Seti@Home optimized science apps and information
Optimized Seti@Home apps => Discussion Forum => Topic started by: Ghost0210 on 23 May 2010, 05:06:19 am
-
Hi, Whilst Seti has been having its recent upload/download issues I've been crunching on several different projects, one of which was Aqua. With the CPU wu's at Aqua the tasks will use mulitple cores, so one task willl use all 6 cores on my AMD, dramatically reducing the run time of a task.
What I was thinking is..... would this be possible on AstroPulse tasks at Seti? And if so, how easy would it be to implement?
For example, an AP task on my cpu takes around 12-14 hours to complete, but if I could run a task on say 3 of my cores, this in theory could cut the overall run time to 4-5 hours.
Would welcome any comments if this is possible to be done
Thanks
-
It requires some (maybe even big) efforts to make such multicore app effective.
Cause there is no hurry with returning AP tasks (instead of Aqua project where they wanna result back ASAP) it's no sense to worse performance in sake of fast turnaround.
-
Thanks Raistmer, is it just the development time needed to get the mulitcore app working effectively that would stop this from happening?
I understand that Seti ain't in any hurry to get the AP tasks back, but as I only have cuda cards, and it doesn't seem that AP are going to work on them, I was trying to think of ways that I could speed up the processing of them on my CPU
-
Thanks Raistmer, is it just the development time needed to get the mulitcore app working effectively that would stop this from happening?
I understand that Seti ain't in any hurry to get the AP tasks back, but as I only have cuda cards, and it doesn't seem that AP are going to work on them, I was trying to think of ways that I could speed up the processing of them on my CPU
Actually, Raistmer's OpenCL Astropulse port seems to work quite effectively on GPU, and when his build is finalised/released a nVidia build will likely quickly follow with the help of Devaster. It's true that the low distribution of Astropulse work has slowed the impetus towards Astropulse development.
With respect to Multibeam, the shrinking computational density ( really short processing time on new hardware ) indicates that a new model is needed, possibly processing multiple tasks at once.
When talking about threading & parallelism, it is all about a concept called 'granularity' which is the size of the parts that get processed in parallel. For the current situation, you need to realise that Boinc already effectively processes tasks in parallel, and that improvement on that can only come from implementation oriented toward greater data locality / lower contention. This sort-of study is ongoing, though much slower than CPU manufacturer marketing departments would like us to have ;)
Jason
-
Hi Jason, thanks for the info. Its good news that there could be a cuda AP app coming out at some stage :).
Even better though is that with the current and new hardware it maybe possible to run multiiple mb tasks on the same core
-
Hi Jason, thanks for the info. Its good news that there could be a cuda AP app coming out at some stage :).
Even better though is that with the current and new hardware it maybe possible to run multiiple mb tasks on the same core
I get excited when I think of that indeed, but DO know that there is much work ahead to achieve that. Have fun, these are some good times :D
Jason