Microsoft to use Hammer to combat Sun?

[JonWoodruff]

Coïncidence or not, CNet reports that

"Microsoft is building a high-end feature into Windows for speeding up data access in multiprocessor servers--a feature that to date has been available only in high-end Unix servers, the company said Tuesday. "
The Redmond, Wash.-based software company is working on support for a technology called non-uniform memory access, or NUMA, one method for designing large servers crammed with processors, said Sean McGrane, program manager for Microsoft's top-end Datacenter server.

That's from AcesHardware.com.

AMD's Hammer processors use a NUMA architecture for mult-processing, where each processor has it's own bank of memory and is connected to the other processors with their own banks of memory through a chain or network of point-to-point busses. Problem with that is that if all the memory appears as one big chunk to the OS (which it does), half the time, or a quarter of the time each processor will be fetching data through the other processor's memory controllers over the point-to-point network. (HyperTransport in Hammer's case)

AMD has said (noted on AcesHardware) that OS support for the NUMA archetecture was possible and even necissary, and if Microsoft is supporting AMD then they must be designing a "NUMA controllor" into the Windows memory interface. Why is this significant? Whith a very good design, which Microsoft will have the freedom to do, they will be able to keep the data for each processor's threads in it's own memory bank yeilding near perfect scaling all the way up to lots of processors.

Seems to me that this would be Microsoft's big oppurtunity to penetrate SUN's market for big multi-processor servers, if we don't count Itanium. It will also greatly benefit AMD in the corporate segment. As far as I know, Intel doesn't have anything that gives Microsoft so much scaleability and control over system archetecture. If Microsoft jumps on x86-64 and Hammer with NUMA big time, AMD will benefit big time.

 

[Saem]

RoFL!

One must love Ace's Hardware. The folks there do know quite a bit, but their knowledge dies down when you go outside the realm of x86.

This is as Mr. DeMone often calls it, "x86 arrogance." I don't see AMD or Microsoft putting out any competition against Sun's larger servers. No one I know considers Win2k to be comparable (reliability wise) against Solaris nor would they think that they that Hammer will compete against Sun's machines in reliability. I wonder how so many people think that just because AMD has great performance and they produce a good quality product for their price range they'll be able to take on the world. They certianly didn't do squat against Intel in the higher end market. And look at Intel, even though they have higher reliability and better QA, they still have trouble going far into the high end market. Some how people feel that AMD will be able to do better?

If MS wants to begin to touch Sun, this is best done with Itanium -even though I despise the architecture.

 

[Gubbi]

Hmm, IRC Chris Rijk of Aceshardware is involved with SUN.

And while Solaris is a stable OS, their hardware has a more checkered past. In particular the UltraSparc-1 fiasco, where you couldn't run 64bit apps, without compromising the system (you could whack the system from userland).

More recently their US-2 with 4mb cache didn't have ECC protection which caused *alot* of problems, in particular in their big E-10k.

AMD has built alot of error checking into the Hammer family. It has ECC on all cache levels. It has chipkill to shut down bad main memory, and it has background scrubbing of ECC protected main memory. So the hardware should be at least as reliable as SUN's.

You're right about Windows being a low reliability OS, but this haven't stopped IT managers around the world from adopting Windows yet. But unlike the Itanium, Hammer will be a mass market product, which makes it infinately more interesting for M$.

Cheers
Gubbi

 

[Saem]

Good points, Gubbi.

Just one thing, thought IT managers use Windows, I don't think they use them in the same rolls that one would typically put Large scale Sun machine's into. I think here we're talking about IBM, SUN and used to be Alpha MPUs but now Itanium. These aren't going to be running Winblowz. We're in Unix land and on the web server front we're in Linux territory. I believe there is a Unix coming out for Itanium, though I maybe wrong. Wasn't HP working on something along that line.

But what's more important is that the companies Like IBM and SUN and to add to the list HPaq and DELL either have their own offering and/or have an Itanium offering, along with this they have their own OS and most importantly their tech support.

I don't think any of those companies except for Dell is interested in letting MS getting into the market and having it's way with it. I'm pretty sure said companies don't want MS anywhere near the market and will goto great lengths to keep it that way.

