[tomtom]

Hello,
I'm shite with the latest fads in PC hardware, but I'm wanting to build a box (no keyboard monitor etc..) for some parallell simulation work - but with as many cores as possible... From what I've seen AMD have a 6 core processer, Intel a 8 core one, then there are various combinations of multi-processor motherboard shizzle to multiply this all up...

So I'm hoping to pick the combined mega brain of UKB to see what kind of thing I can get.. budget £1k (max), HDD doesnt really matter (500gb fine etc..) and memory is not crucial (4gb would be fine..). It will be running W7... (sorry Linux folk..). I dont give a ferk what it looks like, its the number of cores that count..

I'm not prepared to go down the GPU route (yet) as thats another realm of programming things to learn (unless someone knows a simple way..) so its more cores at the moment...
TT

[lagerstarfish]

its the number of cores that count..

is hyper-threading allowed?

[tomtom]

its the number of cores that count..

is hyper-threading allowed?

Yes, I think so - though I am a pleb.

SHIT, DOH! I shouldnt have answered. I have now triggered an @lagerstarfish payment charge..

[lagerstarfish]

invoice in the post

[Paul B]

I've got an AMD Phenom X6 and it seems great. There was a site with them all ranked in all out performance, performance per £ etc. It came out pretty highly in the latter. God knows what 1k would get you as mine was ~500 and tears through heavy use of Fluent, Lightroom and CS5 simultaneously, it also doesn't have issues with Vegas 10.

It does however sound like it might take off with a 22cm top fan or something else ridiculous

Maybe if this is University simulation type stuff you should look at chain-ganging a load of ps3's instead as the graphics processing has been found to be fantastic at CFD iterations.

[iain]

I'd second what Paul said. Price per real core AMD win hands down.

Hyperthreading depends on how hard you're working the processor. Each thread on a core really only has 50% of the processing time available to it, as soon as it goes over that it's useless.

If you were looking for more cores than that from AMD/Intel you'd need to think about an Opteron or Xeon dual processor system.

[tomtom]

I've got an AMD Phenom X6 and it seems great. There was a site with them all ranked in all out performance, performance per £ etc. It came out pretty highly in the latter. God knows what 1k would get you as mine was ~500 and tears through heavy use of Fluent, Lightroom and CS5 simultaneously, it also doesn't have issues with Vegas 10.

It does however sound like it might take off with a 22cm top fan or something else ridiculous

Maybe if this is University simulation type stuff you should look at chain-ganging a load of ps3's instead as the graphics processing has been found to be fantastic at CFD iterations.

Thanks Paul (and Iain)- Id seen the AMD reviews and wasnt sure compared to the intel. Bangs per buck it seems to score big time..

The PS3/GPU front is an appealing way (£300 graphics card buys you 60+ processors  - or however they do it) but I'm unsure/put off by the extra shennanigans I'd have to do with the code. ATM, I can simply replace well constructed for-next loops (etc..) with a parallell version and bobs my uncle...

[Paul B]

from what I understand you 'can' do that on a PS3 (and it'll run between 10 and 50 times quicker than on a single core) or if you vectorise it, it'll run at 6 times faster than the prior factor!

Whopping.

[slackline]

How does that all square up with Sony removing the "OtherOS" functionality?  Is a M$-Win variant running on the PS3?

I've come across a papers in the past¹ that compares CPU to GPU and it was favourable, but I think one of the barriers might be the coding as already alluded.

Afraid I know little about performance on multi-cores, I've lowly dual-core at home to satisfy my geeky needs, and for anything more powerful at work I use the "wonderful" ( ) 568 core, 2.3 TB RAM Iceberg part of the White Rose GRID.

¹Davis et al (2010) Real-world comparison of CPU and GPU implementations of SNPrank: a network analysis tool for GWAS. Bioinformatics Advance Access:

[tomtom]

Slackers you're right..
I've seen / read things on using CUDA to program GPU...  From what I've heard its pretty hard to get stuff into CUDA - thuogh there are other ways now.. but I do all my stuff in c#.. I know there is a set of c# libraries/extensions that do GPU under the name of brahma, but its not (at all) clear how all this works etc..
I'm fairly savvy about structuring code/variables to work OK in a parallell environment - but not so sure about the whole GPU thing... and CUDA's less than user friendly operation. But GPU is seductive.. £100 sheets or so can buy a graphics card with 32 procs on it....

[Lund]

Afraid I know little about performance on multi-cores, I've lowly dual-core at home to satisfy my geeky needs, and for anything more powerful at work I use the "wonderful" ( ) 568 core, 2.3 TB RAM Iceberg part of the White Rose GRID.

