Doesn't exactly inspire confidence in your chosen supplier.

No, but I can't name names, that'd be sugaring the pill...

Mistakes happen, but repeatedly?

Welcome to dysfunctional corporate Britain. I'd so love to be home in Byron Bay, but I fear I'd have to sell my soul to afford it.

---

The right box showed up yesterday. OS installation hampered slightly by there only being two USB ports. Couldn't get the "extended" USB port in one of my keyboards to work so ended up with a powered hub for mouse, KB, and USB stick. Got the Intel compilers installed, but some complaints about missing 32-bit libraries. Then to the multicore support -- only to find that Intel only provides support for Red Hat and SuSe, and our server guru had given me Ubuntu to install... There are recipes out there to get it converted and re-compiled but it's going to take me all day Monday and then some. For some reason you need LaTeX installed for compilation, and Intel recommends you use gcc for some modules and their compiler for others!

---

Congrats on finally getting the right box! (You would hope that the purchasing people and suppliers would realise that all of the cryptic hierglyphs on the order do actually mean something...)

That systems sounds like something very much still under develpment. Welcome to the cutting edge of bleeding edge?

The LaTeX is likely needed for document generation as part of the system build...

And is there any special reason for Ubuntu?...

If you want something more directly comparable to Red Hat or SUSE, then you could try any of CentOS, Mageia, or any "rpm"-based distro. Take care over whether 32-bit or 64-bit is needed, or whether 64-bits with 32-bit libraries is the best option.

For comparison, I've recently installed some supplier software on a 64-bit Gentoo server. The supplier install script was supposedly for Red Hat only and for only c2005! Their support were very obviously scared silly of anything not called "Windows". Also their software is 32-bit only... To get that old stuff working on a modern Linux setup required the standard 32-bit versions of the libraries to be installed and two aliases to rename the modern versions of two utilities to appear as the old names. Voila! Their support couldn't believe it would work and didn't believe it was running ok because the CPU usage is so very low. All works fine.

Full kudos to robust operation and backwards compatibility!


Good luck and don't be too disappointed with poor support for something so new.

Happy fast crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

The LaTeX is likely needed for document generation as part of the system build...

Yeah, that's what I figured.

And is there any special reason for Ubuntu?...

It's what our "systems guru" gave me. I'd have used Scientific Linux CERN 6 (which is RHEL 6 repackaged) to be compatible with the experiment but he argues (rightly so) that its kernel and compilers are woefully out of date. A student we have on the project blanched at Ubuntu too, but then found an Intel page that said Phi was compatible with all Linuxes -- turned out he was thinking of the MPSS requirements, and I'd forgotten about the restrictions there too, in the several weeks since I'd read it. :-(

Good luck and don't be too disappointed with poor support for something so new.

Happy fast crunchin',
Martin

Thanks, we'll see how I get on this afternoon...

BTW, I activated the server as a s@h cruncher on Friday night; it's already got 14,200 credits and 64 WUs pending. (Dual 6-core Xeon; 64 GB RAM.)

---

Any advances on OpenCL porting field so far?

---

SETI apps news
We're not gonna fight them. We're gonna transcend them.

Any advances on OpenCL porting field so far?

Not yet. I've reverted to Scientific Linux CERN 6, and re-installed the Intel compilers. Tomorrow I'll attempt to get the Intel MPSS installed; with a bit of luck I might actually be able to install the Xeon Phi tomorrow afternoon (tho' I do have a weekly videoconference to keep an eye on). Once that's installed and working I'll be able to download the OpenCL SDK and start exploring it.
However, if it doesn't happen this week it may be a while yet -- I've been called up for jury duty which will take out at least next week and the week after; if I get a complex case, who knows? Then I might need to go back to Australia for several weeks for a family emergency (depends on when a major surgery is scheduled).
So all in all, don't hold your breath. I'm running as fast as I can but everything is conspiring against me (lost a filling last week, my central heating is shot, and I don't have the time to organise a fix for either; at least I got my front door replaced last week after the police tried to break it down in February, but now I need to get the £1100 cost out of them...)

