Runtime: how about increasing it

[OldChap]

Sam commented about having poor runtime.

I run my card with lots of concurrent wu's to increase my runtime (while I can) then I came to look at the top "members" list and wondered who can tell me....

Which of these "members" really are individuals



My highlighting is just those I think I know

[stoneageman]

Ditto your selection

[nanoprobe]

And FWIW I think Hypernova could do much more but he is not using his GPUs to anywhere near their full potential IMHO after reading he's running 10 GPU tasks and 2 cpu tasks on his HT hex cores with 7970s.

[OldChap]

Maybe he runs Linux??? Sam says he has issues running many concurrent on Linux.

I only ran my 5870's on Linux so I cannot comment

[nanoprobe]

Could be. IDK.

[BlindFreddie]

OldChap: I run my card with lots of concurrent wu's to increase my runtime

[EDIT: The following is incorrect, as pointed out below by Snow Crash & wplachy [/Edit]

Only CPU runtime is counted, not GPU runtime.
While the GPU is crunching an individual WU, the CPU is only spending a low % feeding it, so it's not incrementing its runtime tally much at all. You're running more tasks than when doing CPU WUs, but the total CPU run-time can only be as high as the number of CPU cores/threads times the elapsed time. There's no double-counting!

In fact, running GPU work you'll tally up less run-time than running CPU work, because with CPU work the CPU usage is at 100% 100% of the time, while with GPU work the CPU is often at less than 100% usage.

Hope I'm less confused now ...

[Snow Crash]

talk about confused <wink> ... what you've posted is not true. Please take a look at your device statistics on WCG. If you are configured to concurrently process more tasks than you have actual cpu threads you WILL get more runtime than the number of threads.

[wplachy]

Only CPU runtime is counted, not GPU runtime.
While the GPU is crunching an individual WU, the CPU is only spending a low % feeding it, so it's not incrementing its runtime tally much at all. You're running more tasks than when doing CPU WUs, but the total CPU run-time can only be as high as the number of CPU cores/threads times the elapsed time. There's no double-counting!

In fact, running GPU work you'll tally up less run-time than running CPU work, because with CPU work the CPU usage is at 100% 100% of the time, while with GPU work the CPU is often at less than 100% usage.

Hope you're nice and confused now ...

talk about confused <wink> ... what you've posted is not true. Please take a look at your device statistics on WCG. If you are configured to concurrently process more tasks than you have actual cpu threads you WILL get more runtime than the number of threads.

WCG awards runtime as: CPU tasks get CPU time; GPU tasks get Elapsed time (wall time). To confirm I verified that my devices running 12 concurrent GPU tasks only are being awarded 11-13 days per cal day.

--------------------------------------------------
Bill P

[BlindFreddie]

Yes, you guys are right, I was wrong. (Edited my post).
So you can manipulate your WCG runtime stats by simply varying the no of concurrent GPU jobs, without much effect, if any, on throughput..
Lies, damn lies & statistics ...

[nanoprobe]

Yes, you guys are right, I was wrong. (Edited my post).
So you can manipulate your WCG runtime stats by simply varying the no of concurrent GPU jobs, without much effect, if any, on throughput..

There is limit to how many GPU tasks you can run simultaneously on any given system before you start to lose efficiency. Depends on your hardware. Trial and error with your system(s) is the only way to know what would be the max for you.

[OldChap]

I have only done this on one rig and my observations are as follows:

q6600 @3.3 and 7950 @ 1325

the absolute best AVERAGE time when running 16 wu's concurrently is about 51 seconds per but the offset gets out of sync quickly which results in that average time dropping to 53 seconds or more.

Moving to 24 concurrent wu's best time seems to be in the order of 52 seconds and offset stays longer but still fails and then the average is maybe 54 seconds or more

With 36 concurrent wu's the best time is probably 54 or 55 seconds average but 24 hours later the offset is generally maintained and the average runtime for a wu is still around 55 seconds.

I have just set for 40 concurrent and will try to report my findings

In each instance above my offset is approximately 35-40 seconds. The idea behind that was that for at least some of the running wu's there will not be wu's in cpu mode at both the middle and end phase at the same time.

Remember this is a 4 core non ht cpu running 100% all the time.

I believe that it may be possible when flooding the cpu like this to have settings that you can set and forget for maybe a week or more at a time. So far it is not perfect but looks promising. Does the answer lie in matching the cpu and gpu runtime for each wu perhaps?

[0ne.shot]

I'm now running 32 WUs with an average completion time of ~22.5-23 min per WU. This should come out to around 42-43 seconds per WU with random offsets between WU groups. This is with my 2700k @ 4.5 GHz.

