Running time very differents between two computers

marsinph
marsinph
Joined: 11 Jun 06
Posts: 18
Credit: 67,225,714
RAC: 28
Topic 212832

I have 3 computers, all with a I7-2600K without overclocking
Alle have 8Gb RAM from the perfekt same RAM.
The name of my computers are in fact the model of mainboard.
Also the same SDD, also identically project preferences (crunching, CPU use,....)!
All runs with perfect the same Win7 version. (Ultimate/64 bits)

The only difference is a GPU GeForce 950 and GeForce 660TI) all with 2Gb DDR5, Open CL1.2

For the similar WU ( Gamma ray pulsar binary search#1  on GPU 1.20 FGPR open cl1k Nvidia)
WU : LATeah0511_1196.....

On one it take about 35-40 min to run and on other more than one hour
All computers nothing else runs and are not in use.

Only computer P8Z68V is in use. Strange, this it the one with the less crunching time.Who can explain ?

Who can explain ???

 

archae86
archae86
Joined: 6 Dec 05
Posts: 3,021
Credit: 4,990,453,245
RAC: 3,036,543

The 660 Ti is on the Nvidia

The 660 Ti is on the Nvidia Kepler series, which is architecturally considerably different than the second generation Maxwell series of which your 950 is a member.  

I recall that some efficiency improvements in the Einstein GPU code were much more effective on some generations than others.  Perhaps someone else will comment here on whether Kepler is among the disadvantaged. 

By the way, unless we are logged in to your account, your computer names are not visible to us.  If you want to communicate more clearly, use host numbers instead.  For example, host 12596445.

 

 

Richie
Richie
Joined: 7 Mar 14
Posts: 638
Credit: 1,701,135,370
RAC: 63,134

Yep, Nvidia 600- and also

Yep, Nvidia 600- and also 700-series cards became slow for Einstein at some point. I had GTX 780 but quit using it here because it was much slower than GTX 960 and required much more juice from the wall. That just didn't feel good anymore.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,021
Credit: 4,990,453,245
RAC: 3,036,543

An exception to the 600 and

An exception to the 600 and 700-series comment is the 750 (and 750 Ti) which is actually the first generation Maxwell, not a Kepler, and really not in a series with 760, 770, 780 cards.  While it is an economy card, the performance at introduction was quite respectable for the price, and the power efficiency was exceptionally good.  Even now, several years after introduction, the power efficiency on Einstein work of the 750 is quite respectable, though you can get more out of a host with a higher capability card from either AMD or Nvidia.  Thus the system level power efficiency may not be competitive, even though the card level efficiency is.

On the other hand, a 660 running here will burn a lot of power per unit output.  Possibly it would best be put to another use.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,494
Credit: 65,669,778,319
RAC: 54,168,073

Some points to consider about

Some points to consider about current GPU crunching at Einstein.

It's pretty much all about the GPU.  The CPU supporting it is relatively unimportant as long as you have sufficient cores available to support the GPU when it needs it.  For nvidia GPUs, that means a core not allocated to something else for each GPU task running concurrently.  I didn't see any evidence of a problem in that regard on your hosts.

You have 3 machines with 3 different GPUs - GTX570, GTX660Ti, GTX950.  The rough crunch times are 2400s, 3800s and 2100s respectively.  I'm assuming that you are not running more than a single task on each GPU.  Whilst I've never used those particular models, I have used 550Ti, 650,  650Ti and 750Ti so I'm not at all surprised to see a 570 doing better than a 660Ti and a 950 doing better than both of those.

Former Einstein apps used CUDA and Kepler cards were good performers then.  With the current OpenCL apps (no CUDA apps have been produced) my Kepler cards are poor performers and I've retired them all.  When I last ran them, a 550Ti was faster than either a 650 or a 650Ti (OpenCL app), but the problem was higher power consumption for what was still a very low output.

