Job Ad Hints At Microsoft Move To ARM Servers 138
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."
Re:I don't know.. (Score:3, Insightful)
I wonder how true that quote actually is. I mean, what if you decide to run OSX Server, or order a shitload of iPads... or start buying Sparc boxen (they still make those, yanno...)
"Edge" systems? (Score:4, Insightful)
Here's a thought...most pure data retrieval tasks don't require a huge amount of compute power on the device making the request. If I were operating a datacenter with thousands of hits a second, I'd want to optimize for the ability to hold a session open, then offload the request to either a monster data layer or a midrange layer that brokers requests and caches frequent search results.
Something like a single-board computer (or a really scaled-up thin client :-) ) running a low-power processor dedicated to driving network interfaces that also have their own offloading processors would allow them to scale the access layer way up for less power costs. Reliability would be less of a concern too, because you could have tons of cheap devices for the same costs as a fraction of full servers.
When you scale out, you often don't need the overhead that full servers would give you, because you're limiting the tasks that layer of access has to do.
Or...they just want to see how many smartphones it would take to replace layer one of Bing. :-)
I'm waiting for the announcement of Windows Embedded CE 2011 Datacenter Edition.
"own the decision" (Score:5, Insightful)
MBA-ese for "take the blame"?
Re:So... (Score:5, Insightful)
NT was written [microsoft.com] as a portable OS from grounds up (remember that it had a working MIPS build before x86 build!), and much of that legacy still remains today in OS architectural design, so porting the OS itself shouldn't be hard. The toolchain (compiler etc) is already there to target ARM for CE.
Drivers (third-party ones specifically) might be trickier, though they're still mostly written in C, so for the most part it should be a straightforward recompile.
Re:Have to note as a big 'duh' (Score:3, Insightful)
The answer - to me - is pretty clear. The bottleneck to implementation is the software.
I'd rather have whichever one is more cost-effective, and I sincerely doubt it's going to be the ARM solution.
I'd like to see some evidence that an ARM CPU provides two orders of magnitude better processing power/watt than a Xeon CPU.
Then you might want to consider how much power consumption the order of magnitude greater supporting electronics (motherboards, RAM, switches, etc) is going to add to the ARM solution.
*Then* you might want to consider the cost of people to handle the additional administrative overhead in managing and order of magnitude more machines, and the additional physical space required.
It's not magical (Score:3, Insightful)
You don't just plop down a bunch of ARM processors on a board and magically get suitable performance without scaling out memory architecture and such. The only way in the x86 space that very many core systems get acceptable results is by increasingly sophisticated memory architectures that demand more memory modules in aggregate to allow direct, lightly loaded paths between compute and memory. Those memory modules draw more energy, as does various strategies that put more memory controllers down to lighten the load and more. Scaling general-purpose computing tasks to many small cores simply has some significant challenges that drive up the incremental power requirements as it goes up.
The most performance per watt in pure compute power is currently PowerXCell 8i, which doesn't exist outside of an IBM blade as far as I know. If a datacenter wanted to *really* be serious about performance per watt, I think I'd see more QS22s lying around. Intel is admittedly not the leader in performance-per-watt, but the crown still lies in systems optimized for high resource utilization in a small number of CPU packages.
Re:Have to note as a big 'duh' (Score:4, Insightful)
If you didn't want to run them on the same machine for performance reasons before, virtualization will only exacerbate the problem. Virtualization does not speed up anything, it does not magically make multitasking better (in fact, worse). Virtualization can be used for some security separation without having to think at all about it (I would prefer people do the thinking to make it ok to coexist in the same OS instance, the VM thinking just leads to effectively static linked binaries for everything, but practically speaking, people are lazy) or for running disparate OS apps (i.e. Windows and Linux) concurrently with reduced hardware.
Re:"own the decision" (Score:4, Insightful)