- more than about 4 cores is an expensive waste of time for general purpose processing because it's hard to parallelize the processing effectively.
- if you're doing the same calculation in parallel with little or no shared data more cores is amazing
- graphics processors are optimised for repetitive shit, but usually because they have on chip acceleration to do that kind of thing which you often need to use specific SDKs to use
- the on chip cache size is likely to be critical for large data sets, more so than the number of cores or the speed
- faster chips (i.e. raw per-core speed) means a hotter chip.  Hot chips aren't good in embedded/dense hardware architectures, where usually you have a multi-core chip that runs slower but more... so your intel chips run a lot hotter and lot faster per-core than say a cavium chip but that's because of the application they're designed for
- multi core is much harder to design your application for; quad core is largely fire and forget.

All that said: Intel take some beating right now.  AMD have dropped the ball.  On the low power/lots of cores devices... probably freescale/power PC is best.  (Cavium not great.)

[slackline]

The other paper I couldn't remember last night...

²Sinnott-Armstrong et al (2009) Accelerating epistasis analysis in human genetics with consumer graphics hardware. BMC Research Notes 2:149

[tomtom]

Thanks Guys,
Lund, the code scales really well for parallell tasks - theres little shared memory and I can scale the 'chunks' of loops that are taken off to individual cores/threads pretty easily to reduce the overhead to a minimum..

As I kind of said in the OP, I'd love to do this on GPU - as it'd be faster and cheaper - but I'm put off using CUDA as frmo what I've seen and heard (more heard) its a bitch to get right... I was kind of hoping/wondering if someone had written some nice C# libraries that would do all the shizzle for you. I suspect Brahma does this, but I can't make head nor tail of what or how it works from its website #notgeekyenough...

So my halfway house is to use multi cores... whats Cavium?? (awaits LMGTFY...)
TT

[Lund]

Cavium is just another chip manufacturer - like AMD/intel/etc.

So.  Basically heat isn't an issue for you, which means you don't have to compromise on core speed.  Twice as fast is just as good as twice as many cores right?  So I would think probably intel is the best chip manufacturer.

If I were you... at this point I'd have a look on their site and decide what you want - then go and try to find a board with one on - a development system of some description.  I'd also then realise that these all come without windows - nobody uses windows - and were quite pricey and had a long lead time. :-(

What's the driving factor?  software development cost/budget/timeline?  If money and time and were no object... I would buy an ATCA chassis, put two quad quad core intel CPU processor blades in it and write some amazing distributed software to run on it.  If money was tight... I'd buy a load of cheap regular PCs, and a gigabit ethernet hub, and write a distributed application.

I know nothing about PSX/GPU etc., so cannae help there.

[tomtom]

Theres too much overhead to do multi machine at the moment (its lots of smallish jobs rather than a few larger ones if that makes sense).. Yes, twice the processor speed would be same/better than two cores.. but from what I know, if I spend £500 on a processor as opposed to £100, it wont be twice as fast.. (single core speed) whereas I can get a c.90% speed up moving from a dual core to a quad core..

I know you didnt like the AMD's, but they do a 6 core proc, thats cheap.. Dual 4 core processors is probably where I'm heading though at the moment...
Thanks though - some good ideas...

[slackline]

I know you said you don't want to re-write C# for linux, but there are tools to help, and if you can get access to the aforementioned White Rose GRID via your institutional affiliation you'd have everything at your disposal (even if not already able to register with White Rose, you might still be able to apply for an account and use it).

Perhaps have a peek at...

DotGNU an implementation of the Common Language Infrastructure (CLI), more commonly known as ".NET", includes everything that you need to compile and run C# and C applications that use the base class libraries, XML, and Systems.Windows.Forms. Currently supported CPUs: x86, ppc, arm, parisc, s390, ia64, alpha, mips, sparc. Supported operating systems: GNU/Linux (on PCs, Sparc, iPAQ, Sharp Zaurus, PlayStation 2, Xbox,...), *BSD, Cygwin/Mingw32, Mac OS X, Solaris, AIX.

Mono is a software platform designed to allow developers to easily create cross platform applications. Sponsored by Novell, Mono is an open source implementation of Microsoft's .NET Framework based on the ECMA standards for C# and the Common Language Runtime. A growing family of solutions and an active and enthusiastic contributing community is helping position Mono to become the leading choice for development of Linux applications.

[Obi-Wan is lost...]

I recently got a i7-860 machine through the Dell outlet store, pondered long and hard about building one or getting a no-brand one from Flea-bay etc. In the end didn't have time/couldn't be arsed so got a Optiplex 980 with 3 year warranty. Really impressed with it so far although haven't really put it though it's paces much yet. It did rip a full DVD feature film in 10mins and re-encoded it in another 10min not even breaking a sweat, fun to watch 8 cores doing their stuff (4 physical, 4 virtual/HT). It was virtually silent whilst doing so which was nice and the air coming out the back barely rose above room temp. My only disappointment is that I'd hoped it would ship with 64bit Windows but it came with Win7 Pro 32bit and no media. Currently querying with Dell why I didn't get the option when the spec listed both, may still upgrade if I can be bothered.

