28 July 2012

28th of July

The farm continues to make steady progress. The GTX670's have been running non-stop on GPUgrid and as of this post have climbed to position 34 on the project rankings by RAC (Recent Average Credit).

CPDN
Climate Prediction has been issuing lots of work units in the last week. The last time I looked they had 22,000 EU and 9,900 PNW (Pacific North West) regional models available. This is in addition to the 137,000 models that are out being processed at the moment.

I wonder if the plumbing will be able to cope. That is BOINC and the server hardware they have, as they have a history of server issues particularly around data storage. Their work units are large and produce 12 trickles of 7Mb each and a final file that is typically 32Mb for the PNW models.

FPGA blog
Last week I mentioned the blog started by Terry trying to get FPGA support into BOINC and some science apps converted. I sent him a note during the week, which he has quoted here: http://www.fpgaathome.org/2012/07/26/message-to-fpga-manufacturers/

He has started on trying to get the Seti@home multibeam application running (or at least parts of it) on an FPGA. If anyone is interested in helping you can contact him via his blog (leave a comment) or send an email to terrystratoudakis (at) gmail (dot) com


GPUgrid
We had a little drama on the project this week. One of the new scientists to join the project issued some rather unusual work units which were all failing. BOINC then starts to reduce the amount of work you can get because your host is returning errors and eventually you get down to being only allowed 1 work unit per day until you return a valid result.

The offending work units were flushed out of the system and corrected ones issued. From reports these new ones take a long time to process. I haven't received any of the corrected ones yet so can't say by how much.

19 July 2012

19th of July

Further to yesterdays post about coprocessors. There is a blog regarding using FPGA's for computing and specifically integrating them into BOINC. Its called FPGA at home and can be found at http://www.fpgaathome.org/

In his post of 18th of July he links to an Xcell article in which they state that they used the XC6VLX240T with a MicroBlaze processor. Looking at the recommended pricing on the Xilinx web site for the board and processor it has them costing approx $2500.

I am not against using FPGA's for computing, in fact I think they should be used. If it can do maths when why wouldn't you use it? Its just that the makers seem to have priced them out of reach. Until they decide to get serious the likes of AMD/Intel will control the CPU market and ATI/Nvidia will have the GPU market to themselves.

18 July 2012

18th of July

Not a great deal of activity on the farm this week. Things have been chugging along without any incidents (touch wood).

The GTX670's have managed to get me up to position 36 based upon RAC (Recent Average Credit) on GPUgrid which is what they spend their time processing. Considering I was down around position 150 or so about a month ago this is pretty good. More importantly they are contributing to some useful science.

BOINC testing
We're still testing 7.0.31 which has a few minor changes since 7.0.28. I have reported one issue to do with project initialisation not honouring the exclude_gpu preferences.

Coprocessors
I have been making some inquires with a number of manufacturers regarding coprocessors. I had initial email discussions with Adapteva who have a multi-core CPU, ClearSpeed who make a maths coprocessor and DRC regarding their FPGA offerings. Unfortunately all seem to have priced themselves around $10,000 which makes them too expensive for volunteers to use.


Future plans
At the moment we're still waiting on Nvidia to release a GTX600 series card to replace the 560Ti. I have three GTX560Ti cards so I am keeping an eye on announcements. The main reason is electricity prices went up by another 18% this month, with the carbon tax getting the blame, and my electricty bill is getting rather large now.

I still need to find a case for the Adventure. Once I have one I will also need a machine to run it and might look for a relatively cheap AMD based motherboard. It doesn't need much CPU power as all it needs to do is drive a PCIe x16 slot to feed the Adventure.

04 July 2012

4th of July

The farm continues to run. 4 of the 5 i7's have been running constantly for the last couple of weeks. The two GTX670's have managed to get their RAC (Recent Average Credit) up to 317,000 for GPUgrid (which is all they are running) with maybe a little higher to go before they plateau.

Unfortunately electricity prices have gone up again, with the introduction of the Carbon tax and having to fund infrastructure. The last quarterly bill came in at $650 which is mostly the computers. Just as well I upgraded to the GTX670's as they should save 78 watts per hour from the cards they replaced. Once an equivilent of the GTX560Ti is available as a Kepler-based card I will look at replacing all of them as well, if only to reduce electricity use.

BOINC testing
This week we got to 7.0.31 which has a few minor tweaks since 7.0.28, but nothing major in it. I have it on one machine and its doing fine so far. Speaking of 7.0.28 it has now become the recommended version, so if you are running an older client I would suggest you upgrade.


New project - Asteriods@home
This is a new project. It is still in start up mode. They are computing the shape of asteroids from multiple sightings. They currently only have a Linux science app. Once the bugs have been ironed out they will port this science app to Windows.


Project news - Climate Prediction
They had a 3 day outage because of network problems connecting to their DB machine. When they came back online yesterday they also had some new work for the regional models, about 1800 work units. I managed to manually get 3 of the machines to pick up work and by the time I got to the 4th machine it had all gone. They have a 1 hour backoff after a scheduler request so it took some time to get them all to make work requests.


Project news - GPUgrid
Ignassi, one of their project staff had his docterate defence yesterday. He is now known as Dr. Ignasi Buch. Unfortunately that also means he is leaving the project, so we wish him well.

Their CUDA 4.2 app is now working well with run times reduced by almost half, so what used to be 8 to 12 hour work units are taking 5 hours. I am not sure how much of this is due to using CUDA 4.2 and how much is due to the newer graphics cards, but either way its impressive.