(Rumour) XB2 CPU @ 65nm

[notAFanB]

Simple. Memory limitation.

not simple enough, what do you mean? cacheing the kernal?

Power5 is a shared memory SMP machine; it runs one copy of operating system and addresses are visible by all processors.

this is a problem because? PS: do you mean SMMP?

CELL is a split memory "share nothing" MPP machine, but implemented on single piece of sillicon. Every APU gets its own exclusive memory space which is fairly small, not big enough to support something as large as NT.

first how do you know this with no real architecture released except "it'll be blue gene cutdown"?
what's to stop CELL or any other MultiP from sharing Memory?

second why can't an NT kernal fit into this model?

Both the kernel and process has to fit within 128 KB size for optimal performance, which is obviously impossible with NT but possible with some special purpose kernels.

why 128K? why would you want to the move the NT kernal over like this?
Is this really an insurmountable hurdle? or enough to warrent competitive alternatives?

EDIT: one more thing.

Which NT requires. I was explaining why CELL couldn't run NT.

well yes I can agree on that, however that's not really what I was getting at. Why can't an MDMP system akin to (what we proposes as) CELL run the NT kernal.

break it down to an principal would be appreciated.

 

[London Geezer]

Sorry to be the little ignorant guy here, but what fundamental difference do the Linux and NT kernel have, if one is able to run just fine and the other not-at-all?

 

[maskrider]

People have been forgetting things.

For Win NT kernels, Win NT has been supporting PPC architectures since the beginning (dropped after 3.51).

For loosely coupled PPC/Power architecture, there is always IBM SP/2.

IBM has no problem with SMP or loosely coupled parallel processing systems.

 

[Deadmeat]

...

Sorry to be the little ignorant guy here, but what fundamental difference do the Linux and NT kernel have, if one is able to run just fine and the other not-at-all?

The fundamental difference is that Linux kernel is much more flexible than NT kernel. Linux kernel can run on anything from a wristwatch to supercomputers, while I have yet to hear NT/XP PDA.

 

[Panajev2001a]

Re: ...

The Kernel does not have to reside ikn the APUs' Local Storage,

Which NT requires. I was explaining why CELL couldn't run NT.

but in the "Shared" dRAM...

Which I undestand as pipes to interconnect various microprocesses spawned by the parent process.

CELL is not a multithreaded uni-process machine; rather it is a single threaded multi-process machine.

With NT you would not need the APUs to run the Kernel: I have seen NT run on SMP machines and an SMP set-up is exactly what the PUs would appear like.

Edit: you also have not heard of Windows XP Embedded and Windows CE 3.0 ( soon with either 4.x or 5.x they will push DirectX fully onto its specs ).

 

[notAFanB]

Sorry to be the little ignorant guy here, but what fundamental difference do the Linux and NT kernel have, if one is able to run just fine and the other not-at-all?

why maskrider covers a good point regarding IBM R&D so far, this is pretty much what I was getting at.

is it inappropriate for NT as an MDMP/ MMP type of OS? maybe, but why?


EDIT:

The fundamental difference is that Linux kernel is much more flexible than NT kernel. Linux kernel can run on anything from a wristwatch to supercomputers, while I have yet to hear NT/XP PDA.

that doesn't really explain much, only give room for conjecture. besides it's not like the decision to gofor one OS over the other (dev, porting and what not) happens in a vacumm.

Double EDIT:

With NT you would not need the APUs to run the Kernel and I have seen NT run on SMP machines and an SMP set-up is exactly what the PUs would appear like.

really? may I ask where?

 

[DeanoC]

Re: ...

I cannot see would kind of massive, insurmoiuntable obstacles prevent this. please elaborate and try and avoid the jargon.

Simple. Memory limitation.

If memory really is a problem than it would run MS embedded OS. Sorry to ruin your idea that MS has one peice of huge code/OS, they have at least 3 fully fledged OS's.

Power5 is a shared memory SMP machine; it runs one copy of operating system and addresses are visible by all processors.

CELL is a split memory "share nothing" MPP machine, but implemented on single piece of sillicon. Every APU gets its own exclusive memory space which is fairly small, not big enough to support something as large as NT. Both the kernel and process has to fit within 128 KB size for optimal performance, which is obviously impossible with NT but possible with some special purpose kernels.

Why does all code have to fit into 128Kb? You think you going to run a complex OS on an APU. Personally I'd run the OS on the main CPU, but hey what do I know...

CELL isn't new from an OS point of view, its just another attempt to get multi-tasking working efficiently. Lots of OS work in the 70's was based on much more complicated designs, including multiple kernels protected from each other. Getting an MS OS running on CELL would be a fun exercise for the OS HAL team.

How can you exclude any possibility when a company as powerful as Microsoft is involved? Time is on Microsofts side this time around and a huge market is at stake.