[nanoprobe]

I'm now running 32 WUs with an average completion time of ~22.5-23 min per WU. This should come out to around 42-43 seconds per WU with random offsets between WU groups. This is with my 2700k @ 4.5 GHz.

What video card are you using and what are your clocks?

[EtaCarinae]

It's amazing. Sam has a much higher point production per thread.
I do 12 million points with 195 days/day and sam does 12 million points with 135 days per day.
I am at 61'538 points/day and Sam is at 88'888 points/day. Much better efficiency per thread, as one day runtime per day is basically one thread per day.
Sam's threads are 44% more efficient than mine. That is astonishing as I am running top gear hardware. Maybe Sam runs Linux and I am with Windows.
I know that Linux is more efficient. Maybe I should change. The problem is that I am a Linux illiterate.

[EtaCarinae]

I have only done this on one rig and my observations are as follows:

q6600 @3.3 and 7950 @ 1325

the absolute best AVERAGE time when running 16 wu's concurrently is about 51 seconds per but the offset gets out of sync quickly which results in that average time dropping to 53 seconds or more.

Moving to 24 concurrent wu's best time seems to be in the order of 52 seconds and offset stays longer but still fails and then the average is maybe 54 seconds or more

With 36 concurrent wu's the best time is probably 54 or 55 seconds average but 24 hours later the offset is generally maintained and the average runtime for a wu is still around 55 seconds.

I have just set for 40 concurrent and will try to report my findings

In each instance above my offset is approximately 35-40 seconds. The idea behind that was that for at least some of the running wu's there will not be wu's in cpu mode at both the middle and end phase at the same time.

Remember this is a 4 core non ht cpu running 100% all the time.

I believe that it may be possible when flooding the cpu like this to have settings that you can set and forget for maybe a week or more at a time. So far it is not perfect but looks promising. Does the answer lie in matching the cpu and gpu runtime for each wu perhaps?

OdChap do you mean that on your machine you crunch an HCC GPU WU in 52 seconds with 16 or more threads.
I am between 7 and 8 minutes with a 7970 GPU and a 990X CPU.
Have I missed something !!!!!

[Snow Crash]

OdChap do you mean that on your machine you crunch an HCC GPU WU in 52 seconds with 16 or more threads.
I am between 7 and 8 minutes with a 7970 GPU and a 990X CPU.
Have I missed something !!!!!

I believe OldChap is taking the average runtime for a WU and dividing it by the number of concurrent threads he is running. How many concurrent threads are you running (16)? I'm guessing you may get better overall numbers by upping the qty ... I've pretty much settled on 24 concurrent on my 7950 + 980X so I only have to manually intervene 1 time every 24-36 hours to make sure not too many WUs are bunched up at cpu processing.

GPU OC has a lot to do with the overall runtimes ... a 7950 at 1325 is quite good. Mine maxs out at 1300. At 1325 I'll start throwing Invalids no matter how high I set the volts.

Last thought ... if you have your cache set too high you may be experiencing "High Prioity" fubar ... sometimes BOINC partially forgets I am running concurrent WUs and changes my estimated completion time to be as if I was crunching 1 at a time yet it remembers I am runing 24 "cores" so the result is I get flooded with new WUs until it settles back into the real runtime again (sometimes a couple seconds, sometimes a couple minutes). If I have .5 day cache this does not disrupt processing but if you are much higher you might have WUs preempt WUs already running and in that even the checkpoint is only at 50% so you will lose any time crunched between 0-49% and 51-99%

[OldChap]

Yep, when I talk about AVERAGE runtime then what I mean is .....

I take the total number of threads running and suspend internet until all have completed. Today that is 40 wus so I will end up adding together all 40 times then dividing by 40 for the average time it takes a single thread to complete.(today that is 35 mins and about 13.5 seconds). However during the time it takes a single thread to complete I have completed 39 other threads too so I take the average single thread time and divide by 40 for the AVERAGE time between completed wu's. (52.83 secs for my last 40 today) with no adjustment for about 24 hours.

Is this the best way? I don't really know but it does give me something to work with when I look at daily production:

24*60*60 = 86400 seconds a day which I divide by that AVERAGE to complete a wu time of say 52.83 secs is around 1635 wu's a day so I keep an eye on this ..... It is not helped by the problem of varying PV numbers but overall I see what I need to doing this. Using this method works best for me if I run new settings for a week.

I am sure that I have explained things better than this before now but I hope you get the gist

[bluestang]

Or just take 60 minutes divided by the average total elapsed time in BOINC for a WU and multiply by the number of concurrent WU's your running and then by 24 hours....