It'd be nice if the Linux and associated communities got off their collective asses cleaned up the OS (perhaps standardize the GUI via Java) and then we might see a big improvement for Hammer, unfortunately, Intel has it's hand in this as well and will surely stake a claim for Itanium.

As for Itanium being slow, well Itanium 2 (McKinley) is no slouch, it's pretty darn fast, not to mention it's got the high quality chipsets (something AMD still doesn't have, nVidia, SiS, Via, ALi and even AMD don't ring of super high quality, no matter how much people wish to believe otherwise) and it has a lot of folks developing systems based on it. Not only that, I believe it'll be first out of the gates.

I tend to be of the opinion that the Hammers Claw and Sledge will compete with P4's and Xeon's, respectively.

Now, this is merely what I believe will happen. What I'd like to see happen and is mildly possible is that x86-64 becomes pervassive and EPIC doesn't get a strangle hold.

What I would absolutely love is if someone came up with a great open ISA and a good emulation technique for x86 so that we could move away from x86 poopieness. I don't believe x86-64 is the answer to that, since in the end, it's hack job, IMO.

 

[Gubbi]

Well, we are in agreement. I agree with you that Hammer won't enter the high end (or even the upper end of medium) size server market.

They will however, be completely unbeatable on price/performance in the 2-8 CPU segment. Getting Win XP-64 on these machines are M$ ticket to ride into the (higher end) server market *if* the Hammer platform proves it self.

Even if Hammer fails in the server market, Microsoft is likely to make money on Hammer, as they *will* sell in the desktop market.

As for Linux stability: You are more or less right. Although Linux stability has improved with new Out Of Memory Killer, improved journaled filesystems etc. Only thing left is to make a scheduler, that isn't O.

Or you can run with *BSD; that is what we do in our webserving setup: run numerous Linux boxes (basically PC hardware in pizzaboxes) that connect up to a big Dell box running FreeBSD and the DBMSes. The BSD box handles in excess of 130.000 connections/minute.

Cheers
Gubbi

 

[pascal]

The Opteron (hammer´s oficial name) annoucement: http://www.amdzone.com/releaseview.cfm?ReleaseID=859

Quality is key and Sun quality is very high.
Windows is trash for IT people.
Linux is as stable as low end Solaris.

 

[Nite_Hawk]

linux etc

Hi Guys,

Gubbi: Linux is getting an O(1) scheduler.

http://groups.google.com/groups?hl=en&threadm=linux.kernel.1019501283.939.42.camel%40phantasy.SOMEWHERE&rnum=9&prev=/groups%3Fq%3Dlinux%2Bscheduler%2Bo(1)%26btnG%3DGoogle%2BSearch%26hl%3Den%26scoring%3Dd

Saem: dude, we are working on linux as fast as we can. There's a lot of work to do to put a decent and somewhat consistent UI on it. Sun just hired a consulting firm with about 50 programmers over seas to help work on gnome though, and things are starting to come along. The addition of bidirectional text, AA, and subpixel rendering to gtk will help quite a bit. I'm not really happy with some of the UI designs in gnome, but I figure they'll eventually come around.

Nite_Hawk

 

[Saem]

Nite_Hawk,

Dood, you work at Sun?

I know work is being done on Linux. I suppose this is the usual Linux cycle. A new kernel comes out it's BIG step forward (2.4.x from 2.2.x was a BIG deal) but then after a while the enthusiasm and the "Greatness of Linux" mantra wears thin. Open or Free (I can't keep them straight) BSD have made HUGE strides and their files system, TCP/IP stack and Multi-threading. The 2.5.x kernel is suppose to have some major rework done on multi-threading, but I'm not sure what else is going on. Also, I remember hearing rumours that one might see a JVM come standard with the Linux new kernel, is this true? I'd personally be interested in seeing the I believe it was the "Cavalry" -because it always arrives just in time- kernel module, it JIT compiled Java bytecode into x86 assembly.

