[00:00.000 --> 00:26.240]  Okay, so the aim is to control the source of variation and from a scientific point of
[00:26.240 --> 00:32.320]  view when you are publishing a paper, an independent observer should observe the same
[00:32.320 --> 00:40.520]  results and have the same conclusion and this observation must be a sustainable so it
[00:40.520 --> 00:45.920]  doesn't depend where, when you observe it and maybe not where neither. So it's also
[00:45.920 --> 00:54.120]  collective. So all the question in this scientific framework is how can we redo
[00:54.120 --> 01:01.880]  later and elsewhere so I'm doing something on my laptop and another person will try to
[01:01.880 --> 01:10.280]  redo something like two years later the same thing. So how can we redo later and elsewhere
[01:10.280 --> 01:17.720]  what I've done here and today and this is a big question and the challenge in reproducible
[01:17.720 --> 01:25.520]  research and in science and I think Geeks answer to this question. So what does it mean
[01:25.520 --> 01:31.880]  a computational environment? So for example Alice says using this data you need that
[01:31.880 --> 01:40.840]  C file and GCC 11 to run my analysis. Okay, but what is the source code of GCC 11 and
[01:40.840 --> 01:46.200]  GCC 11 requires some tools for building and these tools, there is also tools require
[01:46.200 --> 01:54.960]  a runtime and this is recursive. So answering all this question is controlling the source
[01:54.960 --> 01:59.680]  of variations so you are controlling your computational environment. So this question
[01:59.680 --> 02:05.600]  is not new in computing, I mean in computing, it's not new. We have solutions. The solutions
[02:05.600 --> 02:13.600]  are package manager and so on. So we have package manager like APTU but there are some
[02:13.600 --> 02:17.880]  issues for example it's difficult to have several versions, difficult to have rollback.
[02:17.880 --> 02:23.440]  We have environment manager like Conda, PIP, module files but there is a kind of issue
[02:23.440 --> 02:30.880]  with transparency, who know what is inside a PIP installed torch. There is module files
[02:30.880 --> 02:37.080]  but how they maintain on your, I mean how they maintain, do you use on your laptop and
[02:37.080 --> 02:44.000]  on another machine? There is a docker file but the docker and docker files are based
[02:44.000 --> 02:55.280]  on previous solutions so also drawback apply also. Geeks in fact is all this solution plugged
[02:55.280 --> 03:03.440]  together and in fact it fix all the annoyance of each. This is what Geeks is. So Geeks is
[03:03.440 --> 03:11.320]  a package manager like APTU, etc. is a transactional and decorative and it produce showable packs
[03:11.320 --> 03:17.280]  like docker images. You can produce virtual machines and you can deploy that. For example
[03:17.280 --> 03:24.240]  you like unseemble or parker and you can build a whole Linux distribution and it's also a
[03:24.320 --> 03:32.520]  scheme library. Geeks is really awesome. Okay we have 20 minutes so I don't speak about
[03:32.520 --> 03:39.960]  that because it's too much. I'll just explain you how Geeks is helping me in my daily job.
[03:39.960 --> 03:47.760]  Geeks you can run Geeks on the top of any Linux distribution so it's really easy to try. If
[03:47.840 --> 03:57.840]  you haven't you should. So Geeks is just another package manager. So yeah you can install, remove
[03:57.840 --> 04:04.080]  without any privilege. This is more than APT from Debian for example. You have the direct
[04:04.080 --> 04:09.640]  creative management so declarative management means that you can have a configuration file
[04:09.640 --> 04:14.440]  and so on transactional so you don't have broken state you can rollback and so on and
[04:14.480 --> 04:19.080]  you have binary substitutes so you are not compiling everything from scratch every time.
[04:19.080 --> 04:25.240]  Okay this is some kind of classical package manager but you have more. You have isolated
[04:25.240 --> 04:33.000]  environment on the fly so you can create an isolated environment with Linux namespace on
[04:33.000 --> 04:39.440]  the fly and this is really helpful to check the dependency of your scientific analysis and you
[04:39.440 --> 04:47.160]  also can use Geeks to produce images like Docker images but without using the Docker file machinery.
