ARM port of FreeNAS planned??

[Ericloewe]

That's not a pony. Unless it's a unicorn pony.

 

[jgreco]

That's not a pony. Unless it's a unicorn pony.

It's a Skittles factory, fool.

 

[Ericloewe]

It's a Skittles factory, fool.

That tends to explain why those things make me nauseous. When I was 10 or so, I actually tried those things. I'd never realized that candy could suck so badly.

 

[Bernard Mentink]

http://www.nextplatform.com/2016/04/28/arm-server-chips-xeon-class-matter/

 

[Ericloewe]

http://www.nextplatform.com/2016/04/28/arm-server-chips-xeon-class-matter/

That addresses nothing.

Any monkey can stick 32 of one thing in one place and add eight more of the other thing. That does not mean it's useful on a real NAS workload.

Currently, there are only two uArches that can hope to compete with Intel: AMD's Zen (unproven) and Apple's Swift (completely different target market) - and Apple's solution has been mostly to throw silicon at the problem, whereas Intel and AMD aren't growing their die sizes significantly.

Plain ol' ARM is tuned for completely different workloads.

The ARM server market is mostly handwaving bullshit. Why do you think the big cloud guys are gobbling up Xeon D? It hits the low power market without being a ridiculous construct with 32 cores that needs eight separate memory controllers and it integrates proven platform features such as Intel 10GbE.

And really, if you need 32 cores, the right product is what was formerly known as Larrabee or a GPU.

 

[jgreco]

The ARM server market is mostly handwaving bullshit. Why do you think the big cloud guys are gobbling up Xeon D? It hits the low power market without being a ridiculous construct with 32 cores that needs eight separate memory controllers and it integrates proven platform features such as Intel 10GbE.

Okay, this is where I think it gets complex. I agree with @Ericloewe about why Xeon D is popular. There's a lot of desire for familiarity within the IT/cloud world.

We're at (and have been at) a stage in "cloud" development where a lot of what's running is no longer Microsoft Windows. People are writing in scripted languages, Python, Java, Node.JS, PHP/XHP, whatever. If you can compile the interpreter/VM, guess what, no one gives a crap what silicon's running underneath. Often that's also true for a compiled language (C, C++, whatever). Hypervisors are a vaguely has-been technology as we move on to containers. Big monolithic CPU has been giving way over the last 20 years to better design by programmers that allow things to be broken up into many bits and pieces to be spread across multiple cores or even an entire network of machines. These things are all things that can work in favor of ARM.

An ARM processor with 64 cores or more could conceivably be a really big deal. Intel's stuff is "good", but the ARM stuff is more efficient overall. It COULD be a killer if only it could get a foothold. But in order to get a foothold, you'd need some serious adoption, and then you'd need to see it actually available on commodity server boards, and we're just not anywhere near seeing this, as far as I can tell.

Either Intel has to fall behind or ARM would need to make a massive improvement of a sort that just doesn't seem to be forthcoming at this time. Until then, it seems like no one's really interested in being an early adopter of a technology that just doesn't really offer much benefit in its current iteration,

 

[Bernard Mentink]

This old news(2014) is interesting too, as HP was one of those "early adopters", they claimed 35% reduction in overall cost of ownership with their ARM servers, that's not insignificant.
http://www.zdnet.com/article/hp-launches-first-enterprise-class-arm-based-server/

and
http://www.nextplatform.com/2016/02/12/arm-servers-qualcomm-maps-out-datacenter-battle-plan/
http://www.zdnet.com/article/windows-server-with-arm-crazier-ideas-have-happened/#!

 

[jgreco]

This old news(2014) is interesting too, as HP was one of those "early adopters", they claimed 35% reduction in overall cost of ownership with their ARM servers, that's not insignificant.
http://www.zdnet.com/article/hp-launches-first-enterprise-class-arm-based-server/

Sure, but the reality is that if this was going to reshape the world, it's had about two years to do so, and doesn't seem to have. Instead, Intel released Xeon D.