Gubbi,

Heh, saying MS will make money is almost as sure as gravity, even it looses money, that's merely temporary, the take another wack and dominate. Kinda sad, I don't personally like MS. If they did have such great products, then they wouldn't need low-level lock down to make sure you can't switch out of them easily and would rather sell the product on it's merits. Oh well, not much can be done for now.

I'm not sure about complete "ownage" in the price performance department. Lets not under estimate the P4 and Intel, 32 bit ALUs, large L1 and trace caches and improve Hyper-Threading are all on the way, just to name a few. But back to just the Hammer. If you take a look at CPU price history, AMD has been inching it's pricing upwards with each new AXP they realease. I think by the time AMD releases the Hammer, price will be rather close with Intel and they may even justify tying it up or beating their counter-parts.

-regards

 

[Gubbi]

I expect Sledgehammers to be priced similarly to Xeons (ie. $600-800). But building the infrastructure to support these chips will be *alot* simpler with the sledgehammers. I expect the price advantage to be at the system level rather than per cpu. And a hammer system will scale better than a Xeon system (unless Intel revamps their bus).

Cheers
Gubbi

 

[Saem]

Oh I see. Then I whole heartedly agree that on that level they could definately hold an advantage. As for the Xeons, do you think Hyper-threading and the 533MHz FSB will make a difference, also note the new dual channel DDR chipset (which was best by the DRD i860, sad really. I wish SLDRAM made it).

 

[Gubbi]

I think Intel and AMD will be neck and neck in the 1 and 2 CPU segment (Intel commanding a higher price because of brand recognition). Higher clock rates and Hyperthreading might give Intel the edge here (side note: Hyperthreading increases pressure on the trace cache, which is already rather small. Of course increasing the size of the trace cache would take Si real estate from an increased lvl 2 cache, and hence detrimental to performance; I would expect larger performance variations for Xeons with different sized (I&D) workingsets).

For the 4-8 way machines the scalability of the Hammers will kick in (>19GB/s aggregate bandwidth for 8-way Sledge).

Anyway. Competition is good, and AMD is ensuring that competition will be *there* for a couple of more years to come (now we need someone to step up to M$).

Cheers
Gubbi

 

[JonWoodruff]

Who's going to sell enterprise Hammer systems?

If hammer is practical in 8-way configurations, (which it seems that it would be, esspecially with a NUMA optimised OS) isn't that a new market for x86-compatible machines? How will those kind of systems be produced? Will Microsoft endorse them and get HP or somebody else to sell those big systems, or will only smaller OEMs push them with Linux as the OS? Or better yet, will SUN themselves adopt the Hammer for low-end servers using Linux?

Basic question: Who is going to sell the 8+ processer boxes that Hammer will make possible, and with what OS? And for what application?

 

[Saem]

Basic question: Who is going to sell the 8+ processer boxes that Hammer will make possible, and with what OS? And for what application?

This is the big question, WHO?

HPaq, Dell, IBM, SGI and Sun are basically all taken. Gateway went all Dell and said Intel only. I'm not sure who's big enough and will take on the Hammer. Everyone has their own or is using someones non-x86 solution.

I suppose AMD might get lucky, if their Linux is good, they might get picked up by some of linux integrators pogolinux and a whole bunch of small guys like that, which might be a proof a concept and if they make enough waves -it'll be hard- they'll a bigger guy will take them under their wing.

 

[JonWoodruff]

It's looking better.

Ok, we've learned a lot in the last few days. First of all, Microsoft really is excited about the Hammer, which means they will push the 64bit hammer platform. Furthermore, according to the register and other places, Michael Dell commented in the "Hardware Heaven" thing that after Hammer, everyone will go to two sources. (!) Sanders had said that Intel would have to do x86-64 because Dell wanted it, and Michael Dell agreed, sort of, by that comment. (!) Dell already doesn't like the Itanium and needs some chip to push up into Sun territory.

So how much of a stretch is it to say that Dell will use Hammer in the super-high end, higher than what Intel can provide x86-wise, with Microsoft? Hard to say if they'll do it, but with Dell talking like that, I don't think it's a stretch. When that Athlon was comming out, Michael Dell was saying it had reliability problems. With the Hammer comming out, he's saying that the whole industry will have to adopt it. (He said that too, by the way.)