[04:47.160 --> 04:57.320]  So you are saying okay nice but the issue with science is about reproducibility so you have a
[04:57.320 --> 05:03.120]  package manager that have all these features but why is reproducible? So the answer is about
[05:03.200 --> 05:11.240]  version. So for example Alice says I use GCC Adversion 11 so you have GCC to change but you
[05:11.240 --> 05:18.240]  need a linker like LD you need binitils and GCC is a compiler but the compiler need NPC and also
[05:18.240 --> 05:29.480]  need the NPC need NPFR and the big question is is the same GCC if we replace this NPF 4.1 by
[05:29.520 --> 05:38.800]  NPF 4 NPFR 4.0 is the same GCC or not and this is the issue that we can have when we are
[05:38.800 --> 05:46.000]  using when we are running analysis we are not controlling within us detail this graph and
[05:46.000 --> 05:52.200]  maybe the difference in the version of this package have a big influence of GCC at the end we
[05:52.440 --> 06:03.760]  cannot know and okay so Geeks in fact the version of Geeks it's fixed by I mean the state of Geeks is
[06:03.760 --> 06:10.840]  provided by Geeks described and this fix the graph this graph it fix the complete collection of
[06:10.880 --> 06:23.000]  packages and Geeks itself so in fact each node specify a receipt and each node specify code
[06:23.000 --> 06:31.080]  source the upstream source but also all the tools that you need to build the package so the compiler
[06:31.080 --> 06:39.080]  the build automation CMake make etc the configuration flags and so on and the dependency that you need
[06:39.320 --> 06:44.480]  to do that so you have a kind of recursive graph and this graph can be really really used for
[06:44.480 --> 06:52.760]  example for skypie it's more than 1000 nodes so it's it's not manageable by end and Geeks provide
[06:52.760 --> 07:04.720]  you a fine control about this graph so collaboration in action is this is what Geeks is helping me
[07:04.960 --> 07:12.160]  concretely every day so I write a manifest manifest it's just a file where I describe my tools for
[07:12.160 --> 07:20.960]  example python skype skipy nimpy or or and so on and I create my environment with Geeks shell so
[07:20.960 --> 07:30.600]  this is nice then I can I can pin this graph and I apply Geeks describe and I pin the graph so now
[07:30.600 --> 07:38.880]  I have another file state-alice which pins this graph but collaboration is about sharing the
[07:38.880 --> 07:45.920]  computational environment so somewhere is sharing one specific graph so in fact if I share these
[07:45.920 --> 07:53.600]  two files the manifest which describe the list of the tools and state-alice which describe the
[07:53.600 --> 08:06.360]  state of the graph okay Blake my collaborator can spawn exactly the same computational environment
[08:06.360 --> 08:14.320]  using the Geeks time machine so you think that Blake and Alice are running exactly the same
[08:14.360 --> 08:23.120]  computational environment and if Carol also knows these two files Carol can run the exact
[08:23.120 --> 08:30.640]  same environment as Blake and Alice so here we have some things that it's really easy so on
[08:30.640 --> 08:39.960]  my laptop I write my I specify the tools that I need python or etc and I specify the state I'm
[08:40.000 --> 08:45.960]  running on my on the laptop then I deploy on the cluster I use just transfer the two files and I run
[08:45.960 --> 08:52.040]  this Geeks time machine command and on the cluster I have the exact same environment that I have on
[08:52.040 --> 08:56.640]  my laptop so there is no question about what can be wrong between my laptop and the cluster because
[08:56.640 --> 09:03.400]  it's exactly the same computational environment and this is a game changer when you are running
[09:03.400 --> 09:07.960]  analysis on different machines where your colleagues are running in different places
[09:07.960 --> 09:19.720]  so Geeks this time machine provides a way to jump in different states temporarily so you can be I
[09:19.720 --> 09:27.360]  mean if you imagine you have the time and and and Alice and Blake or Carol are not on the same state
[09:28.200 --> 09:35.480]  compared to the time but they can jump artificially and temporarily to the same point in time to have
[09:35.480 --> 09:45.520]  the exact same computational environment and this is not possible with any other maybe next and so
[09:45.520 --> 09:52.040]  this is kind of game-changing for me so to have that working very well you need to preserve all
[09:52.040 --> 09:57.920]  the source code and for that you need software heritage which is I took just after and you need
[09:57.920 --> 10:06.160]  to have a backward compatibility of the Linux Linux kernel so and you also need to have some
[10:06.160 --> 10:13.280]  compatibility of hardware for example in five year we can if the hardware are not x86 yeah maybe
[10:13.280 --> 10:17.680]  we cannot jump it back in time but if you have compatibility of that where we have this that
[10:17.720 --> 10:25.840]  work and the question is what we have this size where these three conditions are satisfied so now
[10:25.840 --> 10:32.160]  we have for example this condition and what is the size of this window and from my point of view
[10:32.160 --> 10:44.280]  Geeks is running a quasi unique experiment at I mean real-world experiment a large since 2019 so
[10:44.320 --> 10:50.280]  we will see that this size maybe in five years we will try to to redo something from now and we
