Challenge
log in

Advanced search

Message boards : News : Challenge

Author Message
Profile Kevin
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
Send message
Joined: 27 Jul 12
Posts: 507
Credit: 14,550,449
RAC: 2,436
Message 1078 - Posted: 25 Jan 2013, 23:21:11 UTC

theSkyNet POGS team welcome the challenge crunchers.

You guys are awesome. We're generating more galaxies for you to Crunch over the weekend. Yesterday I had to ramp up the generator as you are sucking down 3 galaxies every 10 minutes. Thats 432 a day! In the past 6 months we'd only processed 2,300 galaxies. You'll do more than that in a week!

The BOINC servers are just about handling the load, but the Database has been struggling. I might have to up it to the next size Amazon Instance

Thanks

____________
Regards
Kevin
-----
International Centre for Radio Astronomy Research

Profile Freddykrug
Send message
Joined: 14 Aug 12
Posts: 13
Credit: 1,237,234
RAC: 95
Message 1085 - Posted: 26 Jan 2013, 16:50:47 UTC
Last modified: 26 Jan 2013, 16:52:47 UTC

When will released the application for GPU, then will be worse :)
Apropos, when it to released?

Profile Kevin
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
Send message
Joined: 27 Jul 12
Posts: 507
Credit: 14,550,449
RAC: 2,436
Message 1087 - Posted: 27 Jan 2013, 9:53:32 UTC - in response to Message 1085.

We've not found anyone who wants to do the port yet :-(
____________
Regards
Kevin
-----
International Centre for Radio Astronomy Research

Ross*
Send message
Joined: 20 Jan 13
Posts: 4
Credit: 9,193,465
RAC: 0
Message 1090 - Posted: 27 Jan 2013, 22:30:19 UTC - in response to Message 1087.
Last modified: 27 Jan 2013, 22:30:50 UTC

We've not found anyone who wants to do the port yet :-(


Hi
perhaps you could use AVX like Primegrid ?

Congrates on the recovery. Best I have seen.

Off topic
Why are we getting these short WUs and my pending has gone up from 870 to 1120?

Cheers
Ross*
____________

Profile Kevin
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
Send message
Joined: 27 Jul 12
Posts: 507
Credit: 14,550,449
RAC: 2,436
Message 1091 - Posted: 27 Jan 2013, 23:12:58 UTC - in response to Message 1090.

@Ross

I must confess I hadn't thought of using Advanced Vector Extensions. I'm not sure how I'd separate out the executables as only a few processors support it at the moment (Intel Sandy & Ivy Bridge; AMD Bulldozer & Piledriver). I'll ask the BOINC community how this could be done as I like the idea.

I checked the work unit logs and we've had a lot of small galaxies go through recently. Essentially they have a small galaxy in the middle of the image and lots of smaller features around the edge. These bits around the edge tell us about galaxy evolution and classification (my research area - which is why you're getting a long answer :-) ). This can tell us about past galaxy interactions and evolution. If we can see faint signals it could be a small dwarf galaxy interacting with the bigger main galaxy. Once the Murchison Wide-field Array and the Australian Square Kilometre Array Pathfinder come fully on line we'll be able to look for tidal interactions between galaxies - which you can only see in Radio frequencies.

If you go and look at what you've been processing you'll find you have been unlucky and got a lot of edges.
____________
Regards
Kevin
-----
International Centre for Radio Astronomy Research

Ross*
Send message
Joined: 20 Jan 13
Posts: 4
Credit: 9,193,465
RAC: 0
Message 1092 - Posted: 27 Jan 2013, 23:38:06 UTC - in response to Message 1091.

Hi
Thanks for the quick and full reply.

AVX has reduced processing time by 40% and more in the Primegrid tasks.

yes some processes are up to it and most new ones have it, with the new Intel chips have AVX2 ?

as I run 9 i7 3930ks AVX is of great interest

How many galaxies are you planning to chart? I have seen photos of some parts of the sky with hundreds of thousands.
Cheers
Ross*

____________

Profile Kevin
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
Send message
Joined: 27 Jul 12
Posts: 507
Credit: 14,550,449
RAC: 2,436
Message 1093 - Posted: 28 Jan 2013, 0:03:48 UTC - in response to Message 1092.
Last modified: 28 Jan 2013, 0:29:07 UTC

@Ross

We're looking at local (in galactic terms) galaxies out to a redshift of around about z=0.05. We have to find areas the GALEX, WISE and Pan-STARRS have all looked at. That'll be a few hundred thousand. If it goes well then we'll slowly expand the catalogue. Pan-STARRS (http://pan-starrs.ifa.hawaii.edu/public/) is a fantastic telescope for this because:


  • It has the largest digital camera in the world at 1.4Gpixels (yes 1,400,000,000 pixels)
  • It is a survey telescope so it can map very large areas of sky to great sensitivity. It looks everywhere, not just where people are interested.


The number of galaxies out there is truly amazing. Carl Sagen expressed wonderment at the vastness of space and time, in his phrase "The total number of stars in the Universe is larger than all the grains of sand on all the beaches of the planet Earth." Turns out from the latest Hubble Deep Field, Ultra Deep Field and GOODS images (http://hubblesite.org/hubble_discoveries/breakthroughs/cosmology) he was absolutely correct.