Look, Intel's in many ways the VHS of the CPU world. It isn't winning as much on merit or technical excellence as it is on volume and standardization. Look at the history. When ARM started to look like it might be a threat, suddenly Intel made a massive leap forward and introduced the Xeon D. Data centers wanted lower power and a certain ratio of cores-to-GB. And Intel's big enough that they can keep up pricing pressure on competitors. Those prices on the Xeons? Do some math and think about it a little. Especially in the Xeon E5 line, you can see that Intel's actually trying to price things to squeeze more out of performance users.

Claims of TCO reduction when introducing a new platform are often convoluted and may optimistically dismiss certain factors. Getting an actual 35% reduction in TCO is significant, but only where you've factored in more than capex and power as opex as part of the "total." As it stands, if I can't get a decent hypervisor that lets me run FreeBSD VM's, it's pretty much a nonstarter for me. Most of the FreeNAS boxes here do not run standalone. The smallest one's got 17 VM's on it. One of the bigger ones, 76. I get cost reduction through virtualization. So far I can only do that credibly on x86/x64. I don't see VMware for ARM out there. Going to an ARM platform without a 100% rock solid hypervisor means increased TCO for me.

Hell, I had to banish all our Opteron hypervisors to their own cluster because there's no way for them to play nice with the Intels. You might be able to get a TCO reduction if you've got a greenfield deployment where you can commit to a homogeneous environment with your own custom software stack, but that's basically Google or Facebook doing a new room in a data center. Adoption by us "small guys" is probably a long way off.

and
http://www.nextplatform.com/2016/02/12/arm-servers-qualcomm-maps-out-datacenter-battle-plan/
http://www.zdnet.com/article/windows-server-with-arm-crazier-ideas-have-happened/#!

Yeah. Well there's been a lot of talk for a lot of years. Smooth-Stone was doing this back in 2008 with their 5 watt cores and still here we are today, years later, no significant progress seems to have been made, there aren't tons of ARM server boards floating around. If one of these companies can actually deliver on something that Intel can't easily match or beat, we might have something interesting going on. Competition is certainly good.

But the way I see it, ARM is the boy who cried wolf for the last 8 years.

 

[jgreco]

And in the news today, Intel throws in the towel on Atom.

http://www.pcworld.com/article/3063...-tablet-markets-after-cutting-atom-chips.html

Which we could take as a signal that it is just as hard for Intel to horn in on ARM's massively successful smartphone and tablet markets as it is for ARM to horn in on Intel's territory.

 

[Ericloewe]

And in the news today, Intel throws in the towel on Atom.

http://www.pcworld.com/article/3063...-tablet-markets-after-cutting-atom-chips.html

Which we could take as a signal that it is just as hard for Intel to horn in on ARM's massively successful smartphone and tablet markets as it is for ARM to horn in on Intel's territory.

I had been wondering what had happened to their grandiose plans for Atom, with the yearly die shrinks. Looks like the answer is "seriously downsized".

 

[jgreco]

I had been wondering what had happened to their grandiose plans for Atom, with the yearly die shrinks. Looks like the answer is "seriously downsized".

I think the writing's been on the wall for awhile, with the sudden screeching halt of Avoton and the introduction of Xeon D that fills a similar niche. I think I even talked about Intel seeming conflicted here awhile back.

So who wants to take bets on how long it'll be before they try to get tablets to adopt the low core count Xeon D parts ... I have to say, I'm not a huge Intel fan, so after the pasting they gave AMD in the late 2000's, I would find it ... amusing ... to see them try something like that, and then have them flail about for another five years.

Xeon D can't compete with ARM at the current pricing though.

 

[Ericloewe]

I think the writing's been on the wall for awhile, with the sudden screeching halt of Avoton and the introduction of Xeon D that fills a similar niche. I think I even talked about Intel seeming conflicted here awhile back.

So who wants to take bets on how long it'll be before they try to get tablets to adopt the low core count Xeon D parts ... I have to say, I'm not a huge Intel fan, so after the pasting they gave AMD in the late 2000's, I would find it ... amusing ... to see them try something like that, and then have them flail about for another five years.

Xeon D can't compete with ARM at the current pricing though.

They kinda already have a Xeon D for small devices, it's Core M. Same uArch, but produced on the power-tuned version of their current process.