---

It doesn't fit. The supplier didn't include the special brackets needed to mount the Xeon Phi...

---

It doesn't fit. The supplier didn't include the special brackets needed to mount the Xeon Phi...

In such extreme exasperation of such brazen continuing ineptitude, may I suggest a completely deadpan Socratic approach be attempted with the whichever neuron challenged supplier?...

Can you deduct costs for your lost time?!


The best of luck,
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

It doesn't fit. The supplier didn't include the special brackets needed to mount the Xeon Phi...

In such extreme exasperation of such brazen continuing ineptitude, may I suggest a completely deadpan Socratic approach be attempted with the whichever neuron challenged supplier?...

Or maybe skip the debate entirely and just pour the conium down his throat ;)

It doesn't fit. The supplier didn't include the special brackets needed to mount the Xeon Phi...

In such extreme exasperation of such brazen continuing ineptitude, may I suggest a completely deadpan Socratic approach be attempted with the whichever neuron challenged supplier?...

Or maybe skip the debate entirely and just pour the conium down his throat ;)

I think it might be Treason to do that to a Lord of the Realm.

---

Lots of Linux threads all ready to be unleashed -

How Intel's MIC Co-Processors Work On Linux

The Xeon Phi with up to 61 cores and 244 threads while packing up to 16GB of onboard memory makes it quite a beast. Each Xeon Phi card runs Linux and is IP addressable and can deliver 1 TeraFLOPs double-precision peak performance. Intel's Xeon Phi supports offload, native, and symmetric programming models...


Good luck for getting everything together!

Happy very fast crunch in'
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

Good luck for getting everything together!

Latest from the sucrose team is that we should have the brackets on Tuesday. Let's hope The Old Bailey gives me some afternoons off, or I'll have to work nights to make good The Professor's pledge to The Student that he can start using the device "by the end of October".

---

, or I'll have to work nights to make good The Professor's pledge to The Student that he can start using the device "by the end of October".

Did you actually state the year?

We're having extensions done at work- walls coming down, walls going up, holes knocked in walls. That sort of thing.
It was meant to have been mostly done by the end of Sept. Then it was meant to be mostly done by the end of October. We assumed they meant this October, at the present rate it'll probably be done by Oct 2014. Maybe.

---

Grant
Darwin NT

, or I'll have to work nights to make good The Professor's pledge to The Student that he can start using the device "by the end of October".

Did you actually state the year?

We're having extensions done at work- walls coming down, walls going up, holes knocked in walls. That sort of thing.
It was meant to have been mostly done by the end of Sept. Then it was meant to be mostly done by the end of October. We assumed they meant this October, at the present rate it'll probably be done by Oct 2014. Maybe.

Sympathies. Been there, done that, didn't really want the t-shirt...
Worst was last year when they decided to replace the lift which features early in Kubrick's Clockwork Orange -- for several weeks longer than planned. I got used to plodding up three storeys with a 3 kg laptop on my back. I've excerpted the CO scene and made a video of the new lift going up and stopping at all floors ("This lift is going up. Please mind the doors. This is floor three.") but I haven't yet worked out how to join them together -- everything I've tried likes one format but not the other!

---

I've excerpted the CO scene and made a video of the new lift going up and stopping at all floors ("This lift is going up. Please mind the doors. This is floor three.") but I haven't yet worked out how to join them together -- everything I've tried likes one format but not the other!

Did you try Avidemux (I use Avidemux 2.5.6-1)
http://fixounet.free.fr/avidemux/
http://avidemux.berlios.de/
http://avidemux.sourceforge.net/

---

 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 

Thanks for the links, I'll have a look.

---

Latest from the sucrose team is that we should have the brackets on Tuesday. Let's hope The Old Bailey gives me some afternoons off, or I'll have to work nights to make good The Professor's pledge to The Student that he can start using the device "by the end of October".