Because Microsoft is a software company and its strength comes from the hundreds of millions of lines of code it has accumulated, not hardware. Can't run NT? Forget it.

So the includes all PocketPC's, Dreamcast, phones?

Even Xbox isn't NT, just shares some code.

 

[Deadmeat]

...

> not simple enough, what do you mean? cacheing the kernal?

The private address space of each APU is not large enough to support NT Kernel image. Exceeding the 128 KB limitation will result in a massive performance drop.

> first how do you know this with no real architecture released except "it'll be blue gene cutdown"?

I understand how Blue Gene works. I also understand how CELL works from reading the patent application.

> what's to stop CELL or any other MultiP from sharing Memory?

Because Blue Gene/L nodes don't share memory. That's not how Blue Gene(Both L and Cyclops) works.

> second why can't an NT kernal fit into this model?

If you can fit NT kernel image into a 128 KB space, I think you will be made the next OS project manager at MS.

> why 128K?
Past 128 KB APU has to resort to off-chip memory reading.

> why would you want to the move the NT kernal over like this?
Of course no one wants to. THis is why MS chose Power5.

> Why can't an MDMP system akin to (what we proposes as) CELL run the NT kernal.

Technically CELL could. Just not fast and efficiently enough.

 

[Fafalada]

The private address space of each APU is not large enough to support NT Kernel image. Exceeding the 128 KB limitation will result in a massive performance drop.

Rrrrright - because every OS call performs several GFlops worth of uhmm... er... what exactly?

 

[Panajev2001a]

Sorry to be the little ignorant guy here, but what fundamental difference do the Linux and NT kernel have, if one is able to run just fine and the other not-at-all?

why maskrider covers a good point regarding IBM R&D so far, this is pretty much what I was getting at.

is it inappropriate for NT as an MDMP/ MMP type of OS? maybe, but why?


EDIT:

The fundamental difference is that Linux kernel is much more flexible than NT kernel. Linux kernel can run on anything from a wristwatch to supercomputers, while I have yet to hear NT/XP PDA.

that doesn't really explain much, only give room for conjecture. besides it's not like the decision to gofor one OS over the other (dev, porting and what not) happens in a vacumm.

Double EDIT:

With NT you would not need the APUs to run the Kernel and I have seen NT run on SMP machines and an SMP set-up is exactly what the PUs would appear like.

really? may I ask where?

Last I knew Windows NT/2000/XP had licenses for more than 1 Processor ( up to 16 CPUs ) for quite a while.

We had some servers at work with licenses for 2 CPUs or more and they were runnign Windows 2000.

 

[notAFanB]

>The private address space of each APU is not large enough to support >NT Kernel image. Exceeding the 128 KB limitation will result in a >massive performance drop.

fair enough, any programming solutions to minimise or attack this problem in the industry? how about academia?

>I understand how Blue Gene works. I also understand how CELL works >from reading the patent application.

I don't possess any understanding of how bluegene works the above was merely an example.

>Because Blue Gene/L nodes don't share memory. That's not how Blue >Gene(Both L and Cyclops) works.

an that has what exactly to do do with my question below?

> what's to stop CELL or any other MultiP from sharing Memory?

>The private address space of each APU is not large enough to support >NT Kernel image. Exceeding the 128 KB limitation will result in a >massive performance drop.

OK that makes sense.

any solutions or appraoches to tackle this problem come to mind?

>If you can fit NT kernel image into a 128 KB space, I think you will be
>made the next OS project manager at MS.

answer my question please if you could.


>Past 128 KB APU has to resort to off-chip memory reading.

Why is NT kernal 128K?

>Of course no one wants to. THis is why MS chose Power5.

the question was would you move an NT kernal over like this?


>Technically CELL could. Just not fast and efficiently enough.

not efficiently? maybe, after all it's not like the current Kernal and hardware evolution don't go hand in hand. But what are the solutions and ways to attack this issue?

 

[Deadmeat]

...

Quick FAQ. How BlueGeneL/CELL works.

In "shared nothing" massively parallel machines like BlueGeneL/CELL, there are two kinds of kernels, the general purpose I/O kernel to govern the behavior and I/O of overall system and a very simple application kernel to support the execution of single process within a processor address space. The I/O kernel is typically the linux, while the application kernel is system-specific and non-standard.

In a normal Linux system, when the application programmer initiates the "fork and exec" chain of command to start a new process, the new child process is setup within the application space of same kernel that hosted original parent process.

In case of BlueGeneL/CELL, the new child process is setup within the application space of a tiny application kernel, which is then uploaded into the one of application processors(APU) and executed. The purpose of this application kernel is to support the execution of that single process.

When you understand this you understand how CELL works.

 

[Panajev2001a]

Forget about the APUs.

Imagine CELL as 4 simple nice PUs ( PE = 1 PU, 1 DMAC and several [in the Broadband Engine as described in Suzuoki's CELL patent we see 8 APUs per PE] APUs ) which you think as your CPUs.

