GW GPU O2MDFS2_Spotlight work

archae86
archae86
Joined: 6 Dec 05
Posts: 3,070
Credit: 6,013,962,517
RAC: 2,530,461

robl wrote:0 invalids, 0

robl wrote:
0 invalids, 0 errors.  Credit granted as promised. 

Yes.  It appears that sometime during the night here in New Mexico (morning in Germany, which is more relevant), the validator was turned back on, and the previous "validate error" results were reassessed.

My test machine currently has 52 valid, 114 pending, 0 error, 0 invalid, and 0 inconclusive.  For mine the credit awarded was in all cases 1000 cobblestones.

vsral
vsral
Joined: 25 Jul 20
Posts: 8
Credit: 327,078,256
RAC: 136,207

Seems like my RX580 doesn't

Seems like my RX580 doesn't really like these WU's. GPU usage is all over the place (including weird noises from the gpu).

Credit 1000 for over twice the runtime as GRP WU's (3465 credit).

Are my settings of or why could this be? My GPU is doing two at a time because it was more efficient that way for GRP WU's. I changed the ratio back to 1 for GW WU's but my task manager had already downloaded a huge pile of WU's. Best not to abort these I take it?

Richie
Richie
Joined: 7 Mar 14
Posts: 652
Credit: 1,702,978,594
RAC: 29

vsral wrote:Are my settings

vsral wrote:
Are my settings of or why could this be? My GPU is doing two at a time

Your runtimes are normal for that card. It takes roughly the same 2100 secs for my RX 580 to run these tasks two at a time. These take longer to run and give less credit than GRP tasks. That's just how it is. For anybody wanting to maximize his credit production GRP would be the first choice. GW tasks are nice if you are interested for the sake of supporting GW science.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 1,834
Credit: 14,931,169,192
RAC: 40,475,635

vsral wrote: Seems like my

vsral wrote:

Seems like my RX580 doesn't really like these WU's. GPU usage is all over the place (including weird noises from the gpu).

Credit 1000 for over twice the runtime as GRP WU's (3465 credit).

Are my settings of or why could this be? My GPU is doing two at a time because it was more efficient that way for GRP WU's. I changed the ratio back to 1 for GW WU's but my task manager had already downloaded a huge pile of WU's. Best not to abort these I take it?

These GW tasks are much more CPU bound than the GR tasks, and GPU utilization will be directly related to the CPU power available. I would make sure you have some spare cores available on that old Xeon to help support the GPU task. If you're running CPU work on that processor also, you should restrict it so that some cores are free to support the GPU task.

_________________________________________________________________________

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 1,834
Credit: 14,931,169,192
RAC: 40,475,635

Richie wrote:vsral

Richie wrote:

vsral wrote:
Are my settings of or why could this be? My GPU is doing two at a time

Your runtimes are normal for that card. It takes roughly the same 2100 secs for my RX 580 to run these tasks two at a time. These take longer to run and give less credit than GRP tasks. That's just how it is. For anybody wanting to maximize his credit production GRP would be the first choice. GW tasks are nice if you are interested for the sake of supporting GW science.

you both are running old and weak X58 era Nahelem/Westmere processors. IPC isnt great, and if you aren't leaving some processor available to support the CPU component of the GPU job, it will restrict performance. make sure you aren't maxing out the CPU while running these GW jobs.

 

on my nvidia cards, and coupled with good CPU resources available, I see 75-90% GPU utilization, and runtimes are faster than on GR tasks.

as an example:

RTX 2070 + 3.8GHz Xeon E5-1680v2

GR: ~600 seconds

GW: ~410 seconds

 

RTX 2080ti + 5.0GHz i7-9700k

GR: ~340 seconds

GW: ~250 seconds

 

i do agree that the credit reward isn't in line with the work done (when compared to credit per time for GR). if they doubled it to 2000, that would make things more even in my opinion.

_________________________________________________________________________

archae86
archae86
Joined: 6 Dec 05
Posts: 3,070
Credit: 6,013,962,517
RAC: 2,530,461

vsral wrote:Seems like my

vsral wrote:

Seems like my RX580 doesn't really like these WU's. GPU usage is all over the place (including weird noises from the gpu).

Credit 1000 for over twice the runtime as GRP WU's (3465 credit).

Are my settings of or why could this be? My GPU is doing two at a time because it was more efficient that way for GRP WU's. I changed the ratio back to 1 for GW WU's but my task manager had already downloaded a huge pile of WU's. Best not to abort these I take it?

Probably all normal.

The credit rate is indeed much lower for Einstein GW than for Einstein GRP.

As the GW task needs much more CPU support than does the GRP, and as the ratio of the support needs is considerably different in the early parts of the task than later, the application is sensitive to other demands on our CPU (for example CPU BOINC work) much more than GRP.

As to the "huge pile", this was a consequence of fetching GW while the system Task Duration correction Factor was at a low value "learned" from running many GRP tasks, fooling the system into thinking your average time to finish a GW task would be much, much lower than the truth, which it will estimate once the DCF has popped up to the correct one for your system running GW.

If you are allowing both GW and GRP work to run at the same time, you'll find the GRP tasks will complete faster than usual, and the GW "even slower" than otherwise, as a consequence of the differing pace of CPU support pauses.

