Discussion related to cado-nfs

[Jason Parker-Burlingham <jasonp@panix.com>]

On Tue, Sep 22, 2020 at 11:11:46AM +0200, eric.jeancolas@free.fr wrote:

> I made some further analyses on a log.
>
> * when a full run is done without interruptions, the overall WTC
>   (elapsed time for entire factorization) is ok. But when I retart after
>   an interrupt, this counter starts from the restart. That's why I used
>   the Total cpu counter to have an idea of the cpu usage. In the log I
>   studied, Total Usage / Total WTC is 3,68. I've got a 4-core computer,
>   so it's not too weird. But it is not really exact. It's too much...
>
> * in the log, the following is reported : Info:Lattice Sieving: Total
>   time: 666693s. What is this time? From my log, the WTC is 334221,78,
>   close to half this value. When I looked at cpu usage with the top
>   command, I saw two las tasks with two threads. Well, two las processes
>   at 198% CPU.
>
> * when entries in the log are talking about WTC, the reported values
>   are ok.
>
> * I have an entry with Info:Linear Algebra: Total cpu/real time for
>   bwc: 90816.1/23293.6. It seems to be the sum of Krylov + Lingen +
>   Mksol. It is true for WTC. So if there is a sum about total cpu, there
>   should be total cpu value for each of Krylov and Mksol, shouldn't it?
>   If not, I don't understand how the total is computed. 
>
> * when I see Jason's log:
>   https://stdkmd.net/nrr/cont/7/77377.htm#N135_C183 , I see a ratio
>   about total statistics 2.37178e+08/2.53203e+06 = 93.67. It seems a
>   lot, even for a AMD Ryzen Threadripper 2990WX 32-Core. That's why I
>   made this log analysis about time values...

[Re-sending this message because it's a good idea to check the size of
logfile before attaching to email.  Sorry if this is a near-duplicate,
Eric.]

I've uploaded a copy of my logfile for that factorization to
https://storage.googleapis.com/number-field-sieve/run/77377_135.log .

Total runtime was about 3 months, with numerous interruptions during
sieving, krylov and one failure during sqrt.

Some interruptions were because I was experimenting with thread counts,
others because of power failures, and the sqrt interruption because I
had to drop threads from 8 to 4 to not run out of memory.

In particular one interruption during krylov was checking that
utlization and s/iter did not improve by running 64 threads instead of
32.  I also remmeber that total CPU utilization reported by htop during
krylov was fairly low, around 30 to 40%.

Whether this might be of particular help in understanding my CPU and
real time reports, I'm not sure.  I do still have all the files
generated by the factorization, so if parameter snapshot files and so
forth are useful, I can also make those available.

jason