60/35.225*40*24=1635

[OldChap]

I can do the math...it is just not pretty.... usually

[nanoprobe]

It's amazing. Sam has a much higher point production per thread.
I do 12 million points with 195 days/day and sam does 12 million points with 135 days per day.
I am at 61'538 points/day and Sam is at 88'888 points/day. Much better efficiency per thread, as one day runtime per day is basically one thread per day.
Sam's threads are 44% more efficient than mine. That is astonishing as I am running top gear hardware. Maybe Sam runs Linux and I am with Windows.
I know that Linux is more efficient. Maybe I should change. The problem is that I am a Linux illiterate.

If you're using your GPUs like the post I saw at WCG (10 GPU and 2 CPU threads on a hexcore with HT) then you are not using anywhere near the potential of your 7970s. Take just one and try this so you can compare. Set BOINC to receive no new tasks. If you're still using an app_info file finish all the tasks you have in your cache first and delete any CPU tasks that are left when the GPU tasks finish. Use a separate WCG profile for that machine and set it to receive/run no CPU tasks.
Upgrade your BOINC client to 7.0.56 After it starts just shut it down for now.
Copy and paste the text below into notepad. Rename it app_config.xml and place it into the WCG project folder. (That will give you 36 GPU tasks each using 1/3 of a CPU core) Delete the app-info file or at least remove it from the folder. Restart BOINC. Set you cache to 0.50 days to get started. Set BOINC to allow new tasks.
Let Boinc download maybe 100 tasks, set Boinc to receive no more tasks and suspend the ones you have(Including the ones that are running) from the tasks page not the project page. The easiest way to do that is highlight the first task, scroll down to the bottom hold down the shift key and highlight the last one. That should highlight them all then click suspend. Now highlight the first task only and click resume. Every time you see the percentage done start to move highlight and resume the next task. Do this until you have all 36 running. It will take a while but this will keep you tasks spread out for better efficiency. Once you get 36 running mass resume the rest and up your cache to whatever you'd like. Your hexcores should be able to knock out 36 tasks in 25-30 minutes, maybe less. (for a reference my 7970 does 24 tasks every 20 minutes on a 2600k running @ 4.2 GHz. with the GPU @ 115 core clock) Depends on your hardware and any overclocks you might be running. Most 7970s can run @ 1125 core clock on stock voltage. With a GPU configuration only I think you should see 1.3-1.5 mill WCG PPD on that cruncher.

<app_config>
<app>
<name>hcc1</name>
<gpu_versions>
<gpu_usage>.0277</gpu_usage>
<cpu_usage>0.333</cpu_usage>
</gpu_versions>
</app>
</app_config>

[EtaCarinae]

If you're using your GPUs like the post I saw at WCG (10 GPU and 2 CPU threads on a hexcore with HT) then you are not using anywhere near the potential of your 7970s. Take just one and try this so you can compare. Set BOINC to receive no new tasks. If you're still using an app_info file finish all the tasks you have in your cache first and delete any CPU tasks that are left when the GPU tasks finish. Use a separate WCG profile for that machine and set it to receive/run no CPU tasks.
Upgrade your BOINC client to 7.0.56 After it starts just shut it down for now.
Copy and paste the text below into notepad. Rename it app_config.xml and place it into the WCG project folder. (That will give you 36 GPU tasks each using 1/3 of a CPU core) Delete the app-info file or at least remove it from the folder. Restart BOINC. Set you cache to 0.50 days to get started. Set BOINC to allow new tasks.
Let Boinc download maybe 100 tasks, set Boinc to receive no more tasks and suspend the ones you have(Including the ones that are running) from the tasks page not the project page. The easiest way to do that is highlight the first task, scroll down to the bottom hold down the shift key and highlight the last one. That should highlight them all then click suspend. Now highlight the first task only and click resume. Every time you see the percentage done start to move highlight and resume the next task. Do this until you have all 36 running. It will take a while but this will keep you tasks spread out for better efficiency. Once you get 36 running mass resume the rest and up your cache to whatever you'd like. Your hexcores should be able to knock out 36 tasks in 25-30 minutes, maybe less. (for a reference my 7970 does 24 tasks every 20 minutes on a 2600k running @ 4.2 GHz. with the GPU @ 115 core clock) Depends on your hardware and any overclocks you might be running. Most 7970s can run @ 1125 core clock on stock voltage. With a GPU configuration only I think you should see 1.3-1.5 mill WCG PPD on that cruncher.