The PUs are simple, yet functional CPU cores which can run an OS.

So we have, thinking about the Broadband Engine as described in Suzuoki's CELL patent, a 4 CPUs SMP system ( Simmetric Multi-Processor system ).

Windows NT/2000/XP can run on a quad Xeon SMP system, so why should it not run, albeit modified here and there, on a quad PE SMP system ?

( btw, some IBM Cell patent described multiple Processor Elements systems as an SMP configuration )

 

[London Geezer]

The problem here seems to be: do YOU understand that?

Cause we have several devs here telling you you're quite wrong, DM... Nothing personal, just watching this from the outside...

 

[Deadmeat]

...

fair enough, any programming solutions to minimise or attack this problem in the industry? how about academia?

None that I heard of.

I don't possess any understanding of how bluegene works the above was merely an example.

Because CELL is a BlueGene/L derivative.

Why is NT kernal 128K?

???

not efficiently? maybe, after all it's not like the current Kernal and hardware evolution don't go hand in hand. But what are the solutions and ways to attack this issue?

Why are you asking me this?? CELL APU was specifically designed to host a tiny application kernel and a small process. It won't do anything large like NT kernel.

 

[DeanoC]

>The private address space of each APU is not large enough to support >NT Kernel image. Exceeding the 128 KB limitation will result in a >massive performance drop.

OK that makes sense.

any solutions or appraoches to tackle this problem come to mind?

Lets take DM at his word, that NT will not fit into Cell and MS wouldn't consider using a system thats not NT. Then a hypothical MS CELL would simple bang a cheap OS CPU in it, you can license a normal CPU (say for example 32 bit ARM7 core ) thats entire job is to direct the OS and manage the APU's.

Problem solved, OS support isn't even an minor consideration in CPU for Xbox2.

 

[notAFanB]

Re: ...

Quick FAQ. How BlueGeneL/CELL works.

In case of BlueGeneL/CELL, the new child process is setup within the application space of a tiny application kernel, which is then uploaded into the one of application processors(APU) and executed. The purpose of this application kernel is to support the execution of that single process.

When you understand this you understand how CELL works.

How did you link the usage in Bluegene to how CELL works?

also why does this exclude the NT kernal as an alternative? any solutions?

 

[Panajev2001a]

Re: ...

Quick FAQ. How BlueGeneL/CELL works.

In "shared nothing" massively parallel machines like BlueGeneL/CELL, there are two kinds of kernels, the general purpose I/O kernel to govern the behavior and I/O of overall system and a very simple application kernel to support the execution of single process within a processor address space. The I/O kernel is typically the linux, while the application kernel is system-specific and non-standard.

In a normal Linux system, when the application programmer initiates the "fork and exec" chain of command to start a new process, the new child process is setup within the application space of same kernel that hosted original parent process.

In case of BlueGeneL/CELL, the new child process is setup within the application space of a tiny application kernel, which is then uploaded into the one of application processors(APU) and executed. The purpose of this application kernel is to support the execution of that single process.

When you understand this you understand how CELL works.

The APUs get DMAed by the PU a new Program Counter, a new Stack and their job they have to do ( which might consist of a program or a program and some data [the APU cna DMA the data it needs alone if you want it to do so] ).

You can still run the Windows NT/2000/XP style kernel with the PUs and store it in the Shared DRAM which is not all partitioned between the APUs ( the PUs can partition it the way they want to ).

 

[Deadmeat]

...

Windows NT/2000/XP can run on a quad Xeon SMP system, so why should it not run, albeit modified here and there, on a quad PE SMP system ?

A ported NT kernel will run on dual PPC cores inside CELL. But NT does not support the kind of "fork and exec into an alien kernel" mechanism without a massive modificatin.

 

[Panajev2001a]

>The private address space of each APU is not large enough to support >NT Kernel image. Exceeding the 128 KB limitation will result in a >massive performance drop.

OK that makes sense.

any solutions or appraoches to tackle this problem come to mind?

Lets take DM at his word, that NT will not fit into Cell and MS wouldn't consider using a system thats not NT. Then a hypothical MS CELL would simple bang a cheap OS CPU in it, you can license a normal CPU (say for example 32 bit ARM7 core ) thats entire job is to direct the OS and manage the APU's.

Problem solved, OS support isn't even an minor consideration in CPU for Xbox2.

DeanoC, that hypotetical OS CPU is already there... it is called the PU ( it directs the OS and manages the APU ).

[0019] In another aspect, the present invention provides a system and method for the PUs' issuance of commands to the APUs to initiate the APUs' processing of applications and data. These commands, called APU remote procedure calls (ARPCs), enable the PUs to orchestrate and coordinate the APUs' parallel processing of applications and data without the APUs performing the role of co-processors.