Advanced search

Message boards : Graphics cards (GPUs) : Raising the Cache Limit !!!

Author Message
STE\/E
Send message
Joined: 18 Sep 08
Posts: 366
Credit: 268,368,907
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 10597 - Posted: 16 Jun 2009 | 11:38:57 UTC

The Cache Limit for now is @ 4 (If you Have a Quad Core CPU, 8 if you have a i7 Type CPU) & thats fine if you have just 1 Single Core CUDA Capable Video Card but short changes you if you have say a GTX 295 or even Multiple GTX 295's which I plan on having in the next few days & some people have already.

My Question/Proposal is the amount of Cache should Depend on the amount of CUDA Capable Cores you have and not the CPU I don't think. I plan on adding 2 more GTX 295's to a Box that already has a GTX 295/280/260 in it already. The 208 & 260 are going in other Box's and then 3 GTX 295's will be in that Box.

Well that's only going to leave me 2 Spare Wu's in the Cache while the 3 295's Chew on 6 Wu's at a time. What that means is the slighest little Blip in the Sever & I'm out of work in no time. I think a Box with 1 295 in it should have a 8 Wu Cache 2 295's a 16 Wu Cache etc ... :)

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 10611 - Posted: 16 Jun 2009 | 18:51:22 UTC - in response to Message 10597.

I agree.
GPU-Grid WU caching should be dependent on the Clients GPU card number and specs, not their CPU's.

A few simple reasons why you might not get a task:

The server is down/offline/unreachable for a while (GPU-Grid side).
There are no new tasks!!!
Your Internet is down for a couple of hours (temporary ISP issues).
Your wireless router is playing up.
And more annoyingly I have them in my queue:-

I have an 8800GT in one of my systems (only 60GFLOPS)! I dont need 4 tasks. One in the queue is sufficient for that card. Too many tasks are downloaded and then deleted from that system before they even begin. Its just pointless. So its obvious whose cache they should be in!

How about this idea/suggestion?
Calculate the Cache quantity to download according to the average time the computer is on and the Boinc GFLOPS ratings, (or WU turnover). If there are 2, 3, 4 or even 6 ratings (3x295)and the cards are high performers, then download and cache 3 or 4 WU's per rating. If it is a low performer, 1 or 2 in the cache per card rating is enough.

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1947
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 10615 - Posted: 16 Jun 2009 | 23:10:37 UTC - in response to Message 10611.

We will be testing to assign #jobs = 2*#gpus this week.

gdf

STE\/E
Send message
Joined: 18 Sep 08
Posts: 366
Credit: 268,368,907
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 10616 - Posted: 16 Jun 2009 | 23:21:48 UTC

Thank's GDF ... :)

Profile Paul D. Buck
Send message
Joined: 9 Jun 08
Posts: 1050
Credit: 37,321,185
RAC: 0
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 10618 - Posted: 17 Jun 2009 | 4:44:27 UTC

In a way I don't want them to raise the numbers too much in that though I like to have more work cached, I don't want to get so much that I cannot process it in time to get the bonus ... :)

Though on my fastest system with 4 GPU cores running, 4 spare tasks means I have only a total of about 12-14 hours worth of work on hand ...

More interesting is MW can't seem to make up their minds as to what they are going to do yet ... sigh ...

STE\/E
Send message
Joined: 18 Sep 08
Posts: 366
Credit: 268,368,907
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 10623 - Posted: 17 Jun 2009 | 8:46:11 UTC

Paul, I'm not asking for more than what is being given out now, 4 Wu's for a Single CUDA Core Computer. All I'm asking for is if you have say single 295 (2 CUDA Cores) that you get an 8 WU Cache, a Double 295 Equipped Computer (4 CUDA Cores) you get a 16 Wu Cache.

Like any other Project if your getting to much to handle then you can lower your Connect Time and or your Resources to the Project. Why I'm asking this is because either today or tomorrow I'll have a Triple 295 Box that will only be able to have 8 Wu's on it at one time. Six of those Wu's will be running all the time so I'll only have a 2 Wu Cache. Not much backup if the Server hiccups at all IMO ... :)

Profile Paul D. Buck
Send message
Joined: 9 Jun 08
Posts: 1050
Credit: 37,321,185
RAC: 0
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 10627 - Posted: 17 Jun 2009 | 14:06:52 UTC - in response to Message 10623.

