20 WU per core - not per host

Author	Message
meshmar Send message Joined: 29 Aug 07 Posts: 7 Credit: 187,002 RAC: 0	Message 1315 - Posted: 3 Jan 2008 \| 14:21:25 UTC
My 400MHz K6-2 gets 20 WU My 3.2GHz Athlon 64 X2 gets 20 WU It takes an average of 66000 seconds of time for the K6-2 to complete 20 short WU and about 200000 for long WU. It takes the Athlon 1750 for short WU and 4500 for long WU. The reasoning behind the per host limit is time of return of the WU. This is an extreme example, but it shows that issuing 20 WU per host would not affect timely return of crunched WUs ... ____________
ID: 1315 \| Rating: 0 \| rate: / Reply Quote

Webmaster Yoda Send message Joined: 21 Dec 07 Posts: 69 Credit: 7,048,412 RAC: 0	Message 1319 - Posted: 3 Jan 2008 \| 14:52:24 UTC
20 per core/cpu would be good. Some computers (4 cores or more) can go through 20 (short) work units in less than 15 minutes so they have to contact the server very often. If it was 20 per core (provided it could be done), they could still get the work done in about an hour but with fewer connections to the server. It would make no difference to slower, single core machines but could prove more efficient on multi core machines. ____________ Join the #1 Aussie Alliance on MilkyWay!
ID: 1319 \| Rating: 0 \| rate: / Reply Quote

LiborA Send message Joined: 15 Sep 07 Posts: 15 Credit: 5,312,625 RAC: 2,427	Message 1511 - Posted: 16 Jan 2008 \| 14:19:55 UTC
Yes, 20 WUs per core will be better :-)
ID: 1511 \| Rating: 0 \| rate: / Reply Quote

Webmaster Yoda Send message Joined: 21 Dec 07 Posts: 69 Credit: 7,048,412 RAC: 0	Message 1539 - Posted: 20 Jan 2008 \| 2:24:38 UTC Last modified: 20 Jan 2008 \| 2:26:58 UTC
Alternatively, if it has to stay at 20 per host, perhaps the backoff time could be reduced. Milkyway@home 20/01/2008 11:09:18 AM Message from server: No work sent Milkyway@home 20/01/2008 11:09:18 AM Message from server: (reached per-host limit of 20 tasks) Milkyway@home 20/01/2008 11:09:18 AM Deferring communication for 20 min 0 sec Milkyway@home 20/01/2008 11:09:18 AM Reason: requested by project The problem with this setting is that, when only short (2 credit) work units are sent, this computer (and others) goes through them in less than 20 minutes (typically 13-14 minutes on this particular computer). It runs out of work before it is allowed to connect again, so either goes idle or switches to another project (and may not switch back for an hour). On a dual quad (which I don't have), 20 work units might only take 7 or 8 minutes... ____________ Join the #1 Aussie Alliance on MilkyWay!
ID: 1539 \| Rating: 0 \| rate: / Reply Quote

teemac Send message Joined: 3 Jan 08 Posts: 5 Credit: 1,377,460 RAC: 0	Message 1541 - Posted: 20 Jan 2008 \| 2:38:41 UTC
I would also like the limit to changed to at least 20 WU per core. My Quads need more work as they just eat up 20 WU per machine really fast and then BOINC jumps back over to my backup project. This means that MilkyWay is not getting my full attention ____________
ID: 1541 \| Rating: 0 \| rate: / Reply Quote

Cruncher Pete Send message Joined: 30 Sep 07 Posts: 12 Credit: 133,909,841 RAC: 225,358	Message 1542 - Posted: 20 Jan 2008 \| 2:41:55 UTC - in response to Message 1315.
I agree entirely with Webmaster Yoda and Meshmar. This is emphasized more with our time difference around the world. In my case in Australia if a WU is stuck the machine will continue to work the other 16units and will stop as no more work units are available. As it is in the middle of our night I can not physically check and manually adjust thereby loosing many hours of computer time. If the WU's were increased at least our machines could crunch on until all units completed and the server is back functioning normal... ____________
ID: 1542 \| Rating: 0 \| rate: / Reply Quote

CairnsByTheSea Send message Joined: 28 Dec 07 Posts: 1 Credit: 2,521,156 RAC: 0	Message 1543 - Posted: 20 Jan 2008 \| 3:06:16 UTC
On a dual quad (which I don't have), 20 work units might only take 7 or 8 minutes... So, in other words for this scenario, for the default setting for the BOINC client (I think) of 60 minutes for switching on applications Milkyway could lose 52 minutes of CPU time! Hmmmm . . .
ID: 1543 \| Rating: 0 \| rate: / Reply Quote

