Meanwhile, I wonder why the Windows port of GPUGRID loads the CPU so...

Could it be that my GPUs are underutilized?
Or is the Windows software that inefficient?

No, your Linux should be fine!

Trying a short explanation: the GPU crunches for some 10th of milliseconds and when it's done the CPU needs to intervene. If it doesn't the GPU runs dry. Under linux the project uses "nanosleep", don't know if it's a function or library or whatever.. anyway, it allows to send the cpu task into sleep mode and wake it up at a precisely controlled time.
There is no nanosleep for windows and the underlying reason seems to be that the smallest time steps the windows scheduler knows is 1 ms, so the timing control is much less efficient (and the polling has to be more agressive, otherwise GPU performance suffers).

MrS

---

Scanning for our furry friends since Jan 2002

Scott, are your cpu usage numbers set to 100% = 1 core? Otherwise they would look strange.

Was writing the earlier posts fairly quickly before having to leave home to do some work. Anyway, the percentage usage is from the windows task manager processes list.

And the very slow responsiveness you're seeing is due to your relatively slow card (32 shaders vs recommended 50+). Nothing has changed here during the different client versions. But the WUs have changed, now we also crunch more complex models, which take even longer to process (and thus make the lag worse)

MrS

While the effects on former app versions was more evident on the slower card (vs. for example my 9600GSO), these were always relatively modest effects. Also, with 6.55 app versions, this box crunched a few of the different types (US, USPME, GPUTEST, JAN) of work with no real differences to note other than ms/time step and overall run time. It is only with the 6.61 app version that the machine is so badly affected.

Well, here is an interesting twist:

http://www.gpugrid.net/workunit.php?wuid=184242

I thought that work was tied to the application version, but I guess not. The two machines that errored out this unit before I got it list the app version as 6.59 and 6.61, respectively?

6.59 is the linux equivalent of 6.61 for win ;)

Anyway, the percentage usage is from the windows task manager processes list.

You have a dual core cpu, so the default windows behaviour should be to show 100% as both cores under full load and 50% for one core under full load. To exceed 50% one has to go multithreaded. When you say you se cpu usages of up to 73% that would mean 1 and a half core are used. Nobody else reported something like that before for GPU-Grid.

And regarding the other issue: so it seems the different WU types are not causing your massive slow down and it looks like the new client is really to blame for this.

MrS

---

Scanning for our furry friends since Jan 2002

6.59 is the linux equivalent of 6.61 for win ;)

Ah...got it!

You have a dual core cpu, so the default windows behaviour should be to show 100% as both cores under full load and 50% for one core under full load. To exceed 50% one has to go multithreaded. When you say you se cpu usages of up to 73% that would mean 1 and a half core are used. Nobody else reported something like that before for GPU-Grid.

And regarding the other issue: so it seems the different WU types are not causing your massive slow down and it looks like the new client is really to blame for this.

MrS

I have tried to test it to reproduce the more than 50% load, and haven't been able to do so. Maybe it was just something screwy with that particular workunit. I downloaded a new one and it has a load of 46-50% and is fairly constant.

Unfortunately, it still has the continued slow-down problems. Also, I think you were correct in your initial assessment of more difficult work. The last unit I downloaded gives time to completion estimates that are twice what any of the pre-6.61 units were. Old work would averaged about 2.5% complete per hour on my 9500GT, but the current work is running at less than half that speed.

They're trying to compensate for the increased model complexity by including less steps in each WU, so that the total completion time is more or less constant (12h on a 9800GT). This may or may not work well on different hardware, i.e. slower cards could have a disproportionately slower memory interface or something like that. But I guess a factor of 2 is not expected.

MrS

---

Scanning for our furry friends since Jan 2002

...i.e. slower cards could have a disproportionately slower memory interface or something like that. But I guess a factor of 2 is not expected.

I thought about that as a possibility, especially since the 9500GT is a 128-bit card. However, that should mean that on my 9600GSO (192-bit) I would see a more modest effect, but so far it seems to be no different than before the 6.61 apps (both have GDDR3, with the 9500GT having a bit faster memory clock).

I am actually wondering if there could be a CPU pattern emerging. It seems that several of the above notes in this thread, including mine, are with AMD machines and not Intel's (indeed, my 9600GSO is in an Intel box). There are some significant architectural differences between those two, but I have no idea how they might play into differences in CPU - GPU interactions?

I'd say the GPU-CPU interaction is, although time critical, rather limited. I.e. any differences would emerge from differences in the chipsets and drivers rather than the CPUs themselfs.

And I'd say we need more data, statistically relevant numbers, to be certain about such performance changes. Not to disregard your finding and your concerns, but watching the progress bar of one WU is not yet enough ;)

MrS

---

Scanning for our furry friends since Jan 2002

While the effects on former app versions was more evident on the slower card (vs. for example my 9600GSO), these were always relatively modest effects. Also, with 6.55 app versions, this box crunched a few of the different types (US, USPME, GPUTEST, JAN) of work with no real differences to note other than ms/time step and overall run time. It is only with the 6.61 app version that the machine is so badly affected.

V6.61 is a BIG step backwards here too. More CPU usage, noticeably poorer video responsiveness. Instead of improving, things are getting worse. V6.56 worked great on Win32, everything since (including the downgrade to v6.55) has been a step backward for the users IMO.

I wish developers to be lucky and smart ot less than f@h founders! Wish GPUGRID luck in this New Year!

And I wish to crunch on 6.62 app version, with CPU usage less than 10% on one core......

So, what about "low CPU" app for Windowzz?

Well, it appears to have definitely been the 6.61 app. The 6.62 app is now crunching away happily on the 9500GT at the same or better speeds than were observed with 6.55. The CPU load is around 3%, and the machine is fairly smooth in usage with the severe slowdown seen under 6.61 now absent.

My machine is now again usable since v6.62. With v6.61 it was useless except for crunching. CPU usage has gone from 22% on the quad to 1%. Thanks, v6.62 is GREAT!

6.62 is amazing...running in +/- 3-4 mins of Q9950 CPU time and 2x GTX 260 Core 216 Superclocked video cards. My CPU contributions are climbing again. 6.62 looks like a winner to me! Thanks GDF & Co.


BTW, I am averaging 9K of points per card per day and climbing still.

Neil

---

Crunching for the benefit of humanity and in memory of my dad and other family members.