<app_config>
<app>
<name>hcc1</name>
<gpu_versions>
<gpu_usage>.0277</gpu_usage>
<cpu_usage>0.333</cpu_usage>
</gpu_versions>
</app>
</app_config>

Let's see I will do some tests on one machine. But my past experience is that very quickly you start getting unstable systems, with some compute errors and some window restarts. I generally favor les extreme configurations but that last for weeks, months without interruption. But as GPU is coming to an end mybe it is time to try to really become an Hypernova

[nanoprobe]

Let's see I will do some tests on one machine. But my past experience is that very quickly you start getting unstable systems, with some compute errors and some window restarts. I generally favor les extreme configurations but that last for weeks, months without interruption. But as GPU is coming to an end mybe it is time to try to really become an Hypernova

Running this configuration should be no sweat for your machines. I've been running my version since the app_config system came out. Not a single error, invalid or problem. One thing you may need to do is a reg fix for driver timeouts that is a known issue in windows. If you see tasks where the runtime is moving but the completion % is not moving then it has probably had a driver reset. The fix is very simple and I've done it to all my machines. Copy all the text below and paste it into notepad. Rename it fix.reg. Right click on it and open with Registry Editor. You'll get warnings about editing the registry, just click yes, and you're done.

Windows Registry Editor Version 5.00

[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Contro l\Watchdog]
"DisableBugCheck"="1"

[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Contro l\Watchdog\Display]
"EaRecovery"="0"

[my87csx481]

If you're using your GPUs like the post I saw at WCG (10 GPU and 2 CPU threads on a hexcore with HT) then you are not using anywhere near the potential of your 7970s. Take just one and try this so you can compare. Set BOINC to receive no new tasks. If you're still using an app_info file finish all the tasks you have in your cache first and delete any CPU tasks that are left when the GPU tasks finish. Use a separate WCG profile for that machine and set it to receive/run no CPU tasks.
Upgrade your BOINC client to 7.0.56 After it starts just shut it down for now.
Copy and paste the text below into notepad. Rename it app_config.xml and place it into the WCG project folder. (That will give you 36 GPU tasks each using 1/3 of a CPU core) Delete the app-info file or at least remove it from the folder. Restart BOINC. Set you cache to 0.50 days to get started. Set BOINC to allow new tasks.
Let Boinc download maybe 100 tasks, set Boinc to receive no more tasks and suspend the ones you have(Including the ones that are running) from the tasks page not the project page. The easiest way to do that is highlight the first task, scroll down to the bottom hold down the shift key and highlight the last one. That should highlight them all then click suspend. Now highlight the first task only and click resume. Every time you see the percentage done start to move highlight and resume the next task. Do this until you have all 36 running. It will take a while but this will keep you tasks spread out for better efficiency. Once you get 36 running mass resume the rest and up your cache to whatever you'd like. Your hexcores should be able to knock out 36 tasks in 25-30 minutes, maybe less. (for a reference my 7970 does 24 tasks every 20 minutes on a 2600k running @ 4.2 GHz. with the GPU @ 115 core clock) Depends on your hardware and any overclocks you might be running. Most 7970s can run @ 1125 core clock on stock voltage. With a GPU configuration only I think you should see 1.3-1.5 mill WCG PPD on that cruncher.

<app_config>
<app>
<name>hcc1</name>
<gpu_versions>
<gpu_usage>.0277</gpu_usage>
<cpu_usage>0.333</cpu_usage>
</gpu_versions>
</app>
</app_config>

Holy crap. 36 simultaneous tasks???? Please forgive my ignorance. I have zero time for playing/reading/learning....for at least a month now....and probably not for another month. I am not familiar with EtaCarinae's systems. Would you recommend this same app_config for a 990x with a 7970? (cpu is at 173 x 25 = 4325, gpu core seems stable at 1240)

[Snow Crash]

Holy crap. 36 simultaneous tasks???? Please forgive my ignorance. I have zero time for playing/reading/learning....for at least a month now....and probably not for another month. I am not familiar with EtaCarinae's systems. Would you recommend this same app_config for a 990x with a 7970? (cpu is at 173 x 25 = 4325, gpu core seems stable at 1240)

ET is running 990X w/ 7970 so yes, crank your concurrent count up

[nanoprobe]

Holy crap. 36 simultaneous tasks???? Please forgive my ignorance. I have zero time for playing/reading/learning....for at least a month now....and probably not for another month. I am not familiar with EtaCarinae's systems. Would you recommend this same app_config for a 990x with a 7970? (cpu is at 173 x 25 = 4325, gpu core seems stable at 1240)

You may have to back off on the GPU a tad when running 36 tasks. 1240 is pushing it pretty hard. Keep an eye on your WCG results page for errors/invalids.