Webmaster Yoda Send message Joined: 21 Dec 07 Posts: 69 Credit: 7,048,412 RAC: 0	Message 3510 - Posted: 24 May 2008 \| 13:12:24 UTC
Any further news on progress in this regard? Please?
ID: 3510 \| Rating: 0 \| rate: / Reply Quote

Author

Message

meshmar
Send message
Joined: 29 Aug 07
Posts: 7
Credit: 187,002
RAC: 0

My 400MHz K6-2 gets 20 WU
My 3.2GHz Athlon 64 X2 gets 20 WU

It takes an average of 66000 seconds of time for the K6-2 to complete 20 short WU and about 200000 for long WU.
It takes the Athlon 1750 for short WU and 4500 for long WU.

The reasoning behind the per host limit is time of return of the WU. This is an extreme example, but it shows that issuing 20 WU per host would not affect timely return of crunched WUs ...
____________

ID: 1315 | Rating: 0 | rate:

Reply Quote

Webmaster Yoda
Avatar

Send message
Joined: 21 Dec 07
Posts: 69
Credit: 7,048,412
RAC: 0

20 per core/cpu would be good. Some computers (4 cores or more) can go through 20 (short) work units in less than 15 minutes so they have to contact the server very often.

If it was 20 per core (provided it could be done), they could still get the work done in about an hour but with fewer connections to the server.

It would make no difference to slower, single core machines but could prove more efficient on multi core machines.
____________
Join the #1 Aussie Alliance on MilkyWay!

ID: 1319 | Rating: 0 | rate:

Reply Quote

LiborA

Send message
Joined: 15 Sep 07
Posts: 15
Credit: 5,312,625
RAC: 2,427

Yes, 20 WUs per core will be better :-)

ID: 1511 | Rating: 0 | rate:

Reply Quote

Webmaster Yoda
Avatar

Send message
Joined: 21 Dec 07
Posts: 69
Credit: 7,048,412
RAC: 0

Alternatively, if it has to stay at 20 per host, perhaps the backoff time could be reduced.

Milkyway@home 20/01/2008 11:09:18 AM Message from server: No work sent Milkyway@home 20/01/2008 11:09:18 AM Message from server: (reached per-host limit of 20 tasks) Milkyway@home 20/01/2008 11:09:18 AM Deferring communication for 20 min 0 sec Milkyway@home 20/01/2008 11:09:18 AM Reason: requested by project

The problem with this setting is that, when only short (2 credit) work units are sent, this computer (and others) goes through them in less than 20 minutes (typically 13-14 minutes on this particular computer). It runs out of work before it is allowed to connect again, so either goes idle or switches to another project (and may not switch back for an hour).

On a dual quad (which I don't have), 20 work units might only take 7 or 8 minutes...
____________
Join the #1 Aussie Alliance on MilkyWay!

ID: 1539 | Rating: 0 | rate:

Reply Quote

teemac
Send message
Joined: 3 Jan 08
Posts: 5
Credit: 1,377,460
RAC: 0

I would also like the limit to changed to at least 20 WU per core.

My Quads need more work as they just eat up 20 WU per machine really fast and then BOINC jumps back over to my backup project.

This means that MilkyWay is not getting my full attention
____________

ID: 1541 | Rating: 0 | rate:

Reply Quote

Cruncher Pete
Send message
Joined: 30 Sep 07
Posts: 12
Credit: 133,909,841
RAC: 225,358

I agree entirely with Webmaster Yoda and Meshmar. This is emphasized more with our time difference around the world. In my case in Australia if a WU is stuck the machine will continue to work the other 16units and will stop as no more work units are available. As it is in the middle of our night I can not physically check and manually adjust thereby loosing many hours of computer time.
If the WU's were increased at least our machines could crunch on until all units completed and the server is back functioning normal...
____________

ID: 1542 | Rating: 0 | rate:

Reply Quote

CairnsByTheSea
Send message
Joined: 28 Dec 07
Posts: 1
Credit: 2,521,156
RAC: 0

On a dual quad (which I don't have), 20 work units might only take 7 or 8 minutes...

So, in other words for this scenario, for the default setting for the BOINC client (I think) of 60 minutes for switching on applications Milkyway could lose 52 minutes of CPU time!

Hmmmm . . .

ID: 1543 | Rating: 0 | rate:

Reply Quote

Webmaster Yoda
Avatar

Send message
Joined: 21 Dec 07
Posts: 69
Credit: 7,048,412
RAC: 0

Any further news on progress in this regard? Please?

ID: 3510 | Rating: 0 | rate:

Reply Quote