Forum > Discussion Forum
Latest Seti news.......
msattler:
Just heard from Eric again.........
They are still trying to get jocelyn up to speed to handle mork's duties.......not sure when or how well she might perform once configured.
Richard Haselgrove:
--- Quote from: msattler on 19 Oct 2010, 06:27:42 pm ---Just heard from Eric again.........
They are still trying to get jocelyn up to speed to handle mork's duties.......not sure when or how well she might perform once configured.
--- End quote ---
They'll need to keep the query rate down - both the number of queries, and the size of the result recordsets.
Quota is a bad tool for that. If it's high, the database, and the recordsets, get huge. If it's low, the client work request rate gets huge - probably two or three work requests per completed task (not done an exact count). With a multi-CUDA host finishing a task every couple of minutes, that'll reduce Jocelyn to a heap of smoking ashes the first day.
My judgement would be to keep quota (to keep the size down), but to add a scheduler server-requested backoff of something realistic - say 10 minutes. That wouldn't be any hardship to multi-CUDA - report five, get five, instead of report one, get one - but it would slow down the request rate dramatically.
msattler:
--- Quote from: Richard Haselgrove on 19 Oct 2010, 06:52:36 pm ---
--- Quote from: msattler on 19 Oct 2010, 06:27:42 pm ---Just heard from Eric again.........
They are still trying to get jocelyn up to speed to handle mork's duties.......not sure when or how well she might perform once configured.
--- End quote ---
They'll need to keep the query rate down - both the number of queries, and the size of the result recordsets.
Quota is a bad tool for that. If it's high, the database, and the recordsets, get huge. If it's low, the client work request rate gets huge - probably two or three work requests per completed task (not done an exact count). With a multi-CUDA host finishing a task every couple of minutes, that'll reduce Jocelyn to a heap of smoking ashes the first day.
My judgement would be to keep quota (to keep the size down), but to add a scheduler server-requested backoff of something realistic - say 10 minutes. That wouldn't be any hardship to multi-CUDA - report five, get five, instead of report one, get one - but it would slow down the request rate dramatically.
--- End quote ---
Whatever it takes to keep things alive until the cavalry arrives........
You won't hear any complaining from me.
Have you passed your suggestions on to Eric?
Sounds like mork is nothing but a parts donor in the future, but Eric says they might try to put him back online temporarily as the replica DB.
Richard Haselgrove:
--- Quote from: msattler on 19 Oct 2010, 06:55:54 pm ---
--- Quote from: Richard Haselgrove on 19 Oct 2010, 06:52:36 pm ---
They'll need to keep the query rate down - both the number of queries, and the size of the result recordsets.
Quota is a bad tool for that. If it's high, the database, and the recordsets, get huge. If it's low, the client work request rate gets huge - probably two or three work requests per completed task (not done an exact count). With a multi-CUDA host finishing a task every couple of minutes, that'll reduce Jocelyn to a heap of smoking ashes the first day.
My judgement would be to keep quota (to keep the size down), but to add a scheduler server-requested backoff of something realistic - say 10 minutes. That wouldn't be any hardship to multi-CUDA - report five, get five, instead of report one, get one - but it would slow down the request rate dramatically.
--- End quote ---
Whatever it takes to keep things alive until the cavalry arrives........
You won't hear any complaining from me.
Have you passed your suggestions on to Eric?
--- End quote ---
No, I reckon he's got more then enough in his inbox to keep him busy, let alone get on with the real work. He doesn't need shrapnel from the peanut gallery.
If the community can come up with some consensus advice, based on real understanding of large-scale databases and operating system comms infrastructure, then it can go back to Eric in one of the comms channels which is already open. Yours seems to be working well at the moment.
Richard Haselgrove:
He got the message....
http://setiathome.berkeley.edu/forum_thread.php?id=61828&nowrap=true#1043291
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version