Open source collaboration has worked wonders to refine and shape intellectual property for ARM Holdings and their RISC reference designs. IBM hopes to capture some of that success and reinvigorate its Power processor architecture by joining forces with other industry leaders via the OpenPOWER Consortium.
Continue Reading This Article
Enjoy this article as well as all of our content, including E-Guides, news, tips and more.
Can open source data center systems boost IBM Power processors?
Some industry observers are skeptical and say the open source route is akin to abandoning customers without directly killing a failing product line.
"At the moment a product or solution is almost dead, and the owner of the technology doesn't want to dedicate resources to its development anymore, it is open sourced ... which too often means, 'Dear customer, go help yourself,'" said Sander van Vugt, Linux author, trainer, consultant and SearchDataCenter contributor.
Open source projects have an uphill battle, particularly in the physical hardware world, said John Treadway, Cloud Technology Partners' products and software leader, who is also a SearchDataCenter contributor.
"Rarely is a declining platform rescued from the abyss by such [initiatives as this consortium]," Treadway said. "The IBM Power Architecture, like all RISC-based systems, has been losing relevance for years, and it's unlikely that any open source move will garner more than a niche response. This is particularly true with OpenPower, which lacks an open governance model on the core technology."
However, enterprise IT teams like RISC-based systems for specific tasks or appliances, and Google's involvement in the OpenPOWER Consortium suggests that the IBM roadmap for its Power Architecture is as a collaborative alternative to traditional x86 servers from the likes of Intel and AMD.
OpenPOWER Consortium duties, limitations
In the consortium, IBM retains full control of the Power Architecture and the Power Instruction Set, similar to the ARM Holdings model -- so don't expect to see consortium members adding or deleting their own unique instructions.
A uniform architecture and instruction set form the processor core, avoiding member friction and prohibitive specialization. Members can add other devices such as graphics processing units (GPUs) and other specialized controllers to the chip, keeping the design homogeneous and interoperable while limiting interactions and patches between the chip design and OS.
The resulting systems must be more compelling to enterprise buyers than x86 and established RISC platforms based on ARM's Cortex-A57 design, such as AMD's Seattle RISC processor, shipping in 2014.
To start, the consortium can help IBM deliver its Power8 chip in late 2013 or early 2014, with the other vendors modifying IBM's processors and firmware. The Power8 offers 12 cores, each with eight threads, running at 4 GHz, 96 MB of L3 cache and an extension bus called the coherently allocated processor interface to interconnect graphics processors and other specialized devices.
In OpenPower, the GPU will come from Nvidia, and Mellanox will provide the networking components. Tyan designs and builds system motherboards and can translate newly developed components into hardware platforms for servers and other systems.
The resulting systems will run on open source OSes such as Linux.
Designing servers that data centers want
Google, the only consortium member that is a major consumer of servers and systems, would find open source designs with energy and performance advantages over traditional x86 servers extremely attractive. One can speculate that Google will -- at least initially -- drive the consortium's design goals.
Traditional processors focused on core tasks such as accessing memory contents or moving instructions through the processor's instruction pipeline, so performance upgrades and system-level server features came in at the motherboard level. These include graphics acceleration, disk or PCIe bus access, and other ancillary functions.
With denser chip-level transistors and more on-chip integration, processors can share a chip with GPUs, peripheral bus controllers (PCIe or even USB) and network interconnects. These system-on-a-chip designs are simpler and less expensive for server makers because there are fewer chips to connect and test externally on a motherboard.
However, there is a disconnect between what server vendors are developing and what data center teams want, said Doug Feltman, director of systems and applications at 24 Seven Inc., a staffing company with data centers in New York and California.
"It seems like they throw a lot of new features and functionalities out there to see what sticks," he said. "Often a new version of a hardware product comes out, and they've taken away the features that you thought were the best thing about the product in the last version."
What would data center directors want from open source servers?
"Power consumption is a new area that matters to us -- it never mattered much before," Feltman said. "I also look at new servers and try to find ones with features that I can exploit. There are features like the ability to boot from a USB that sound like a ... great way to reduce failures."
Meredith Courtemanche, site editor for SearchDataCenter, contributed to this report.