In perusing the tasks lists for all your hosts, I notice you are aborting around half the tasks you are receiving.  Please change your work cache settings to control how much work is downloaded rather than continually aborting the excess.  If you are aborting just to see if a new batch of tasks might perform better than the previous batch, you are wasting your time and, more importantly, you are wasting server resources which are already under considerable strain.  In general, all GPU tasks for all data series have much the same crunch time so you wont get an appreciable change by doing an exchange.

If you wanted to see a big improvement in output for less power consumed, you could achieve that by replacing the 660Ti with a relatively cheap, low power, modern card.  I've just tried out an Asus RX 560 which uses motherboard power only - no PCIe power cable needed.  It's taking about 1100s per task - around 75-80 tasks per day.  The CPU is a G4560 Pentium dual core (2 cores, 4 threads).

 

Cheers,
Gary.

marsinph
marsinph
Joined: 11 Jun 06
Posts: 18
Credit: 67,225,714
RAC: 28

Hello, I have use a "not

Hello,

I have use a "not logged" comùputer, my computer are visible. I also checked my privacy settings. It is "on"
So here the links to josts omputers :
Z77 : https://einsteinathome.org/host/12596725

P8Z68V :  https://einsteinathome.org/host/12596445

aximus : https://einsteinathome.org/host/12617680

 

[img]http://boincstats.com/signature/-1/user/8912/sig.png[/img]

marsinph
marsinph
Joined: 11 Jun 06
Posts: 18
Credit: 67,225,714
RAC: 28

Hello Gary. Yes I aborted a

Hello Gary.

Yes I aborted a lot. Why ?
I also run Seti (10 days and 10 additional days.

My Einstein preferences are set to styore 0.5 days and no any additionnal time.
But, if I let do, I receive WU for much more than 0.5 days.
So I abort, to not let the WU going to "out if time". So my aborted are very fast available for other users.
Doing so, itr reduce the amount of pending tasks.

%20sQAAAABJRU5ErkJggg%3D%3D

 

 

 

3R49N6vzXXAAAAAElFTkSuQmCC

[img]http://boincstats.com/signature/-1/user/8912/sig.png[/img]

Holmis
Joined: 4 Jan 05
Posts: 1,118
Credit: 1,049,271,676
RAC: 947,585

marsinph wrote:Hello

marsinph wrote:

Hello Gary.

 

Yes I aborted a lot. Why ?
I also run Seti (10 days and 10 additional days.

 

My Einstein preferences are set to styore 0.5 days and no any additionnal time.
But, if I let do, I receive WU for much more than 0.5 days.
So I abort, to not let the WU going to "out if time". So my aborted are very fast available for other users.
Doing so, itr reduce the amount of pending tasks.

The work cache setting is a global setting that applies to all your projects, it can not be set differently for different projects. Boinc will communicate the changes between projects and the latest set preference will be used.
The only way to achieve what you describe above without aborting a lot of tasks is to micro manage by using "No new tasks" in Boinc and increasing and decreasing the cache settings.

marsinph
marsinph
Joined: 11 Jun 06
Posts: 18
Credit: 67,225,714
RAC: 28

Hello Holmis. It is what I

Hello Holmis.
It is what I already do !
The generic preference are used on host 12617680. But I changed advanced setting to 1 day WU
I mean : now new tasks, then, on Boinc Manager / preference : 1 day and no additional.
Then manual update. Then Boinc Manager wait 60 second to the next synchro.
I wait, Then I check again my preference on the Web. It is on "one day" and no additionnal.
The, I accept new task in Boinc manager. Manual synchro. And then I get each time about 180 WU of about
a little more than one hr (64 min) computation time. With a deadline about 10-12 days.
So about 9 days WU received. OK. But !!!
Because I run also Seti (50% time each,) I got so WU for 18 days with a deadline of about 10 !!!
On my 2 other hosts, on SETI, I use the "home" preferences with 10 days and 10 addtitional.

