Microsoft turns to Altera silicon to host data-centre-virtualisation algorithms
The paper describes how the company relied on distributed reconfigurable fabrics (Altera’s Stratix V D5 FPGAs) to accelerate the processing of large amounts of data, more specifically by porting part of its search engine Bing across several FPGAs.
The reconfigurable fabric, dubbed Catapult, is embedded into each half-rack of 48 servers in the form of a small board with a medium-sized FPGA and local DRAM attached to each server.
"The performance requirements for today’s large data centre workloads are outstripping what general-purpose servers can provide, so we ran a pilot using Altera technology to deliver more acceleration than software running on servers alone," said Doug Burger, director of Client and Cloud Applications in Microsoft Research’s Technology division.
"We set a performance target that would be a significant throughput gain, while simultaneously permitting more advanced search ranking models to be run. Compared to a pure software implementation, our reconfigurable acceleration fabric permitted a 90 percent improvement in throughput at each ranking server, with great system stability."
Based on the results, Bing plans to roll out FPGA-accelerated servers to process customer searches in one of its data centres starting in early 2015.
Microsoft is relying heavily on Intel’s Xeon processor packs for its servers. And it is interesting to note that during the same week at the Gigaom Structure’14 conference, Diane Bryant, Intel’s data centre division lead engineer presented the future of Xeon processors with integrated FPGAs in a single package, socket compatible to the company’s standard Xeon E5 processors.
“The FPGA provides our customers a programmable, high performance coherent acceleration capability to turbo-charge their critical algorithms. And with down-the-wire re-programmability, the algorithms can be changed as new workloads emerge and compute demands fluctuate”, Bryant explained. Intel expects this new hardware combination could deliver over 20X performance gains compared to more traditional ASIC-based solutions.
“Our innovation extends beyond silicon as we collaborate with industry partners to accelerate the move to Sofware Defined Infrastructure. Our early work with Facebook to define a rack-scale architecture is gaining momentum with cloud service providers, telco service providers and hosters” wrote Bryant in her blog.
Now, would it be too much speculation to guess that Microsoft, a heavy weight in the cloud, is only one of many other possible customers for whom Intel would happily design custom chips, further pushing the balance in favour of Altera?
Mike Strickland, Senior Manager for Strategic and Technical Marketing at Altera didn’t want to confirm if indeed, Microsoft and Intel’s separate announcements could be interpreted as a “design-win in the making” for Altera’s FPGA fabric into Intel’s future Xeon processors.
“Intel’s announcement is an indication that people are interested in the concept of accelerating their data centres with FPGAs”, he conceded. “I think that in the next year or two, they would probably build 2.5D chips for a cost-effective integration” he added.
That falls in line with Intel and Altera’s joint announcement in March of a collaboration for the integration of 14nm Tri-Gate Stratix 10 FPGAs (manufactured at Intel’s fabs) with heterogeneous technologies into a single system-in-a-package.
“Altera’s work with Intel will enable the development of multi-die devices that efficiently integrates monolithic 14 nm Stratix 10 FPGAs and SoCs with other advanced components, which may include DRAM, SRAM, ASICs, processors and analogue components, in a single package.
The integration will be enabled through the use of high-performance heterogeneous multi-die interconnect technology. Altera’s heterogeneous multi-die devices provide the benefit of traditional 2.5 and 3D approaches with more favourable economic metrics” said the official statement.
In Altera’s view of the software defined data centre, reprogrammable FPGAs are set to play a central role, enabling a more flexible infrastructure with software defined allocation and prioritization of virtualized computing, networking, and storage resources.
In effect, each FPGA acts as a node controller, for different algorithms to be accelerated across different FPGAs as shared resources in a data centre.