Register FAQ Search Today's Posts Mark Forums Read

 2019-05-24, 14:11 #1728 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 420710 Posts Currently, https://www.mersenne.org/workload/ does not indicate any expiration dates for TF or P-1 assignents. Could you please add those? (Blank fields create the illusion these won't expire.) Last fiddled with by kriesel on 2019-05-24 at 14:14
2019-05-25, 21:17   #1729
Serpentine Vermin Jar

Jul 2014

1100110010002 Posts

Quote:
 Originally Posted by James Heinrich According to the data I'm seeing, DC's should be around ~55M right now, but the function that fetches available assignments isn't finding anything below 94M, which isn't right. I'll need George or Aaron to look into that.
I've been diving into the manual GPU assignment page. In my testing, I also found that if you just use the defaults and don't specify anything, it will take a VERY long time to query, thanks to a horrible aggregate function in the underlying query.

I mocked up a workaround for that and should have that in place soon. And I think just like whatever change James made, I also bumped up the defaults to go 1-million past the "high water mark".

What's this high water mark? Well, every day, the server updates some stats that basically have the exponent at the leading edge - basically, where the category 4 exponents start.

e.g. it's currently 52140822 for DC and 91389612 for first-time.

There is also a 90-day high water mark that is a projection of where that cat 4 work will be, 90 days from now. Those are currently 54143477 for DC and 92722631 for LL.

The default manual GPU assignments will only look for DC or LL work up to those thresholds that need additional TF.

And what, you may ask, constitutes an exponent that needs additional TF? Well, there's a lookup table involved that specifies, for different GPUs, what bit level is appropriate for different ranges of exponent size, although currently it seems to always use the GTX670 thresholds.

Those thresholds are:
Code:
bit	first_exponent
71	25000000
72	33000000
73	41000000
74	50000000
75	65000000
76	84000000
81	500000000
88	999999999
The way to read that would be, everything from 25M to 33M should be TFd to 71 bits, and between 65M-84M it should be 75 bits. There's a big gap there from 84M to 500M where we're only specifying 76-bits as appropriate. It's probably time to update this table with more interim levels.

The reason nothing shows up below about 93M is that everything below that has already been TF'd to those levels (or the exponent is assigned for LL/PRP testing and isn't available for TF assignments).

In my test code, bumping the top end so it can get exponents up to 93.7M would make another 8000'ish exponents available, most of them needing to go from 74 to 76 bits.

2019-05-25, 22:23   #1730
Serpentine Vermin Jar

Jul 2014

327210 Posts

Quote:
 Originally Posted by Madpoo Those thresholds are: Code: bit first_exponent 71 25000000 72 33000000 73 41000000 74 50000000 75 65000000 76 84000000 81 500000000 88 999999999
EDIT: I forgot this part... if there is already one LL/PRP test done, then the "optimal" TF level is increased by 1 bit. The theory behind that, I'm sure, is that finding a factor at this point only saves one test, not two, so they don't need to be factored as far.

Last fiddled with by Madpoo on 2019-05-25 at 22:23

 2019-05-25, 22:49 #1731 ixfd64 Bemusing Prompter     "Danny" Dec 2002 California 29×79 Posts The manual GPU assignment page isn't working. It takes a long time to process my request before showing a message saying no assignments are available. Last fiddled with by ixfd64 on 2019-05-25 at 22:50
2019-05-25, 23:12   #1732
Serpentine Vermin Jar

Jul 2014

327210 Posts

Quote:
 Originally Posted by ixfd64 The manual GPU assignment page isn't working. It takes a long time to process my request before showing a message saying no assignments are available.
Yup, that's what I just said, LOL:
Quote:
 I also found that if you just use the defaults and don't specify anything, it will take a VERY long time to query, thanks to a horrible aggregate function in the underlying query.
I'm about to implement the fix. Just did a last round of testing.

EDIT: It's updated. Try again and let me know how it went. It tests fine for me but having someone else confirm would help my comfort level.

Last fiddled with by Madpoo on 2019-05-25 at 23:42

2019-05-26, 15:42   #1733
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

101578 Posts

Quote:
 Originally Posted by Madpoo EDIT: I forgot this part... if there is already one LL/PRP test done, then the "optimal" TF level is increased by 1 bit. The theory behind that, I'm sure, is that finding a factor at this point only saves one test, not two, so they don't need to be factored as far.
Hmm, did you mean to write decreased by 1 bit? If saving two primality tests justifies factoring to bit level n, saving one primality test would justify half that factoring effort, which is bit level n-1.

2019-05-27, 04:23   #1734
Serpentine Vermin Jar

Jul 2014

63108 Posts

Quote:
 Originally Posted by kriesel Hmm, did you mean to write decreased by 1 bit? If saving two primality tests justifies factoring to bit level n, saving one primality test would justify half that factoring effort, which is bit level n-1.
I'm sure that's what I meant... my brain was looking at the code on the site and there's a plus one happening for double-checks, but of course that depends on which side of the equation it's plus one'ing.

 2019-05-27, 20:35 #1735 masser     Jul 2003 wear a mask 19·73 Posts Manual Reservation Specification Request? Would it be possible to add an additional option to the manual reservation page? I would like to specify the exponent range for P-1 reservations AND only reserve exponents that have already been factored to 76 bits. I would like to reserve about 20 exponents at one time, without reserving each exponent one-by-one.
2019-05-29, 04:14   #1736
Serpentine Vermin Jar

Jul 2014

23·409 Posts

Quote:
 Originally Posted by masser Would it be possible to add an additional option to the manual reservation page? I would like to specify the exponent range for P-1 reservations AND only reserve exponents that have already been factored to 76 bits. I would like to reserve about 20 exponents at one time, without reserving each exponent one-by-one.
For some reason I thought the server only hands out P-1 work if it has already been TF'd to some minimum level for that size exponent.

I could very easily be wrong about that though.

2019-05-30, 04:22   #1737
masser

Jul 2003

19·73 Posts

Quote:
 Originally Posted by Madpoo For some reason I thought the server only hands out P-1 work if it has already been TF'd to some minimum level for that size exponent. I could very easily be wrong about that though.
I reserved about 20 exponents in the 97M range. A few were trial-factored to 76 bits, but most were only TFed to 73 bits. I was hoping the server would only hand out the exponents at highest bit levels for P-1, but I guess not.

Last fiddled with by masser on 2019-05-30 at 04:22

2019-05-30, 14:52   #1738
Serpentine Vermin Jar

Jul 2014

23×409 Posts

Quote:
 Originally Posted by masser I reserved about 20 exponents in the 97M range. A few were trial-factored to 76 bits, but most were only TFed to 73 bits. I was hoping the server would only hand out the exponents at highest bit levels for P-1, but I guess not.
Hmmm.... maybe it's due to the difference in what is considered "optimally" TF'd between CPUs and GPUs.

 Similar Threads Thread Thread Starter Forum Replies Last Post ewmayer Lounge 39 2015-05-19 01:08 ewmayer Science & Technology 41 2014-04-16 11:54 cheesehead Soap Box 56 2013-06-29 01:42 cheesehead Soap Box 61 2013-06-11 04:30 Dubslow Programming 19 2012-05-31 17:49

All times are UTC. The time now is 01:12.

Fri Aug 7 01:12:09 UTC 2020 up 20 days, 20:58, 1 user, load averages: 1.35, 1.42, 1.40