CADO NFS

[bur]

An update about my problem with the parameter files. In the README I found this section (fortunately before I submitted a bug):

The difficulty here is that when `cado-nfs.py` uses a parameter file
  supplied on the command line, it does not automatically insert into the
  parameter set the options that are necessary for running jobs.
  Therefore, we need to add these options:

./cado-nfs.py --parameters ./parameters/factor/params.c60 90377629292003121684002147101760858109247336549001090677693
 slaves.hostnames=localhost slaves.nrclients=2

Actually it's sufficient to supply slaves.hostnames=localhost, but then it will only launch 1 client. Which value is recommended for a 10 physical cores i9-10900k? When I don't specify a parameters file and thus don't need slaves.nrclients it appears to automatically select 5 clients.

Do you guys not specify the parameters file or why did no one recognize this problem? I'm just asking out of curiosity, not as an accusation. ;)

[charybdis]

Originally Posted by bur

Actually it's sufficient to supply slaves.hostnames=localhost, but then it will only launch 1 client. Which value is recommended for a 10 physical cores i9-10900k? When I don't specify a parameters file and thus don't need slaves.nrclients it appears to automatically select 5 clients.

5 clients is fine. I find that if you assign more than 4 cores to a single client the lattice siever doesn't get up to 100% CPU usage, though this may vary by CPU type.

Do you guys not specify the parameters file or why did no one recognize this problem? I'm just asking out of curiosity, not as an accusation. ;)

My usual CADO invocation line looks like

./cado-nfs.py ../cadojobs/2_2750M/2_2750M.params

where the .params file specifies not only the parameters but also N, workdir, the addresses of clients to start on other machines via ssh, the paths to those machines' CADO installations, etc.

[EdH]

I only supply a params file for SNFS jobs. All GNFS jobs use the default ones, which have been modified locally, mostly using VBCurtis' versions.

As to clients, I run scripts on the clients that check thread count and use override t <count> to use all threads. On my 4c/8t machines they show up to 79x% use for the las process in top. On my 12c/24t machines, I do break them into three, 8 thread processes that each show 79x% in top, but those machines have more memory.

[bur]

Ok, thanks for the info. When I have some spare time I'll look into setting up scripts to do ECM, then cado and finally msieve for linear algebra. Could be interesting to setup.



In that regard, should the filtering also be done with msieve or is that fine with cado?

[charybdis]

Originally Posted by bur

In that regard, should the filtering also be done with msieve or is that fine with cado?

CADO filtering used to be compatible with msieve linear algebra, and indeed the documentation would suggest that it still is, but I recall reading that more recent changes to CADO have made the filtering format incompatible with msieve.

[bur]

I'll give it a try, right now cado is set to do the sieving and filtering with tasks.linalg.run=false. If it fails I can always redo the filtering with msieve, right?


workdir is set to a directory other than /tmp/.