/cloud marketplace. Rackspace’s Aaron Sullivan presented his company’s view on the need to incorporate POWER8-based systems into an open cloud environment (in order to address future scaling requirements) – and he followed his POWER8 server discussion with projections on future cloud software and infrastructure development. My take-away from Sullivan’s pitch was this: Rackspace is firmly committed to integrating POWER8-based servers into its future cloud offerings – a move that I believe other cloud service providers (like Microsoft and Amazon) will eventually have to mimic if they wish to stay competitive with Rackspace from a price and headroom perspective. (Note: like Rackspace, Google has also identified POWER as an important architecture – and is working on its own POWER8-based solutions).
- Several vendors showcased POWER8-based solutions with integrated NVIDIA graphical processing units (GPUs) – including Cirrascale with its 1S Power8/NVIDIA GPU Server, and IBM with its S824L POWER8/NVIDIA GPU Server. There is a major server redesign taking place in the computing community (I’ve been writing about this shift since 2011) – and now I’m seeing some big steps forward in POWER-based “accelerated systems.”Other highlights at the Summit included presentations by Alterix (a company that now makes an FPGA for POWER environments); Mellanox (a maker of fast adapter/switching environments that exploit the remote direct memory access protocol – described in this Clabby Analytics report); and an NVIDIA discussion about its Tesla Accelerated Computing Platform. Further, PMC Sierra discussed using NVM Express technology (an optimized, high performance, scalable host controller interface) to accelerate FLASH/solid state drive performance – as well as using POWER8’s coherence attached processor interface (CAPI) to accelerate data center application performance. Finally, presentations from Canonical (an Unbuntu Linux associate); by compiler company PGI; and by Xilinx (a discussion of key-value stores) covered what is going on in the application design space within the OpenPOWER Foundation.
As noted above, Clabby Analytics has been covering the evolution of new types of server designs since 2011. What we are seeing are designs that employ several different types of microprocessors in order to exponentially improve computing performance. We are also seeing faster busses and the introduction of new interfaces that speed data delivery to/from these processors. Further, we’re seeing great improvements in networking speed and reduced latency, thanks to the use of protocols such as remote distributed memory access (RDMA).
Some of the design changes in server hardware that are leading to greatly accelerated performance include the use of GPUs, CPUs and/or FPGAs within the same server environments. In this 2011 white paper we described an appliance known as the IBM DB2 Analytics Accelerator that uses field programmable gate arrays (FPGAs) and Intel processors to offload complex query processing from mainframes. In October, 2013, we profiled a vendor by the name of VelociData that uses processors to accelerate communications (delivering data at line speed to workload processors); that uses graphical processing units (GPUs) to accelerate parallel processing; and that uses x86-based processors to process serial aspects of a given workload.
That same month, we also described a processing environment known as “The Now Factory” that uses FPGAs to feed internal Intel processors that handle parallel and serial tasks. (During the OpenPOWER Summit I was allowed to visit the GPU Technical Conference [hosted at the same site] where I saw several POWER/NVIDIA accelerated server models – as well as several x86-based NVIDIA-based accelerated servers. Accelerated systems with GPUs and FPGAs are finally becoming mainstream in commercial server design).
As for networking improvements, in November, 2014, we described in this report how new interfaces are streamlining FLASH memory and storage interactions with processors in this report – paving the way for comparatively lower cost in-memory FLASH configurations. In February, 2015, we wrote a report that described how a RDMA is being used to facilitate data transfer from the memory in one computer to the memory of another at high speed with little central processing unit (CPU) involvement. We also noted in this report that, as a result of the increasing use of this protocol, enterprises are now buying faster network switches in order to execute certain workloads more quickly (Melanox, an OpenPOWER founding member and leader in high-speed switches confirmed this trend). So accelerated network performance is also becoming mainstream.
As for streamlined interfaces between storage and the CPU, we wrote in this report about IBM’s CAPI interface for POWER8 which breaks down barriers between the CPU and underlying hardware. That makes it possible for large banks of FLASH memory or coprocessors such as FPGA’s and GPUs to interact in a low-overhead manner with POWER8 processors.
All of this acceleration activity in server design is leading to the creation of ever more powerful, ever more scalable server environments. And many of the components to build accelerated POWER8-based systems designs were shown and/or discussed at the OpenPOWER Foundation Summit.
One theme that I constantly heard at the Summit was “ultimately, technology is interesting – but solutions are more important.” One of the first lessons that I learned in the computing industry (as a sales rep back in the 1970s) was: “customers buy solutions.” So I agree with the presenters who emphasized that applications which perform functions users need are the real end game. I saw a few applications at the Summit that could take advantage of some of the advanced accelerated systems designs – but, at future Summits, I want to see many, many more.
In a way, however, I disagree with those who overemphasized the importance of applications. I still see the POWER8 processor as one of the most powerful processors in the industry (see this report for why) – and with the accelerators being developed in the OpenPOWER ecosystem, it is going to be hard for other processors (including Intel and ARM silicon) to match some of the acceleration speeds that will be achieved by POWER8 and successive generations. So yes – applications are key; but don’t forget that underlying infrastructure can also deliver Quality-of-Service and performance advantages that cannot be achieved using other types of platforms.