[img]http://boincstats.com/signature/-1/user/8912/sig.png[/img]

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,494
Credit: 65,669,778,319
RAC: 54,168,073

marsinph wrote:It is what I

marsinph wrote:
It is what I already do !

If you are prepared to 'micro-manage' by adjusting local work cache settings and using NNT, there are a few things to understand about how it all works.

Quote:
... I changed advanced setting to 1 day WU
I mean : now new tasks, then, on Boinc Manager / preference : 1 day and no additional.

When you do this, the new setting is recorded in a file called global_prefs_override.xml which is stored locally in your BOINC directory.  Local pref values take precedence, so it does not matter what is stored for that parameter on the website.  The website value will not be changed - it will always be ignored while there is a value in the override file.

Quote:
Then manual update. Then Boinc Manager wait 60 second to the next synchro.
I wait, Then I check again my preference on the Web. It is on "one day" and no additionnal.

I can only repeat - A local pref change using BOINC Manager will remain local and will NOT change a value on the website.  It will just 'override' a website value whatever it was.

Quote:
The, I accept new task in Boinc manager. Manual synchro. And then I get each time about 180 WU of about
a little more than one hr (64 min) computation time. With a deadline about 10-12 days.

If you truly have set a local value of 1 day, you cannot get 180 tasks if they have an estimate of around 1 hour.  You would get a total of around 24 tasks over several consecutive work fetch events.  If a task was incorrectly estimated at something like 8 minutes rather than 1 hour, you could end up with around 180 tasks.  The deadline for these tasks is 14 days, not 10-12.

Quote:
So about 9 days WU received. OK. But !!!
Because I run also Seti (50% time each,) I got so WU for 18 days with a deadline of about 10 !!!
On my 2 other hosts, on SETI, I use the "home" preferences with 10 days and 10 addtitional.

There is likely to be something wrong with the way you are setting local preferences.  It also looks like the host might not be in the 'location' (venue) you think it is in.  You say 'generic' but the work fetch behaviour looks more like 'home'.  If both generic and home have 10/10 settings on the website, that could explain it if you haven't set local prefs correctly.

You should be able to use local prefs and NNT to achieve what you want.  Location is relatively unimportant.  All you really need to do is make sure Einstein is set to NNT any time you want to change the local work cache setting to more than 1 day.

What I would do is make a local setting change through BOINC Manager (I assume you are correctly saving the local change?) and then browse the global_prefs_override.xml file to check that the new value is recorded there.  For a 1/0 days setting, that file should contain the lines

<work_buf_min_days>1.000000</work_buf_min_days>
<work_buf_additional_days>0.000000</work_buf_additional_days>

If you can make a change through the BOINC Manager interface and then see the changed value in the override file, then everything should work the way you intend.  Once you have used ANT (Allow New Tasks) for Einstein to get a day's worth, immediately set NNT again and leave it there while you get whatever you want for Seti.  Never change to ANT until you have double checked that your local setting is back on 1 day.

 

Cheers,
Gary.

mikey
mikey
Joined: 22 Jan 05
Posts: 7,734
Credit: 621,900,487
RAC: 104,008

marsinph wrote:Hello

marsinph wrote:
Hello Holmis.
It is what I already do !
The generic preference are used on host 12617680. But I changed advanced setting to 1 day WU
I mean : now new tasks, then, on Boinc Manager / preference : 1 day and no additional.
Then manual update. Then Boinc Manager wait 60 second to the next synchro.
I wait, Then I check again my preference on the Web. It is on "one day" and no additionnal.
The, I accept new task in Boinc manager. Manual synchro. And then I get each time about 180 WU of about
a little more than one hr (64 min) computation time. With a deadline about 10-12 days.
So about 9 days WU received. OK. But !!!
Because I run also Seti (50% time each,) I got so WU for 18 days with a deadline of about 10 !!!
On my 2 other hosts, on SETI, I use the "home" preferences with 10 days and 10 addtitional.

You need another pc to run just Seti so you can stop micromanaging.

Comment viewing options

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