Job Ad Hints At Microsoft Move To ARM Servers

An anonymous reader passes along a brief EE Times note on a suggestive Microsoft job ad. ARM is explicitly mentioned, as are solid-state disk drives as an area of experimentation in the quest to reduce power consumption; but Intel does not get a mention. Here is the ad. "Microsoft is looking for senior software development engineer to help with its Bing data centers, potentially running them on ARM hardware, according to an EE Times article. Whoever gets the job 'can own the decision on the hardware that we use,' the job description said, and added that power management is a key aspect of the job. ... Microsoft was reportedly experimenting with the Intel Atom microprocessor in February 2009 with a view to creating a green low-power data center. One issue discussed then was the Atom microprocessor lacked performance compared with other Intel processors and that therefore any power saving might be negated by the need for more processors to carry a given computational load."

Comments

Product Experimentation

[statusbar]

Not necessarily dropping intel:

To provide sufficient server and networking capacity, the Autopilot Hardware team is involved in Data Center planning, new hardware expirementation including SSD and ARM

They are just doing expirementation (s.i.c.) !

--jeffk++

Maybe not for the server hardware itself

[BadAnalogyGuy]

ARM is severely underpowered, even when comparing to Atom. So it doesn't make *any* sense that MS is considering it as a server platform.

However, ARM excels at low power consumption and mobility. This would allow a new array of "server helper" devices that had needed quick handling of light tasks. Maybe something like packet routing or on the fly network topology auto-configuration. Another concept could be mobile cache points which would be somehow networked to the main servers and provide "smart caching" of data for light user requests.

Who knows. But to think that ARM is going to somehow best Intel's chips in the server market is crazy.

Re:Maybe not for the server hardware itself

[Grishnakh]

The important metric is MIPS/Watt. It doesn't matter if the CPU is "underpowered" if you can run a bunch of them in parallel to get the same performance as a Xeon, and still get better power consumption. Remember, the work they're doing is highly parallelizable, so outright clockspeed isn't very important.

However, I don't have any MIPS/Watt figures available for ARMs or Xeons, so this is idle speculation. If I were to take a guess, however, I'd guess that, given a real-world workload, the Xeons would probably beat the ARMs because of many factors: cache size, context-switching time, etc. If it were more economical to run a datacenter on tons of low-power ARMs, Google probably would have already done it by now.

As for craziness, remember, this is Microsoft we're talking about: the company that thought SongSmith would actually be a commercial success instead of a complete joke. Given their combination of big successes and utter failures, they seem to be quite neurotic.

Re:Maybe not for the server hardware itself

[Skapare]

Back when I was building a Linux distro for an ARM platform (specifically IXP435) I found it to have maybe about 1/3 of the power of x86 CPUs of the time, but running so cool that the CPU didn't need a heat sink and didn't get so hot I couldn't put my thumb on the CPU. And that was after running a regression test suite for 20 hours. ARM definitely is a win for the MIPS/Watt metric.

BTW, it would be scientifically simpler to just refer to this metric as "millions of instructions executed per joule of energy converted to heat" (would roughly equates to a gain of information in exchange for a loss of information).

Re:Maybe not for the server hardware itself

[ravyne]

Actually, current ARM processors are more than competitive with Intel's low-power offerings -- Arm Coretex A-8 cores have been benchmarked to match or exceed the performance of Intel Atom processors in integer performance while suffering a 25% clockspeed disadvantage, and while consuming around 1/20th the power. Floating point performance does indeed lag behind even Intel's Atom (its also one of the focus areas for the Cortex A-9 core), but is not a big requirement for server or database tasks.

There was some intersting reseach not too long ago which paired low-power x86 chips (Geode LX at the time, IIRC), Solid-state storage (in the form of compact flash) and a RAM-based caching of solid-state contents. About 10 of these boards were then clustered, running a distributed database application. As I recall, there were some serious performance and power-savings advantages compared to a single larger, multi-core x86 server setup. The primary advantage, in my oppinion, was that the ammount of available bandwidth, both to storage and to working RAM, combined with intelligent caching, was paired in much more favorable ratios to the power of the CPU. In short, the reseach found that a cluster of modest machines turns out to be competitive, even better-than, a single powerful server in terms of cost, power-consumption, heat disipation, and even in performance.

Microsoft is keen to realize that a modern ARM core is quite well suited to match modern I/O limitations. There's no point building a large system when the requests it's going to serve will be bottlenecked by I/O.

Have to note as a big 'duh'

[Junta]

One issue discussed then was the Atom microprocessor lacked performance compared with other Intel processors

Atom and ARM are great platforms when you don't need much processor in one spot. I.e. many embedded applications and a lot of consumer electronics. They need some processor, but not a lot. 'desktop/server' processors are optimized for a higher load and just don't scale down. Note that ARM isn't inherently low power, it's just the instruction set everyone in the world has rights to implement, and Intel pretty well dominated everything but an emerging low power market. You have a lot of innovation and skill at implementing 'just-enough' processors that simply picks ARM out of convenience.

In the data center context, things change. The notorious energy consumption of the low-power processors come to nothing when you can arbitrarily consolidate workload onto as few processors as possible. The economies of scale of the mainstream desktop/server platforms deliver are far greater than tiny low power devices.

In terms of MS experimenting with it, expect nothing to come of it. It will fail like Atom did in their experiment before. Assuming a long shot, expect nothing to change externally, even if MS discovers ARM is great for their data centers, they cannot readily win a market that centers around lower cost, lower energy, lower performance non-x86 compatible parts. They have a golden example of a company thinking their technology intrinsically drives the industry making a drastic change to discover they were wrong. Intel thought they dictated the terms of the industry, but Itanium simply failed to transform the market without quality x86 compatibility. This was the golden opportunity for AMD to swoop in with an alternative and make huge gains. MS is in the exact same situation, 99.9% of their clout is the environment of existing Windows apps. Microsoft has tried time after time various platforms to reach the same endgame of no success. If the new architecture in *theory* provided more performance, sufficient to emulate x86 instructions, then it would stand a remote chance, but going to lower performance platform renders this impossible. In a really long shot, MS gets a lot of really nice ARM hardware on the market, and then has to compete with Linux on its own merits rather than ecosystem of applications. It's nearly suicide to risk your largest leverage point unless the industry is imminently making you irrelevant even if you stick to your guns.

Re:So...

[fuzzyfuzzyfungus]

It does; but one of their mechanical engineers came to a rather brilliant realization:

A WinCE PDA is almost exactly the same size as a data tape. With modest modifications(consisting largely of forcing the work experience kid to run lots and lots of docking cables) an industry standard tape silo can be turned into a gigantic WinCE/ARM blade farm. If a node stops responding, the robot retrieval arm pops it out, presses the reset button, and pops it back in again. Since the OS is in ROM, boot is short and downtime is minimal.

Re:Netbooks!

[Entrope]

The OMAP3530 (which is in the IGEPv2) is a cool part in a lot of ways, but it would be boneheaded to put it in a data center. Because it doesn't have any high-speed interconnects -- gigabit Ethernet, PCI Express, RapidIO, or the like -- it isn't suited for most network-intensive applications. Marvell has a variety of systems-on-chips that do have ARM cores, running at higher speeds than the OMAP parts, and with high-speed interfaces on the chip. Other vendors probably have similar offerings; those are the ones that Microsoft would probably want to look at.