Your recent WUs take 24+ hours on my modest 660s and 750TIs.

Do they have to be so process intensive??

I am only getting these WUs now - looks like 660 and 750ti are no longer deemed suitable for GPUGRID. Pity.

Im getting them on my 650ti's , they take about 36 hours to complete. Kwityerbellachin but I am getting credit for them

My GTX 750 Ti's are running a little warmer too - up to 58 C, which is 2 C more than usual. But that is still quite cool. They are there to be used as much as possible.

These GERARD_FXCXCL12 tasks are longer than most, but they are 8-12h on the fastest cards (GTX980, GTX780Ti and some Titan models).

For comparison my GTX970 took 13.75h:
27 Mar 2015 | 12:13:29 UTC 28 Mar 2015 | 6:31:38 UTC Completed and validated 49,556.33 49,170.73 255,000.00

Running under W7 I'm subjected to the WDDM overhead and I was also running Einstein on my integrated gpu, 5 CPU tasks and because I had another GPU in the system at the time I was working at PCIE2x16, rather than PCIE3.
I guess on an ideal Linux setup the 970 could have been ~20% faster and completed in around 11h, which would make cache settings irrelevant (so long as you run 24/7 and only used the GPU for GPUGrid).

As for smaller cards, there have not been very many large tasks recently. While I don't know what the researchers plans are on that front, task size is generally at most 8-12h on the fastest cards, so the likes of a GTX750Ti should be able to return such tasks within 48h (for +25% credit) for some time.

Tomba, if you see any tasks running on your GTX750Ti, you could manually suspend them in Boinc Manager and restart them in an order which forces them to run on your GTX770 and thus complete inside 24h.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Even my GTX 660 (with the core clocked at the reference 980 MHz) takes 24 1/2 hours on these. But that is OK. These aren't the fastest cards, and the science should not be held back because of it.

PS - My GTX 660 Ti likes them much better, at 19 hours. It is one of those cases where more cores helps significantly.

My 750Ti finished one of these in just under 24 hours. Unfortunately, I won't be receiving the full bonus due to more than 24 hours having passed since the task was downloaded. Still, don't count these cards out yet!

Edit: I did put a small overclock on the cards to see if I could reduce the run-time. I was able to go from 90k seconds to a little over 85k seconds with a +91MHz core overclock and a +175MHz memory overclock while still at stock voltage.

My 750 Ti crunched one of these wus in a little more at one day (86812 s) at 1306 MHz.

I did a GERARD_FXCXCL12_LIG_1214901 on my oc 780ti in 7.71 Hrs but does not show up in THE TOP PERFORMERS IN BATCH list even though thats faster than any on the list (so far).

Also the Top average performers (last week long runs) has the same user in 8 9 and 10 spot?



I also just posted this on the SERVER... FORUM under Performance tab thread where it should be.

Your recent WUs take 24+ hours on my modest 660s and 750TIs...

When I first saw how long my modest 570 was taking to complete e1s4_2-GERARD_FXCXCL12_LIG_731951-0-1-RND5224, I thought it had downclocked like it sometimes did long ago. Alas, I was wrong... :-(

---

Tomba, if you see any tasks running on your GTX750Ti, you could manually suspend them in Boinc Manager and restart them in an order which forces them to run on your GTX770 and thus complete inside 24h.

Now that's a great idea! Thanks.

While we are on the subject of great ideas, maybe there is still an optimization possible for the Maxwell cards that would allow the GTX 750 Ti to complete them in 24 hours? This seems to be a good motivation for it, if more of the long Gerards are coming down the road.

It's called Linux.
Other things that can expedite the process are not using the CPU for CPU work, using SWAN_SYNC, slightly overclocking, setting higher process priority, affinity and using process lasso might also help, as might using an iGPU for the display.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

SWAN_SYNC

That's something I've been trying to find more information on in the forums but have had no luck. What is it and how does one enable it?

SWAN_SYNC is an environmental variable.
Basically, it will force the task to use a full CPU core/thread.

Last I heard is that on Linux you need to use SWAN_SYNC=1

https://www.gpugrid.net/forum_thread.php?id=3890&nowrap=true#38355

Try some searches for SWAN_SYNC + Linux.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Thanks. That link set me on a trail that eventually answered my question. I now have Swan_Sync=0 set up on my 750Ti rig, and while the tasks are still not using a full CPU core each, the GPU usage is much smoother and 1% or 2% higher than before. Curious to see how this affects my times.

Definition of Swan_Sync

Setting up Swan_Sync (Windows)

(Both linked posts courtesy of skgiven)

Managed to get one in under 24 hours on 750Ti FTW. I'm running overclocked on W7 w/ core clock at 1410MHz and memory clock at 2950MHz. I enabled SWAN_SYNC somewhere around halfway through the the task.

For SWAN_SYNC to work you need to restart the computer, after you create it. You also need to specify in Boinc Manager to leave at least 1 CPU core/thread free to benefit, but leaving more than one free improves things even more. On a 4core/8thread system there is usually no benefit in using less than 50% of the CPU (4threads) and not much difference between using 4 and 5 threads. I normally only use between 5 and 7 threads 62.5% to 87.5% of the CPU.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

For SWAN_SYNC to work you need to restart the computer, after you create it. You also need to specify in Boinc Manager to leave at least 1 CPU core/thread free to benefit, but leaving more than one free improves things even more.

Is using SWAN_SYNC any more beneficial than reserving 1 CPU core per GPU using app_config?

Last time I checked doing both was only slightly more advantageous (2 or 3%) than just freeing up a CPU core/thread (or two), but if you are just missing a deadline then it might help. Freeing up a thread or two could help more, but once you've done that you might as well use SWAN too.
The benefits of SWAN_SYNC could depend on the task type and your system (CPU performance and GPU). Unfortunately there are so many task types around now it could take weeks to get enough results to present accurate measurements.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help