Dears, we are starting to take full advantage of the "long" queue. I've sent a batch of 700 WUs named "AB*", which should be 8-12h on the fastest cards.

I started a (still preliminary) batch of experiments to study the behavior of a protein likely implied in Alzheimer's disease.

By the way, if you happen to see "MDIO warnings" on redefined parameters, they are harmless.

Thanks guys for the data. We are still tweaking WU sizes and credit factors. When things settle, do expect 8-12h and a credit factor of 1.5-2 (in addition to the one for fast return, if any). Maybe I'm optimizing WUs too much ;)

Concerning machine crashes, sorry to hear. BTW, I normally run the WUs on 275s in addition to Fermis.

Hi Ben.
Your longtask times seem to be all over the place; 19K, 22K, 55K, 36K.
I see a fairly constant 26000sec +/- 5% on my GTX470's. Some of this may be explained by different task types, but 19K and 22K for two TONI_AB1 tasks is quite far apart and 55K for a TONI_SMDTRYP7LONG10 sounds a bit long, I ran one and it was only slightly longer (under 28Ksec).

As you have a 500 series Fermi, I think it might be throttling itself (dropping speed if it gets too warm). If not perhaps your Boinc settings or system usage is causing such high variation in times.

Note that when you are using swan_sync and configure Boinc to only use 99% of the CPU cores (threads in your case), the GPUGrid task will use a full thread so your CPU usage will remain at 100% overall. After you set swan_sync you should restart the system. Might be an idea to check the last checkpoints of your tasks first.

Check the temperature of your system and GPU especially. You should be using temp controlling software to keep the card at a reasonable temperature (60-70deg). I suggest MSI Afterburner, as this allows you to configure the fan speed acording to the temperature (it increases as the temperature rises, and decreases when it should). A simpler but less controllable program is EVGA Precision. Both can be downloaded from freeware sites such as http://majorgeeks.com/ - restart required.

This IBUCH_2_mutEGFR task type is taking about 12h on a GTX470 (XPx86, swan_sync, free i7 thread). I think this is a good size given that it just creeps into the estimated 8-12h range and the card is at the lower end of the top Fermi cards.

As nenym said, use swan_sync=0 and free up a CPU core in Boin Manager.

For how to see the FAQ: Best configurations for GPUGRID thread.

Some of the IBUCH tasks tend to be a bit less resource intensive than the other tasks, but 40% utilization is too low, so improve your settings.

One other thing, Vista and Win7 tend to be about 11% slower, due to the Windows Display Driver Model (WDDM), than XP; and that's after the system is optimized for GPUGrid crunching.

Did you also configure Boinc to only use 50% of the CPU?
I get 83 to 84% GPU utilization on my GTX470s when running the long IBUCH tasks.

The G105M is not powerful enough to run GPUGrid tasks.
It only has 8 shaders, you need about 100.

Boinc requests tasks based on the estimated finish time of existing work and other things such as your cache size. Boinc has to calculate how long a task will take and get a general idea of how long tasks are for each project. Until your card runs a task Boinc cannot tell how long it will take.

If GPUGrid implemented a strategy of only allocating tasks on the basis of the GFlops peak reported by Boinc then some Fermi's might not get tasks because not all Boinc versions report the performance correctly.

NVIDIA GeForce GTS 250 (511MB), Clock rate: 1.84 GHz, driver: 27061
Windows XP Home x86 Edition, Service Pack 3, 2GB RAM
Athlon 64 X2 Dual Core Processor 4400+

SWAN: FATAL : swanBindToTexture1D failed -- texture not found

ERROR: file c:\cygwin\home\speechserver\gpumd2_c\src\pme\CPME_cufft.cpp line 73: cufftExecC2C (gridcalc2.1)
called boinc_finish

SWAN: FATAL : swanMemcpyDtoH sync failed

ERROR: file c:\cygwin\home\speechserver\gpumd2_c\src\pme\CPME_cufft.cpp line 73: cufftExecC2C (gridcalc2.1)
called boinc_finish

ERROR: file c:\cygwin\home\speechserver\gpumd2_c\src\pme\CPME_cufft.cpp line 73: cufftExecC2C (gridcalc2.1)
called boinc_finish

So, lots of cuda fft errors and a few others on a GTS250. This is about power for the course for that GPU.
You probably have a better chance of finishing the shorter tasks, but most GTS250’s are just not good to crunch with here. You might want to plan for a replacement.

One of the researchers might have a better idea, but I would guess it would be useful up until about 2 months after the next generation of NVidia cards. So I would say it will be next year some time before it's obsolete. Obviously this depends on NVidia's Kepler release (i'm guessing it will be in the first half of next year before you can buy one).

To some extent the CC1.3 cards are naturally fading from the project; a GTX 295 is now 2years and 4months old (quite old for a GPU). As you would expect, many have upgraded to newer cards, and this is likely to continue. A couple of months after the next generation of NVidia cards appear, very few users will still be crunching with a CC1.3 card; Kepler should eventually double the performance of comparable Fermi cards, never mind the CC1.3 cards.

Even if the Kepler cards are up to PCIE3 standards, they should be backward compatible for older PCIE2 and probably PCIE boards, albeit at some loss (gaming, video editing) - 50% bandwidth. For crunching the bandwidth restrictions would be less noticeable.

Different generations are hard to compare, but I would equate a GTX285 with a GTX580, which offers a bit more than a GTX480 (more equivalent to a GTX280). Even now software for Fermi cards is still developing. So the performance should improve again slightly. Win7 is at least 11% slower than XP. When the Kepler cards turn up they will probably triple the performance of their equivalent GTX 200 card. So until that happens a high end GTX 200 card will remain a good card.

I would expect that a GPUs within 1/10 of the power of the top one will be useful. If kepler will have 1024 cores, then anything above 128 cores should be fine for the short workunits (acemd2). For acemd_long workunits 1/3 of the top card is probably acceptable, so over 240-300 cores once kepler is out.

gdf

Quirino Cuccioli, per i compiti a lungo la GT 220 non è abbastanza veloce.

en. The GT220 is not fast enough for long tasks.

Ton: some other guy got 24.417 ms per timestep (roughly double the speed) on a 470 for a similar task (result 4034521). Don't know what happened but should be transient.

Buona domanda. E 'possibile, ma solo in laboratorio. Il problema è che BOINC è limitato.

En. Just saying that we cannot run one app over 3 GPU's due to the limitations of Boinc, but it is possible in the labs. Pity, 3 GT220's might do well enough (144 shaders combined), ditto for the GT240 and similar cards.