 

[c32767a]

Apologies if this turns out to be troll kibble, but.. I didn't see anyone mention it yet..

I'm always amused when ARM vs x86 comes up in the context of servers that move a lot of data.. I have yet to see an ARM implementation that can support the plumbing necessary to build a platform with good I/O capacity.. I see arm chips boasting 8 PCIe v3 lanes vs 40 lanes on Broadwell-EP.

Suitability is measured by more than just ghz and threads.

 

[jgreco]

Apologies if this turns out to be troll kibble, but.. I didn't see anyone mention it yet..

I'm always amused when ARM vs x86 comes up in the context of servers that move a lot of data.. I have yet to see an ARM implementation that can support the plumbing necessary to build a platform with good I/O capacity.. I see arm chips boasting 8 PCIe v3 lanes vs 40 lanes on Broadwell-EP.

Suitability is measured by more than just ghz and threads.

In general, a NAS doesn't make use of that much. PCIe 3.0 is 8Gbps per lane, which means you could potentially have a pair of 10Gbps ethernet and disk I/O and still be pretty good even on just 8 PCIe lanes. Since most NAS platforms are still on 1GbE this seems like a lesser issue.

From a design perspective, though, the Xeon D integrates the two 10GbE and has a lot more:



And the linked article does a pretty good job of explaining why the Xeon D is a real ARM killer. "The market climate just grew quite a bit more hostile for ARM-based server SoCs, which will have to justify themselves against a much more formidable x86-based incumbent."

And that's true. Aside from the price, this is ultimately something that could kill off even the E3 Xeons. What's the point of high TDP CPU's when you can get something gorgeous like this?

At this point, it isn't clear that this is the ultimate solution for a FreeNAS system. FreeNAS does very well with higher clock speeds, and the Xeon D exchanges more cores for a lower clock. However, ARM was never trying to go up against the high clock Xeons. Down here, the only advantage ARM is really likely to have is cost, and it looks like Intel may be preparing to go to battle on that front as well. With Xeon D going up to 16 cores and 32 threads, there are lots of options.

 

[Bernard Mentink]

You guys might want to have a look at: http://www.linleygroup.com/cms_builder/uploads/x-gene-3-white-paper-final.pdf
The x-gen-3 stomps all over Xeon-D especially with 5x memory bandwidth .... yep ARM really is dead ;)

 

[jgreco]

You guys might want to have a look at: http://www.linleygroup.com/cms_builder/uploads/x-gene-3-white-paper-final.pdf
The x-gen-3 stomps all over Xeon-D especially with 5x memory bandwidth .... yep ARM really is dead ;)

So crap out a readily available motherboard that has usable hardware and I'm sure someone might consider a port FreeNAS to it. Until that happens, and it shows no signs of happening, this is basically geek wankery of the VHS-vs-Betamax variety.

"5x memory bandwidth" is totally meaningless for a NAS anyways. Give me twice the memory at half the bandwidth and I'm twice as happy.

As for "stomps all over Xeon-D", even the linked article doesn't seem to believe that... "X-Gene 3 is on track to surpass Xeon D and even many Xeon E5 processors in per-socket throughput while delivering respectable per-thread performance" ... but even with that having been said, the reality is less optimistic. If you look at the TDP numbers, a very different story is told.

Anyways, there's no ARM port of FreeNAS planned because there's nothing realistic to run it on.

 

[Ericloewe]

Yeah, I'm not going to believe that an unknown team scored a major design hit with their microarchitecture that puts it close to Skylake. Unless they actually sell a real product, that is.

There are three companies out there that can do that: AMD, Apple and Intel. And I'm not so sure about AMD and Apple is in a whole different market.

 

[Mirfster]

Reminds me of "Solar Freakin Roadways!" I will believe it when I see a truly working production model I can buy. ;)

 

[INCSlayer]

Reminds me of "Solar Freakin Roadways!" I will believe it when I see a truly working production model I can buy. ;)

wow i remember those they looked soo cool

 

[Ericloewe]

Reminds me of "Solar Freakin Roadways!" I will believe it when I see a truly working production model I can buy. ;)

Nice analogy.
https://www.youtube.com/watch?v=obS6TUVSZds

With numbers!