Well, the brackets arrived on Thursday. Unfortunately I'm now empanelled on a case and we ran the hearing until 1630 both Thu & Fri so I wasn't home until after 1800. Went in yesterday and fitted the brackets and installed the Phi, flashed the system, and got as far as being able to log into the co-processor:

[eesridr@XeonPhi-mic0 eesridr]$ cat /proc/cpuinfo
processor        : 0
vendor_id        : GenuineIntel
...
processor	: 239
vendor_id	: GenuineIntel

I thought there was supposed to be a gcc compiler on the co-pro but no sign of it so I've yet to compile BOINC on it. Also I ran out of time to install the OpenCL SDK so at present BOINC on the host says "no usable GPU". That will change as soon as I get off jury duty!

---

I thought there was supposed to be a gcc compiler on the co-pro but no sign of it so I've yet to compile BOINC on it. Also I ran out of time to install the OpenCL SDK so at present BOINC on the host says "no usable GPU". That will change as soon as I get off jury duty!

Well, since then I installed the OpenCL SDK, but there was a kernel upgrade to install. I found a newer version of the MPSS (v3.1) so I deleted the old one -- which removed some of the OpenCL packages -- upgraded kernel, installed new MPSS, flashed the system and got it running. But OpenCL wouldn't install.
A bit of searching on the Intel site showed that there isn't an OpenCL SDK for the 3.1 MPSS yet.
I tried compiling a version of BOINC to run natively, but configure and icc don't play together nicely. I ended up manually changing all the FLAGS variables in the Makefiles to explicitly include the -mmic flag and make ran for a while and then aborted, unable to find openssl/ssl.h include file.
I then discovered that there is a modified gcc toolchain installed with MPSS and worked out how to use that with configure. Unfortunately, it stopped because there was "no usable version of libcurl" -- not surprising as the installed version wouldn't be compiled to work with MIC. As suggested, I used --disable-client to get around this (tho' together with --disable-server & --disable-manager I wasn't sure what would get built!). configure then ran to completion but again make failed as unable to find the same include file.
On the bright side, I've been running some private benchmarks -- basically testing a parallelised pseudo-random-number generator by binning 50 billion numbers in each thread into a 100-bin histogram. Since there are 12 CPU cores (dual 2.00 GHz Xeon E5-2620), I tested instances with 1 to 12 threads on both the host and the co-proc. The times rose slightly with threads, but the CPU version was about 6 times faster than the Phi. Beyond 12 threads (I only went up to 180) the CPU time increased linearly with multiples of 12 threads, while the Phi time increased slowly (mainly I suspect because of increasing competition for memory access to the array that holds the PRNG state). As you'd expect, around 72-84 threads the CPU execution time became greater than the co-processor time.

---

On the bright side, I've been running some private benchmarks -- basically testing a parallelised pseudo-random-number generator by binning 50 billion numbers in each thread into a 100-bin histogram.
....

How optimised for the CPU compared to the Phi would those routines be?

---

Grant
Darwin NT

On the bright side, I've been running some private benchmarks -- basically testing a parallelised pseudo-random-number generator by binning 50 billion numbers in each thread into a 100-bin histogram.
....

How optimised for the CPU compared to the Phi would those routines be?

Pretty much the same; using the latest Intel compiler with the same compiler switches except adding -mmic to get the native Phi code. There's no scope (unless I sit down and change the algorithm) for using the vector instructions -- it's all single double-precision additions -- so the advantage of 512-bit registers doesn't come into play. There's a factor of two in the clock speeds, too. Also the Phi doesn't use out-of-order execution so to get best performance you'd like to have all four threads on each core in operation, but with 180 threads I didn't see much difference in the execution times using the different methods of thread affinity. The new MPSS knocked about 15% off the real time, but reported the same user time (19h40m in 6m40s). I'm gradually moving on to more realistic benchmarks.

---