Downloading sufficient tasks for a fe2w days while I have no internet connection.
log in

Advanced search

Message boards : Macintosh : Downloading sufficient tasks for a fe2w days while I have no internet connection.

Author Message
Baz
Send message
Joined: 11 Aug 14
Posts: 2
Credit: 33,378,902
RAC: 0
Message 21573 - Posted: 21 Oct 2015, 11:56:39 UTC

My apologies if this has been covered before or this isn't the right place to ask (I never use forums) but here goes. I have a desktop Mac (Mac mini (Late 2012), 2.6 GHz Intel Core i7, 16 GB 1600 MHz DDR3, Intel HD Graphics 4000 1536 MB) that I connect to the internet by tethering to my mobile phone. I'm going away tomorrow for about 5 days & want to leave sufficient Collatz tasks for that time. I run Collatz exclusively on the GPU & the GPU exclusively runs Collatz (i.e. I'm not looking for CPU tasks). In computing preferences I have it set to give me enough tasks for 10 days but there only ever seem to be enough for about 2. Is there any way to "force" the server to send me more tasks? I don't really want the GPU sitting idle whilst I"m away as this will be (potentially) it's most productive time. Thanks in advance for your help.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 21619 - Posted: 27 Oct 2015, 13:59:43 UTC - in response to Message 21573.

If the WU size were to remain the same and I was to increase the number you were allowed to download in advance, the WUs in progress would increase by 2.5 times which means the server has to keep track to 2.5 times as many results and each time a computer connects, it needs to check 2.5 times as many WUs one at a time as it compares the list the server thinks your host has to the list your host thinks it has. With machines that have bad configurations (e.g. haven't yet installed the latest C++ runtimes) they can download an entire 5 day cache, have every single one error out within a few seconds, and 3 minutes later ask for another 5 days of work. BOINC will eventually (after about 20,000 WUs or so it seems) stop sending them work but in the mean time hundreds of thousands of WUs are trashed and need to get resent. In addition, the BOINC client slows down (and requires more and more CPU) the more WUs that are in progress. So, keeping the WUs the same size and increasing the number you can download results in both the server and the client having problems.

If I increase the size, then none of the CPUs can get work because it will take them soooooo long to complete the WU that the deadline would already be past.

If I make a new app (e.g. large_collatz_sieve) then the "set it and forget it" crowd will complain that they don't have to edit their preferences in other projects and get all pissed off because their CPU crunched for 5 days on a large WU and wont' finish in time but still want credit for the work done. It also means the feeder daemon on the server has to split the number of WUs between two apps. It further divides that by the plan classes for each app. The result is that the plan classes with just a few hosts (e.g. pre Android 4.1 ARM devices which run the non-pie WUs) will have an over abundance of WUs and the nvidia and AMD GPUs will keep seeing the "no work available" error in the BOINC event log.

I won't even go into how useless it is for BOINC to use FLOPS estimates to calculate integer application run times which result estimates that are up off by a factor of up to 100.

What I really need to do is come up with a way to determine how many numbers per hour a processor (GPU or CPU) can process and then create a custom WU for that processor on the fly. I'd also need a way to "match" the processors so that if it errors out on one machine, another which uses WUs of a similar size could finish it. e.g. download one and only one small WU to start and use that as a benchmark to determine how large future WUs should be. Then use the returned results going forward to determine the numbers crunched per hour going forward. That would allow it to adjust automatically to those who feel the need to run 200 BOINC projects at a time or only crunch for a few hours a day. Unlike the stock market, past performance would be used to judge future results. There would have to be a way for a machine to request a "reset" of the stats so that challenges or major changes to what is being crunched and for how long could be adjusted more easily.


Post to thread

Message boards : Macintosh : Downloading sufficient tasks for a fe2w days while I have no internet connection.


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.