[question] ps3 vs "normal" server

[Samuell.Sk]

I just wanted to ask if someone tried to install ydl (yellow dog linux) on "fat" ps3 and tried to run there mango.

Another question is would cell powered ps3 benefit from over 3 - 4 times processing power (in floating points) than any other i7 cpu when running mango ?

This cloud be more of a question for skilled dev likve Vladimir (yes we all worship him for over 2- 3 years now) since i bet he will know if mango use more or less float points.

I know that cell cpus are beasts when it comes to a numbers (our local medical facility has cluster of 8 of those using them for medical purpose). Their times have drooped from week/s of processing to hours/days.

Note to anyone who will be contributing to this post:

Dont mention here 256mb ram of ps3, or that you need ram to handle thousands of users.... that is not the point of this post. Interested in a "raw" comparsion in this situation.

[faramir118]

Mangos isn't parallelization friendly, so I don't think it would scale any better on a PS3 than it would on an i7.

[Patman128]

You would basically have to rewrite mangos to get any decent performance boost (or even to break even with an x86.) Like faramir said, there really isn't that much room to scale on it. Mangos is just too monolithic.

[Samuell.Sk]

Mangos isn't parallelization friendly, so I don't think it would scale any better on a PS3 than it would on an i7.

Please, and for future replies, read the question first and note on the end if possible. This is question about flops (if mangos is doing them enought if it will benefit more from cell cpu or not) Not comparing single/dual/multi threading applications / cpus. Your answer belongs to the "note part" which is not required nor desired.

[Samuell.Sk]

You would basically have to rewrite mangos to get any decent performance boost (or even to break even with an x86.) Like faramir said, there really isn't that much room to scale on it. Mangos is just too monolithic.

Please, and for future replies, read the question first and note on the end if possible. This is question about flops (if mangos is doing them enought if it will benefit more from cell cpu or not). Not about if a majority of code must be rewriten to benefit from it. Your answer belongs to the "note part" which is not required nor desired.

[Lynx3d]

I bet they did read it, but you apparently have no clue how Cell actually works...

"Normal" C/C++ code doesn't just "run on Cell" after throwing a compiler at it. Well it does, but not on the SPUs that make the vast majority of its transistors, but on the meager PPE that has the performance of maybe a Pentium3 at best.

Using the whole Processor means writing new code that needs to be compiled with a special compiler and run in one thread for each SPU.

But let's just be generous and still call that hypotetical rewrite "MaNGOS", would it perform well?

Probably not...

MaNGOS consists of two relevant parts really, a very branch heavy logic core, and some float heavy geometry interaction (vmaps mainly). The SPEs maybe could do the float parts (although i have yet to see a convincing raytracing demo on cell...) but the SPEs suck at branch-heavy code, and the the PPE is hardly capable to compare with even one core of a current Intel/AMD CPU when it comes to logic code...

[Patman128]

Please, and for future replies, read the question first and note on the end if possible. This is question about flops (if mangos is doing them enought if it will benefit more from cell cpu or not). Not about if a majority of code must be rewriten to benefit from it. Your answer belongs to the "note part" which is not required nor desired.

If you want to know how many FLOPS mangos uses, just run your own benchmarks. I don't think asking vladimir or some other developer to speculate on it is very desirable (not that vladimir is one for idle speculation anyway.)

[Samuell.Sk]

I bet they did read it, but you apparently have no clue how Cell actually works...

"Normal" C/C++ code doesn't just "run on Cell" after throwing a compiler at it. Well it does, but not on the SPUs that make the vast majority of its transistors, but on the meager PPE that has the performance of maybe a Pentium3 at best.

Using the whole Processor means writing new code that needs to be compiled with a special compiler and run in one thread for each SPU.

But let's just be generous and still call that hypotetical rewrite "MaNGOS", would it perform well?

Probably not...

MaNGOS consists of two relevant parts really, a very branch heavy logic core, and some float heavy geometry interaction (vmaps mainly). The SPEs maybe could do the float parts (although i have yet to see a convincing raytracing demo on cell...) but the SPEs suck at branch-heavy code, and the the PPE is hardly capable to compare with even one core of a current Intel/AMD CPU when it comes to logic code...

Just for the record. I7 980XE has 107.55 GFLOPS, ps3 with its 1 core and 7 SPEs has over 200.

This answer was needed. Thank you.

Topic can be closed.

[FragFrog]

I think your estimate for the PS3 is a bit low - I find sources indicating it's more like 400GFLOPs. However, AMD's HemlockXT 5970 GPU for example rates about 4TFLOPS, roughly ten times as fast. But I very much doubt you can use any of that power with Mangos' current design, even if you could get it to run.

Might be interesting to see if it can be optimized for CUDA though, too bad none of our servers have a decent GPU