Paul, I'm not asking for more than what is being given out now, 4 Wu's for a Single CUDA Core Computer. All I'm asking for is if you have say single 295 (2 CUDA Cores) that you get an 8 WU Cache, a Double 295 Equipped Computer (4 CUDA Cores) you get a 16 Wu Cache.

Like any other Project if your getting to much to handle then you can lower your Connect Time and or your Resources to the Project. Why I'm asking this is because either today or tomorrow I'll have a Triple 295 Box that will only be able to have 8 Wu's on it at one time. Six of those Wu's will be running all the time so I'll only have a 2 Wu Cache. Not much backup if the Server hiccups at all IMO ... :)

Noted ...

And I am envious ... :)

STE\/E
Send message
Joined: 18 Sep 08
Posts: 366
Credit: 268,368,907
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 10636 - Posted: 17 Jun 2009 | 20:51:08 UTC - in response to Message 10627.
Last modified: 17 Jun 2009 | 21:22:17 UTC

Paul, I'm not asking for more than what is being given out now, 4 Wu's for a Single CUDA Core Computer. All I'm asking for is if you have say single 295 (2 CUDA Cores) that you get an 8 WU Cache, a Double 295 Equipped Computer (4 CUDA Cores) you get a 16 Wu Cache.

Like any other Project if your getting to much to handle then you can lower your Connect Time and or your Resources to the Project. Why I'm asking this is because either today or tomorrow I'll have a Triple 295 Box that will only be able to have 8 Wu's on it at one time. Six of those Wu's will be running all the time so I'll only have a 2 Wu Cache. Not much backup if the Server hiccups at all IMO ... :)

Noted ...

And I am envious ... :)


Well it's not working out like I thought it would, apparently the Motherboard (Gigabyte Extreme X58) isn't up to running 3 GTX 295's, at least not all 6 Cores Anyway. Unless somebody knows something I don't know I'll have to pull 1 and replace it with the GTX 280 that came out of it.

The Motherboard will only recognize 5 of the Cores & the Fan on the 3'rd/Bottom Card isn't even Spinning (Not Good). It sees 5 of the Cores in BOINC & the System Properties/Hardware/Device manager/Display Adapters Display Page but for some reason doesn't see the 6'th Core ???

I think it's the Middle Card that it's not reading boath cores on because when I swapped out the Bottom 295 with a 260 when it booted up it was reading the 260 but only 3 of the 4 295 Cores. I'm going to put the 2 295's in the Top & Bottom Slots & the 260 or 280 in the Middle & it should read all 5 Cores that way.

Profile [AF>Libristes>Jip] Elgran...
Avatar
Send message
Joined: 16 Jul 08
Posts: 45
Credit: 78,618,001
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 10665 - Posted: 19 Jun 2009 | 11:59:39 UTC - in response to Message 10636.

Raising the cache limit test in progress.

STE\/E
Send message
Joined: 18 Sep 08
Posts: 366
Credit: 268,368,907
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 10679 - Posted: 19 Jun 2009 | 14:39:01 UTC - in response to Message 10665.

Raising the cache limit test in progress.


I guess it is, but it's to High now or something else isn't right. My 4 CUDA Core i7 has about 24 Wu's on it which isn't bad but one of my 2 CUDA Core Box's has close to 50 on it. I set to NNW for now and will have to Abort some of them I think so they can be sent out to someone else
...

Profile [AF>Libristes>Jip] Elgran...
Avatar
Send message
Joined: 16 Jul 08
Posts: 45
Credit: 78,618,001
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 10692 - Posted: 19 Jun 2009 | 17:43:18 UTC - in response to Message 10679.

According to this topic Task Quota Change? , it seems to be a bug in the new server software.

Post to thread

Message boards : Graphics cards (GPUs) : Raising the Cache Limit !!!

//