Poly select and CADO sieving for 2,1165+ C217 - Page 6

frmky · 2020-02-17, 21:24

For 2,1165+, the 16e tasks will use significantly more memory. We can't afford to keep it artificially low for this one. It will likely be a bit over 2GB per core.

pinhodecarlos · 2020-02-17, 21:35

Quote:

Originally Posted by frmky

For 2,1165+, the 16e tasks will use significantly more memory. We can't afford to keep it artificially low for this one. It will likely be a bit over 2GB per core.

Make an announcement, double the points for this case only, etc ...I'm happy up to 4 GB/thread.

VBCurtis · 2020-02-18, 02:23

Quote:

Originally Posted by frmky

For 2,1165+, the 16e tasks will use significantly more memory. We can't afford to keep it artificially low for this one. It will likely be a bit over 2GB per core.

Greg-
Please post parameters here once you decide them- lim's and mfb's/lp's. I don't think your client is 34-bit capable, so I imagine we're on 33/33 for LP.
I'd like to start sieving very small Q locally on CADO. I'll use A=32, which is equivalent of 16.5e (40% larger sieve area). I have just one 20-core machine available at present, so I won't get far, but I can start at Q=10M and contribute some relations.

frmky · 2020-02-18, 21:39

Quote:

Originally Posted by VBCurtis

Greg-
Please post parameters here once you decide them- lim's and mfb's/lp's. I don't think your client is 34-bit capable, so I imagine we're on 33/33 for LP.
I'd like to start sieving very small Q locally on CADO. I'll use A=32, which is equivalent of 16.5e (40% larger sieve area). I have just one 20-core machine available at present, so I won't get far, but I can start at Q=10M and contribute some relations.

After testing I'll stick with the suggested parameters:
rlim: 536000000
alim: 536000000
lpbr: 33
lpba: 33
mfbr: 96
mfba: 66
rlambda: 3.7
alambda: 2.8

I usually start at 20M, but I can up that a bit if you wish.

VBCurtis · 2020-02-18, 22:28

Thanks!
How about you start at 40M, and I'll run Q=5-40M on CADO? I think I'll get yield 3-6x higher than ggnfs at those smaller Q (plus the extra 40% from using a larger siever), since CADO is fine with sieving Q values below lim.

I'll get CADO fired up later this week, and will post yield and sec/rel data once I have a reasonable sample.

henryzz · 2020-02-19, 23:29

Quote:

Originally Posted by VBCurtis

Thanks!
How about you start at 40M, and I'll run Q=5-40M on CADO? I think I'll get yield 3-6x higher than ggnfs at those smaller Q (plus the extra 40% from using a larger siever), since CADO is fine with sieving Q values below lim.

I'll get CADO fired up later this week, and will post yield and sec/rel data once I have a reasonable sample.

The lower you go the faster it gets in general. Of course, this has to be balanced with increased duplication. The non-standard options of adjust-strategy might help keep the duplication rate down(different q will get different logI and logJ combos). I would be interested in hearing the results of any test sieves. Is it possible to estimate duplication rates based on test sieving?

VBCurtis · 2020-02-20, 00:27

Not a good sample yet, but at Q=5M after Q=4000 interval I have yield around 6.0, roughly 3.5 sec/rel (with all hyperthreads in use on ECM or CADO, so roughly 2.0 cpusec/rel), just under 20GB per 10-threaded sieve process.

This is with CADO A=32, which is the equivalent of I=16.5. I'm running two 10-threaded processes until I reach Greg's starting Q. I'll update yield and sec/rel in a day or two after I've searched 50kQ or more.

If yield is cut in half by 40M compared to 5M, I would still get something like 140-160M relations from Q=5-40M; however, that would take me 7-8 months on 20 threads. I'll add at least one more client later, and if anyone with 20GB free memory is interested I can open CADO access in the exact-same manner that we did on the C207 team sieve from last summer.

Note that Greg may run out of Q to sieve with ggnfs; if that looks to be the case after test-sieving, he should move his starting value to Q=50M or 60M and we should ask for client-CADO-help.

Is 1200 million raw relations a reasonable target?

SethTro · 2020-02-20, 09:22

I have 20GB free and am happy to point 16 cores at it.

swellman · 2020-02-20, 11:57

I can contribute some cores as well.

pinhodecarlos · 2020-02-20, 12:21

Will only contribute through NFS@Home, might bring some friends to the mix CADO and NFS@Home.

