IRC channel logs


back to list of logs

<civodul>--check is bound to fail, at least because of the reproducibility bug
<rekado>civodul: that was my conclusion when I did a reproducibility analysis of pigx-sarscov2-ww, yes.
<rekado>annoyingly, I can’t seem to find my emails about this
<zimoun>civodul, Yes for sure. But it fails with getarch_2nd.c:12:35: error: ?SGEMM_DEFAULT_UNROLL_M? undeclared
<zimoun>What annoys me is the weirdness. On machine A, «guix time-machine --commit=4b1538e6ef -- help» fails and it passes on machine B. Bah, I do not know.
*rekado runs “guix time-machine --commit=87e7faa2ae641d8302efc8b90f1e45f43f67f6da -- build openblas”
<rekado>still building
<rekado>build phase complete
<rekado>all done
<zimoun>rekado: x86_64, right?
<zimoun>on foreign distro or Guix System?
<rekado>on Guix System, x86_64
<rekado>I sent a reply with the output of lscpu
<rekado>I’ll try again on a different server, foreign distro, with Intel(R) Xeon(R) CPU E7-4870 v2 @ 2.30GHz
<rekado>(I have to move .config/guix/channels.scm out of the way)
<zimoun>for 7357b3d7a52eb5db1674012c50d308d792741c48, It works for Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz but not for Intel(R) Core(TM) i7-10700K CPU @ 3.80GHz; running both on Ubuntu (not the same version though)
<rekado>on that machine I just got a substitute
<rekado>will do “--check --no-grafts” to see if the build fails early
<rekado>it’s now building
<zimoun>Yes, I got the substitute too, and the failure is --check, --no-grafts.
<rekado>if this was like Konrad’s case it would have failed by now
<rekado>does Konrad’s failure happen on a CPU that is unknown by the old openblas because it is too recent?
<rekado>wow, the build is done
<zimoun>I do not know.
<rekado>of course “--check” failed, because the output differs
<zimoun>bah it is probably the CPU. For 87e7faa2ae641d8302efc8b90f1e45f43f67f6da, it passes on Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz but fails on Intel(R) Core(TM) i7-10700K CPU @ 3.80GHz.
<rekado>so… running the build in qemu… is that still reproducibility or is it silly?
<zimoun>I do not know.
<zimoun>civodul, rekado: I have recently read the Skylake bug about OCaml
<zimoun>is it possible for something similar for GCC?
<civodul>i pushed a couple of tweaks to the Guix-HPC report
<civodul>the term "open science" had to be present
<civodul>in the sense that it's good to relate our geeky work to the broader context
<civodul>(i hate this tendency to qualify anything as "open", it's meaningless, but hey)
<rekado_>I thought you meant it didn’t pass some external PR threshold
<civodul>heh no :-)
<zimoun>civodul: the report LGTM! Even, all the achievements are great! :-)
<civodul>zimoun: hi! good, thanks for checking!
<civodul>i'll put it online after lunch
<rekado_>thank you!
<zimoun>civodul: how do I know the codename of cpu for --tune?
<zimoun>For instance, lscpu provides me the nam ei7-10700K but then I am going to to know that the product is formerly Comet Lake.
<zimoun>Then does not list Comet Lake. Does it mean the CPU is not supported?
<civodul>zimoun: you can try the "guix shell cpuinfo -- cpu-info"
<civodul>or just run "guix build --tune gsl -n" for instance
<civodul>but yes, different compilers support different CPUs
<zimoun>Arf, I do not understand all this business… cpu-info tells me: Microarchitectures: 8x Sky Lake. But the official Intel doc calls it Comet Lake.
<zimoun>And “guix build --tune gsl -n” says «guix build: tuning gsl@2.7 for CPU broadwell».
<zimoun>Pic et pic et colegram ;-)
<civodul>is it a brand new Intel CPU?
<rekado_>pages 25, 26, and 27 are a breeze to read!
<rekado_>“We gave demonstrations of what Guix brings to scientific workflows are now available and we expect to continue to show that reproducible scientific workflows are indeed a possibility” — the “are now available” feels out of place
<civodul>ah damn it
<zimoun>civodul: I do not know if the CPU is new. The machine is recent, from December. It is 10th generation when Tiger Lake is 11th; and Tiger Lake is listed.
<civodul>rekado_, zimoun: could you make sure your institutes retweet or somehow relay the info?
<civodul>rekado_: oh yes, the blank pages are needed for the printed A5 booklet
<civodul>i recommend printing it, you'll see!
<civodul>and, as a bonus, it lets recipients of the printed copy take notes about their favorite parts
<rekado_>civodul: I’ll ask Altuna to tweet it; usually the comms department retweets then