Microservers are targeted at light workloads such as serving static HTML content. These small servers generally have a power consumption of below 45W and can be densely packed to scale up capacity and meet demand.
Early ARM-based microserver boards have been based around its 32-bit Cortex-A9, such as the Calxeda EnergyCore boards used in HP's microserver test platform in its Project Moonshot initiative.
ARM's entry proper into server space will begin towards the end of this year when AppliedMicro releases the X-Gene platform, the 64-bit ARM-based microserver board.
But when 64-bit ARM-based microservers arrive they will be competing with similar offerings from Intel. Intel launched its 64-bit Atom Centerton system-on-a-chip (SoC) board — which combines CPU, memory and system I/O on a single integrated circuit — into the microserver market late last year.
By the time the X-Gene reaches the datacentre it will most likely be up against servers based on Centerton's successor, Avoton, which Intel says will combine enterprise features such as the Error Correction Code (ECC) memory support already found in Centerton with a wider range of I/O functionality built into its SoC.
ARM's uphill struggle
One of the biggest hurdles that ARM has to overcome in order to capture the microserver market, according to Rich Fichera, principal analyst serving infrastructure and operations professionals for Forrester Research, is to increase the size of the server software stack for its platform.
"There's not a lot of software for them, it's going to take a few years to build out the ecosystem," he said.
"My gut feeling is if ARM is less than twice the power per watt of the x86 chip, it's not going to catch on" — Rich Fichera, Forrester
"In the short term if you want to make a lot of money selling servers, you've got to have an x86 server," he said.
ARM has to demonstrate a clear performance to power consumption lead if it is to persuade organisations to go through the pain of switching to a new server hardware platform, rewriting software and the challenge of managing a heterogeneous server estate based on both x86 and ARM Risc chip architectures, said Fichera.
"My gut feeling is if ARM is less than twice the power per watt of the x86 chip, it's not going to catch on. It needs a big advantage. It needs to be three times the throughput per watt or better to make a big impact," he said.
Performance wise, HP reports tests under Project Moonshot have demonstrated relatively similar performance per watt between microservers based on ARM Cortex A9 Calxeda EnergyCore SoCs and Intel's Centerton boards in a range of areas.
"We've done workload comparisons in areas like a Hadoop server, caching server, video streaming," said Dave Chalmers, chief technologist for HP for EMEA.
"For generic applications like Hadoop they are very similar in their performance levels. There are nuances of difference to the workloads and we do a lot of work with customers to work out which of the processors is suited to them."
Power consumption for ARM-based server implementations, such as Calxeda EnergyCore with its thermal design power of 5W, is reported to be below that of the lowest power Centertons. However both ARM and Intel dislike the comparison as they say it is not like for like, with ARM citing features missing from the Centerton's SoC, such as the Ethernet controller, and Intel the difference between the Centerton's 64-bit and the Cortex A9's 32-bit architecture.
Server software support for ARM's Risc architecture is improving. Common server software stacks like LAMP and OpenStack have been run on ARM hardware, as have applications for the Hadoop distributed storage and compute system.
However the target workloads for ARM-based microservers are unlikely be the day-to-day operations of big enterprise, due to the performance of microservers for these kind of workloads and companies' reliance on third-party packaged software designed to run on traditional x86, rather than ARM Risc-based servers.
The Facebook effect
A prime market for ARM microservers could be the web giants — the likes of Facebook and Google, as they not only have vast numbers of web-facing servers but they also already run bespoke software and hardware stacks, so have the in-house skills to port code to work on ARM's Risc architecture.
"An architecture such as ARM has the potential to radically increase the amount of compute power we can get from the energy we consume and the money we spend" — Frank Frankovsky, Facebook
"If you're Facebook, you just take your Linux source off the web and you build your software stack yourself. These companies have a lot of software resources themselves and all of that code is written in high-level languages so it is pretty portable," said Ian Ferguson, VP of segment marketing at ARM.
Facebook has proven open to customising its datacentre design, from the server chassis and rack designs through to its computing hardware and software. Not only does the firm run a largely custom software stack in its datacentres, it also launched the Open Compute Project to crowd-source ideas for server chassis, rack and computing hardware designs that will allow it to run its datacentres more efficiently.
Speaking at the Open Compute Summit earlier this year Frank Frankovsky, chairman of the Open Compute Foundation and vice president of hardware design and supply chain at Facebook, praised the potential benefits that ARM-based servers could bring to the social network.
"An alternative processor architecture such as ARM, coupled with open-source software, has the potential to radically increase the amount of compute power we can get from the energy we consume and the money we spend," he said.
But it doesn't appear that large web companies are contemplating a wholesale switch to microservers, which will limit the size of the opportunity for ARM. Jason Waxman, general manager in Intel's datacentre and connected systems group, said that, while these companies were interested in microservers, their overriding demand is for improved efficiency among higher-powered server processors; those sitting in the 95W-130W range.
"Across the large number of datacentres that sit at big web scale, far and away the propensity over the past couple of years has been to deliver more performance for that power envelope versus doing something radical like cutting power usage by 90 percent and taking a hit in terms of performance," he said.
The changing face of the server market
The traditional way for a chip maker to get their hardware into datacentres was through OEMs like HP or Dell. Intel says it has signed deals with 15 different OEMs to distribute servers based on its low-power chips, which also include its low-power Xeons.
HP has committed to using both ARM and Intel-based boards inside microservers developed under Moonshot, and Dell last year began offering its Copper servers, based on the Marvell Technologies Armada XP processor.
"Certainly a lot of the time spent by my team is with Facebook and the end users" — Ian Ferguson, ARM
From ARM's perspective the shift towards large web companies buying their own custom-built hardware means the likes of Facebook are as important as the HPs of the world in determining which platforms will be successful in the microserver market.
"We're working with HP, they are the market leader in servers by volume, but we do see new opportunity out there. Certainly a lot of the time spent by my team is with Facebook and the end users. It's really those guys that have to be convinced that their software is going to run better on ARM than on x86," said Ferguson.
"The choice of who's going to provide the hardware is going to be made by the likes of Amazon, eBay, Facebook and Tencent. That's why it makes sense for us to work directly with those guys."