[slackline]

I've sussed tomtom out, he's getting us to do his homework whilst...

Heading to Wubicon soon, eta 3 ish if anyones about?? T

Hope this isn't what he's teaching his students! 

[tomtom]

:p the lot of you

DotGNU an implementation of the Common Language Infrastructure (CLI), more commonly known as ".NET", includes everything that you need to compile and run C# and C applications that use the base class libraries, XML, and Systems.Windows.Forms. Currently supported CPUs: x86, ppc, arm, parisc, s390, ia64, alpha, mips, sparc. Supported operating systems: GNU/Linux (on PCs, Sparc, iPAQ, Sharp Zaurus, PlayStation 2, Xbox,...), *BSD, Cygwin/Mingw32, Mac OS X, Solaris, AIX.

Mono is a software platform designed to allow developers to easily create cross platform applications. Sponsored by Novell, Mono is an open source implementation of Microsoft's .NET Framework based on the ECMA standards for C# and the Common Language Runtime. A growing family of solutions and an active and enthusiastic contributing community is helping position Mono to become the leading choice for development of Linux applications.

not checked out the first - but have tried with Mono and not got far (didnt support the libraries I wanted - and benchmarks said it ran slower than dot net...)...  I know its for research, but I have a following of about 10-20 around the world who all use, and update the Windoze version - people seem to find C# easy to modify/use/tinker with etc...

I recently got a i7-860 machine through the Dell outlet store, pondered long and hard about building one or getting a no-brand one from Flea-bay etc. In the end didn't have time/couldn't be arsed so got a Optiplex 980 with 3 year warranty. Really impressed with it so far although haven't really put it though it's paces much yet. It did rip a full DVD feature film in 10mins and re-encoded it in another 10min not even breaking a sweat, fun to watch 8 cores doing their stuff (4 physical, 4 virtual/HT).

Whats the score with actual vs virtual cores? I'm not sure I understand this (sorry..).. this could make all the difference I guess.. hmm sounds interesting..

But.. from my own research.. there is an AMD OPTERON (I feel I should put on a darth vader voice!) processor with 12 cores (@2g-thingies) going for £400... Problem is the only motherboard that supports it (actually supports dual processors as well) is £400.. whack in memory and a cpu cooler and its getting on for a G.
But, I have found a AMD Pheonom bundle of 6 core 3.2Gthingoolies proc, 4gb ram and cooler etc.. for £230 inc the Vodka And Tonic.. This may be a good place to start...

[Paul B]

thats pretty much why I went AMD.

[tomtom]

thats pretty much why I went AMD.

Yes - sorry, I meant to acknowledge your reply in my post above *doffs cap to Paul B*...

[Obi-Wan is lost...]

Whats the score with actual vs virtual cores? I'm not sure I understand this (sorry..).. this could make all the difference I guess.. hmm sounds interesting..

Can't say I fully understand it myself but in a nutshell...

For each processor core that is physically present, the operating system addresses two virtual processors, and shares the workload between them when possible

http://en.wikipedia.org/wiki/Hyper-threading

[slackline]

Pedantically (what else) but I think its worth mentioning as I was once confused as to why AMD processors weren't hyper-threaded...

 Hyper-threading is just a brand name by Intels implementation of Simultaneous Multi-threading (as the wikipedia article Obi links to opens with).

[tomtom]

Ordered an Phenom X6 3.2 motherboard, memory bundle today (£236)... see how it goes..
Thanks for all the advice, T

[tomtom]

6 cores up and running great. My work's c.50% faster on 6 cores than 4.. so good news

[iain]

Nice one 

[tomtom]

Piss flaps. Something seems to have gone wrong.. blue screen of death (with a variety of errors) at seemingly random crash times.. tried re-installs (of XP and W7) and no difference. Bugger. I suspect this may be a hardware issue. Any thoughts?

[slackline]

Check the RAM modules by taking all but one out, one at a time.

Get a Live Linux disc (e.g. Linux Mint and boot from that after you've checked each module and see how stable that is.  There are tools for checking hardware, not sure which ones they are or where you'll find them off the top of my head, will check later.

[tomtom]

Check the RAM modules by taking all but one out, one at a time.

Get a Live Linux disc (e.g. Linux Mint and boot from that after you've checked each module and see how stable that is.  There are tools for checking hardware, not sure which ones they are or where you'll find them off the top of my head, will check later.