My advice to you:

1. If credit is your primary concern, just don't run GW--you'll get less.

2. Avoid mixing GW and GRP--choose one or the other.

3. During the transition, set your queue request very low--say 0.1 day, to avoid nasty surprises.

4. With your 580 and your CPU with 6 physical cores scheduled as 12 by hyperthreading, I strongly suspect you'll find processing of these GW tasks to be much more efficient at multiplicities higher than 1.  I suspect you'll see 2 far better than one, 3 a good bit better still, and quite possibly more steps up.  Try them, but remember to keep the queue request low enough to avoid gross over-fetch.

Richie
Richie
Joined: 7 Mar 14
Posts: 652
Credit: 1,702,978,594
RAC: 29

I forgot that I was actually

I forgot that I was actually running these with -50% power limit. I will need to see today what the runtimes will be without any limiting.

vsral
vsral
Joined: 25 Jul 20
Posts: 8
Credit: 327,078,256
RAC: 136,207

Ian&Steve C. wrote: you both

Ian&Steve C. wrote:

you both are running old and weak X58 era Nahelem/Westmere processors. IPC isnt great, and if you aren't leaving some processor available to support the CPU component of the GPU job, it will restrict performance. make sure you aren't maxing out the CPU while running these GW jobs.

You just hurt my CPU's feelings :(. Didn't know a w3680 was that bad. It has 3958 floating points MIPS/CPU vs +/- 4800 for a r5 2700x in the cpu benchmark of the boinc manager.

I do have other cores running Rosetta as my 'main' project and just wanted my rx580 to do something useful as well.

 

archae86 wrote:

The credit rate is indeed much lower for Einstein GW than for Einstein GRP.

As the GW task needs much more CPU support than does the GRP, and as the ratio of the support needs is considerably different in the early parts of the task than later, the application is sensitive to other demands on our CPU (for example CPU BOINC work) much more than GRP.

If you are allowing both GW and GRP work to run at the same time, you'll find the GRP tasks will complete faster than usual, and the GW "even slower" than otherwise, as a consequence of the differing pace of CPU support pauses.

My advice to you:

1. If credit is your primary concern, just don't run GW--you'll get less.

2. Avoid mixing GW and GRP--choose one or the other.

3. During the transition, set your queue request very low--say 0.1 day, to avoid nasty surprises.

4. With your 580 and your CPU with 6 physical cores scheduled as 12 by hyperthreading, I strongly suspect you'll find processing of these GW tasks to be much more efficient at multiplicities higher than 1.  I suspect you'll see 2 far better than one, 3 a good bit better still, and quite possibly more steps up.  Try them, but remember to keep the queue request low enough to avoid gross over-fetch.

Good to know that the lower credit is not only due to my current suboptimal setup. Science first, credits secondary, but when I can't get these GW WU's to play nicely with Rosetta on the other cores I will stick to GRP I'm afraid.

I see in the Status section (0.9 CPUs + 0.5 AMD GPUs) but the task manager only gives 1 cpu for both running GW WU's.

And after changing some stuff my GPU utilisation is now stuck on a steady 35%. Well, at least the frog sounds are gone now.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 1,834
Credit: 14,931,169,192
RAC: 40,475,635

Richie wrote:I forgot that

Richie wrote:

I forgot that I was actually running these with -50% power limit. I will need to see today what the runtimes will be without any limiting.

just make sure you have some spare CPU resources to feed the GPU. don't try to run the CPU at 100%

 

vsral wrote:

You just hurt my CPU's feelings :(. Didn't know a w3680 was that bad. It has 3958 floating points MIPS/CPU vs +/- 4800 for a r5 2700x in the cpu benchmark of the boinc manager.

I do have other cores running Rosetta as my 'main' project and just wanted my rx580 to do something useful as well.

dont trust the BOINC benchmark, it's pretty meaningless. without a doubt an R5-2700X runs circles around the W3680. 

as I said to Richie, make sure you have some spare CPU available. if you are trying to run the CPU at 100% you will bottleneck everything hard. it has also come to my attention that BOINC doesnt properly account for CPU resources <1. so that 0.9 CPU effectively means 0 to BOINC. so BOINC thinks you have a spare core available when you really do not since the GW job is likely using a full CPU core to feed the GPU (it is in my case, and most times even using more than 1 core, 110-150% of a thread = 2 threads being used).

 

my first recommendation would be to use an app_config.xml file to force BOINC to account for the full 1.0 CPU thread being used. that way your CPU% setting in the compute preferences will be more accurate to what you want and what is actually accounted for. then set the CPU use percentage to ~90% to prevent over committing the CPU. you can still run Rosetta alongside, but you might need to give up 1 thread from Rosetta and give it to Einstein.

_________________________________________________________________________

Peter Hucker of the Scottish Boinc Team
Peter Hucker of...
Joined: 12 Aug 06
Posts: 570
Credit: 333,011,693
RAC: 2,368,899

Richie wrote: Peter Hucker

Richie wrote:

Peter Hucker wrote:
Could it be the first card has a newer instruction set?

Previous batch of GW GPU tasks didn't work with GCN 1.0 cards (like R9 280X) either... so I'm quite sure that's the reason.

Can the server tell what cards we have or what instruction set they have?  If so it shouldn't send gravity to those cards.  Also the program itself should abort with an error, not just run indefinitely.  Surely if the instruction required is not on the chip, it should give up!

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.