VBCurtis · 2020-02-20, 18:27

Roughly one day on 20 (hyper)threads got me 60kQ and 362,000 relations. Yield, then, is 6.0 at Q=5M. I changed to 3 10-threaded instances today.

I copied the setup parameters for non-local access from the C207 job file; I hope it works!
Same info as y'all used for 2330L, except the port number is now 35029 (randomly assigned, possible it might change if CADO crashes and is reset).

Default number of sieve threads is 10, so most of you will need to set that with the override switch that we used on 2330L.

Using my 30 threads, I'll complete 1MQ in about 11 days. That means I'm good for 8-10MQ. Sounds like Greg won't queue this for a couple weeks, so we have a little time to decide how many Q we can get done with this mini-team-effort, and then tell Greg what Q to start at.

2020-02-20, 12:21	#65
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 1001101010011₂ Posts	Will only contribute through NFS@Home, might bring some friends to the mix CADO and NFS@Home. Last fiddled with by pinhodecarlos on 2020-02-20 at 12:23

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Poly select and planning for 2,2330M	VBCurtis	Cunningham Tables	70	2021-02-16 06:32
Poly select and planning for 2,2210M	swellman	Cunningham Tables	51	2020-03-22 22:09
Poly select and test-sieving for RSA232	VBCurtis	Operation Kibibit	25	2020-01-07 01:57
YAFU Poly Select Deadline	amphoria	YAFU	22	2016-09-17 09:47
Starting NFS skipping poly select	jux	YAFU	5	2016-01-02 01:01

2020-02-17, 21:24	#56
frmky Jul 2003 So Cal 4077₈ Posts	For 2,1165+, the 16e tasks will use significantly more memory. We can't afford to keep it artificially low for this one. It will likely be a bit over 2GB per core.

2020-02-18, 22:28	#60
VBCurtis "Curtis" Feb 2005 Riverside, CA 2⁸·19 Posts	Thanks! How about you start at 40M, and I'll run Q=5-40M on CADO? I think I'll get yield 3-6x higher than ggnfs at those smaller Q (plus the extra 40% from using a larger siever), since CADO is fine with sieving Q values below lim. I'll get CADO fired up later this week, and will post yield and sec/rel data once I have a reasonable sample.

2020-02-20, 00:27	#62
VBCurtis "Curtis" Feb 2005 Riverside, CA 2⁸×19 Posts	Not a good sample yet, but at Q=5M after Q=4000 interval I have yield around 6.0, roughly 3.5 sec/rel (with all hyperthreads in use on ECM or CADO, so roughly 2.0 cpusec/rel), just under 20GB per 10-threaded sieve process. This is with CADO A=32, which is the equivalent of I=16.5. I'm running two 10-threaded processes until I reach Greg's starting Q. I'll update yield and sec/rel in a day or two after I've searched 50kQ or more. If yield is cut in half by 40M compared to 5M, I would still get something like 140-160M relations from Q=5-40M; however, that would take me 7-8 months on 20 threads. I'll add at least one more client later, and if anyone with 20GB free memory is interested I can open CADO access in the exact-same manner that we did on the C207 team sieve from last summer. Note that Greg may run out of Q to sieve with ggnfs; if that looks to be the case after test-sieving, he should move his starting value to Q=50M or 60M and we should ask for client-CADO-help. Is 1200 million raw relations a reasonable target?

2020-02-20, 09:22	#63
SethTro "Seth" Apr 2019 293₁₀ Posts	I have 20GB free and am happy to point 16 cores at it.

2020-02-20, 11:57	#64
swellman Jun 2012 3,079 Posts	I can contribute some cores as well.

2020-02-20, 18:27	#66
VBCurtis "Curtis" Feb 2005 Riverside, CA 2⁸×19 Posts	Roughly one day on 20 (hyper)threads got me 60kQ and 362,000 relations. Yield, then, is 6.0 at Q=5M. I changed to 3 10-threaded instances today. I copied the setup parameters for non-local access from the C207 job file; I hope it works! Same info as y'all used for 2330L, except the port number is now 35029 (randomly assigned, possible it might change if CADO crashes and is reset). Default number of sieve threads is 10, so most of you will need to set that with the override switch that we used on 2330L. Using my 30 threads, I'll complete 1MQ in about 11 days. That means I'm good for 8-10MQ. Sounds like Greg won't queue this for a couple weeks, so we have a little time to decide how many Q we can get done with this mini-team-effort, and then tell Greg what Q to start at.