As in taking one out trying it, then the other replaceing the first etc.. (process of elimination to see if one of the RAM's is faulty?)

[iain]

As in taking one out trying it, then the other replaceing the first etc.. (process of elimination to see if one of the RAM's is faulty?)

Yup.

I'll let slack--line t recommend an app that can test PC components, I've not really had cause to use many, although you may find the motherboard has it's own diagnostics you can run which is worth doing.

It's also worth trying this to see if any of the bits are running hot. AMD cores shouldn't go over 62c under load apparently, and the cpu fan shouldn't be running too fast either, (a quick google suggests 3-4k rpm at idle)

[Jim]

defo hardware issue, try the ram first and also run speedfan as iain suggested.
I'm not sure how installing linux would help tho?

[tomtom]

Yes, its a bit of a funny one.. first time I set it up and installed the OS (on an old HDD) it worked fine for 4-5 hours on close to full tilt (70-80% load on all 6 cores)... then when I re-booted it all went tits up. I suspected the HDD (it was old and well used) so replaced that today and same grief. It 'seems' to happen more when under load, but has also happened a few times when its not doing much - or installing W7 etc..

I did have a prod on the RAM and one of them wasnt quite seated properly.. didnt seem to make much difference.. I may take them both out & in again to seat them properly.. (it came pre-assembled motherboard/cpu combo). Anyway, its at work so will have to wait until tomorrow.

Thanks for the advice Jim, Iain and Slackers - sounds like I should go up the Hardware route rather than faff with BIOS updates et c.. (which was tomorrows next port of call..)..

[Jim]

I would defo strip the whole motherboard and re-seat everything then run something like OCCT to test stability of cpu

[slackline]

defo hardware issue, try the ram first and also run speedfan as iain suggested.
I'm not sure how installing linux would help tho?

I wasn't suggesting installing Linux, but running from a live DVD (which all sits in RAM if it checks out ok in the first step) to see if hardware fails there too, you sometimes get more information from kernel debugging than a blue screen of death provides.

I'm sure there are probably hardware diagnostic tools on M$ install discs that can be run in a similar manner, but I'm not familiar with them so its not that surprising that I haven't suggested things I don't know about. (the ambiguity in my previous post was because I didn't know exactly what is included in Linux Mint, and is why I've included details of how to install if its not on the disc and an alternative).

@tomtom : Inquisitor looks like a comprehensive tool, if its not already on the Live CD that you've booted from you should be able to install it by firing up a terminal and typing...

Code: [Select]

sudo apt-get install inquisitor

...then see the documentation for various tests (or you could try one of their live CDs).

[iain]

it came pre-assembled motherboard/cpu combo

In that case if the RAM doesn't make any difference then I'd suggest going back to the supplier. There will be nothing more you can do anyway without (probably) invalidating the warranty.

[Jim]

slackers, there's a program called memtest that doesn't need an OS to boot into
http://www.memtest.org/

[slackline]

slackers, there's a program called memtest that doesn't need an OS to boot into
http://www.memtest.org/

Yes I've used that before under GNU/Linux (its available in most suites, but as you say also as a stand alone bootable ISO).  memtest+ is included in Linux Mint, but the functionality is covered by Inquisitor which will check more than just memory.

[tomtom]

Back from work now. Tanks for all the help it seems like one of the memory modules was at fault. When it's removed it's been up and running at 80% for a few hours
Thank you all again
Tt

[tomtom]

Slackers you're right..
I've seen / read things on using CUDA to program GPU...  From what I've heard its pretty hard to get stuff into CUDA - thuogh there are other ways now.. but I do all my stuff in c#.. I know there is a set of c# libraries/extensions that do GPU under the name of brahma, but its not (at all) clear how all this works etc..
I'm fairly savvy about structuring code/variables to work OK in a parallell environment - but not so sure about the whole GPU thing... and CUDA's less than user friendly operation. But GPU is seductive.. £100 sheets or so can buy a graphics card with 32 procs on it....

An update for anyone interested...

I went on a 2 day Cuda training course in late November, which tipped my confidence over the edge to have a try - and it became my xmas break coding project #sadfucker

I've managed to get GPU working via NVIDA CUDA 4.0 within C# using a set of wrappers (???) going by the name of cudafy.net. It was a bit of buggering about, but much simpler in the end than going down the dark path of engaging with writing cuda code in native c++...

Its a bit of a YYFY.. as I've managed to get a 5x speed up compared to a multicore version (an i5 quad core) and a 12x speed up compared to a single cpu on the i5. The work is speeding up a flow model, using a laptop with a lowish end GPU (Nvidia GT555m) which is a fair bit (2-5x) slower than what I can get in a desktop.. Happy days Curry and a beer now.