Current estimates from Hubble and Super-Computer simulations put the number of galaxies at possibly as many as 500 Billion. We have to remember though Hubble can only see visible light. When the Square Kilometre Array comes on line in 2018ish we may well have to increase that as some galaxies are only visible in radio frequencies.
____________
Regards
Kevin
-----
International Centre for Radio Astronomy Research

Daniel Carrion
Volunteer moderator
Project tester
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 24 Nov 12
Posts: 159
Credit: 32,883,143
RAC: 21,858
Message 1094 - Posted: 28 Jan 2013, 1:25:11 UTC - in response to Message 1092.
Last modified: 28 Jan 2013, 2:29:51 UTC

@Ross*

I've actually been experimenting with different gfortran compiler options for compiling the main fit_sed app (just out of curiousity). One of them included the -march=corei7-avx -mtune=corei7-avx option on my i5-3570k. It seem to whack additional instruction in but I didn't see any major improvement. I'm guessing you would actually need to use AVX intrinsics inside the code to get the full benefit instead of relying on automated detection/conversion.

One thing I did have success with was the Intel compiler (ifort) for Linux. Although the binary size went from ~200K to 1.6MB, performance went up quite a bit when using optimised flags. On an AMD machine of mine, a single fit went from ~330 seconds down to ~230. I didn't try on my i5-3570k machine at the time but I suspect it would give good results. I'm going to try it now :).

I noticed that all your machines are Windows so this Linux version I've been testing won't play nice. However, I reckon if someone could get their hands on Intel compiler for Windows, one would be able to compile a fairly optimised fit_sed just from compiler options. The Intel compiler really seems to be doing things right for producing optimised instruction code for various architectures.

UPDATE:

Here's a comparison of a 15 fit run. I couldn't find exact clock speed comparison straight off, sorry.

GNU gfortran compiled (windows) on i7-3930K CPU @ 3.20GHz :
http://ec2-23-23-126-96.compute-1.amazonaws.com/pogs/result.php?resultid=1919950 ~4500 seconds


Intel ifort compiled (linux) on i5-3570K CPU @ 3.40GHz:
http://ec2-23-23-126-96.compute-1.amazonaws.com/pogs/result.php?resultid=1929365 ~1900 seconds

Works out to be > ~120 second difference per fit.

Ross*
Send message
Joined: 20 Jan 13
Posts: 4
Credit: 9,193,465
RAC: 0
Message 1095 - Posted: 28 Jan 2013, 2:26:49 UTC - in response to Message 1094.

Hi
nice work
I am not a programer, just use the apps to the best avantage.
going along the line of a GPU app, perhaps you could look at OPEN CL.
this should then run on ATI and NV GPUs
The GPU apps realy get though that data mountain. If you want to get those other 499 billion galaxies start thinking of how much data that is.
The Boinc coummity will really help as they can see real science in action coming from a profession project.
If you look at some of the other long term projects they have thousands of computers processing the data.
Even PrimeGrid can have over 1000 indivials in a challenge.
Just a few thoughts
Cheers
Ross*
____________

Daniel Carrion
Volunteer moderator
Project tester
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 24 Nov 12
Posts: 159
Credit: 32,883,143
RAC: 21,858
Message 1096 - Posted: 28 Jan 2013, 2:49:35 UTC - in response to Message 1095.
Last modified: 28 Jan 2013, 3:01:17 UTC

@Ross*

I commend your thoughts and you're absolutely right. There's sooo many crunchers out there and a lot of them want to see their hardware being used as efficiently as possible.

going along the line of a GPU app, perhaps you could look at OPEN CL. this should then run on ATI and NV GPUs


I agree...I reckon OpenCL has great promise as it allows execution across heterogeneous platforms including CPUs, GPUs, etc... So you could even see benefits of running OpenCL code on CPUs in some instances. It really does seem like "the" framework for science apps.

I personally feel that the app might need to undergo a port over to C/C++ first. Although it is more than possible to use Fortran to snap-in OpenCL code, it would become quite messy and you'd have to consider how you're going to integrate other arch code as well (e.g. CUDA, VFP, AVX, etc...) down the track. It also helps that Khronos maintain C++ OpenCL binding code. In saying that, I have little experience with Fortran and new arch support may be greater than I know. This is obviously Kevin's call, though. Like you, I'm just a community member throwing out thoughts. :)

All in all, I think it's great that this community is legitimately interested in having optimised science applications as it really does push projects forward. If everyone was just happy with what currently existed, there wouldn't be any progress...Or at least not as much...

Cheers

Daniel

Ross*
Send message
Joined: 20 Jan 13
Posts: 4
Credit: 9,193,465
RAC: 0
Message 1097 - Posted: 28 Jan 2013, 3:47:03 UTC - in response to Message 1094.

Hi
perhaps you could get in contact with some of the programers in PrimeGrid, as they seem to try every angle to optomise their code.

How big of a mountain of data will you have to process over the coming years?

there are crunchers and teams that may want to support you on a long term bases if they knew what the numbers were, and how long.

The best projects are the stable but evoling. plenty of fed back about what is happening and the results. Some crunches like long tasks[ but they must be stable] no point if a couple of days or a week down the tack it turns to custed.

There are some very smart crunches out there and they will look at better/faster ways to do things.

I have gone on enough for today
Cheers
Ross*

____________

Message boards : News : Challenge


Main page · Your account · Message boards


Copyright © 2017 The International Centre for Radio Astronomy Research