[10:50.280 --> 10:56.520]  will fail because something that we but we have this mechanism able to jump in different point in
[10:56.520 --> 11:06.280]  time so software heritage is a long-term source code archive so you collect and preserve all the
[11:06.280 --> 11:13.640]  source code of the world I mean open source code of the world so all github githlab debian and so on
[11:14.360 --> 11:20.080]  and geeks is able to to save the code of the geek package definition so for example the source
[11:20.080 --> 11:28.800]  code of DCC you can geeks use the source code of DCC coming from internet but you can save this
[11:28.800 --> 11:34.480]  directly to to software heritage and the package definition itself and what is really nice is
[11:34.480 --> 11:40.360]  that geeks is able to to fall back if the source disappears so for example you have a product
[11:40.920 --> 11:49.400]  if tomorrow github is down for whatever reason all the paper published with with the line my script
[11:49.400 --> 11:59.160]  is on github and the package is on on github and github is down like githulite I don't remember
[11:59.160 --> 12:07.400]  the name there is many popular platforms that are down and all all break and with this mechanism
[12:08.200 --> 12:26.040]  it is doesn't break so I have five minutes right okay so just geeks is able to so geeks
[12:26.040 --> 12:33.720]  is able to pack everything so you can produce a docker image with geeks so using the manifest
[12:33.800 --> 12:39.880]  you use geeks pack and geeks pack generate the docker image and then if Blake doesn't run geeks
[12:41.080 --> 12:47.080]  she can run the docker and the binary inside the docker are exactly the same than the binary
[12:47.080 --> 12:52.680]  inside the computational environment of Alice and because of the time machine this is reproducible
[12:52.680 --> 13:01.800]  over the time and this is also a kind of game changing in science and in fact a container it's
[13:01.800 --> 13:08.840]  just a format of the archive and geeks is an austic about this container format so you can
[13:08.840 --> 13:15.400]  generate torbol docker singularity there is an experimental debian binary package and yesterday
[13:15.400 --> 13:23.000]  evening there is a patch about supporting rpm package so this is just flexible to every
[13:23.720 --> 13:30.040]  context the key point is to to fully control the binary going inside the container and geeks
[13:30.040 --> 13:33.480]  does this job so this way it's a factory for creating images
[13:35.720 --> 13:41.640]  so geeks is helping me because there is three commands and two files so this is really easy
[13:41.640 --> 13:49.160]  to explain to for example medical doctor and and so on and it's a packing factory when I can deploy
[13:49.160 --> 13:53.800]  on infrastructure where geeks is not running for sharing computational environment and for me this
[13:53.800 --> 14:04.040]  is a two key for two keys for for for open science research and so on and if you want more
[14:05.880 --> 14:12.120]  information there is a this group geeks hpc but it's like many things in geeks the name is not
[14:12.120 --> 14:17.480]  good because it's more like geeks for science than not specific to geeks hpcs geeks for scientific
[14:17.800 --> 14:27.240]  and okay and it's running production so don't be afraid to install geeks I mean it's a there is
[14:27.240 --> 14:33.880]  all this cluster running already geeks more all the laptops and desktop so it's now it's in production
[14:35.720 --> 14:44.760]  so yeah this is uh I mean why the picture geeks and and and science that I would like to have in
[14:44.840 --> 14:47.560]  in the future so thanks
[15:01.320 --> 15:06.200]  so uh yeah but yeah yeah if you're the first one
[15:15.000 --> 15:21.960]  scientists but for example the scientists uh we are trying to reuse uh they can use for example
[15:21.960 --> 15:31.160]  different uh kind different cpu or different architecture how much some kind of uh optimizations
[15:31.160 --> 15:37.880]  for particular cpu can have a impact on having different results can it just like
[15:38.680 --> 15:45.640]  ruin a whole idea they can can they really like if you have idea like how much impact
[15:47.160 --> 15:58.120]  so I have to repeat the question so the question is the so in hpc context you have
[16:01.080 --> 16:05.160]  you have performance that depends on the architecture and you can have micro optimization
[16:05.160 --> 16:10.680]  for specific architecture so how geeks deal with that for the first question and how will
[16:10.680 --> 16:15.000]  we do that for reproducibility is the second question I think from my understanding is the two
[16:15.000 --> 16:27.800]  question I don't know the difference of the performances about the the micro optimization
[16:27.800 --> 16:34.360]  this is a job of the researcher in the field to say this micro optimization provides this
[16:34.360 --> 16:41.480]  performance improvement what I can say what geeks does to manage this micro optimization so
[16:42.120 --> 16:50.360]  Ludo is is is giving a talk in hpc dev room about the the the tune package transformation that
[16:50.360 --> 17:06.680]  I can ask when we are speaking about reproducibility we can ask if this micro
[17:06.680 --> 17:13.160]  micro optimization or I mean our fit the reproducible and the scientific method
[17:13.160 --> 17:16.440]  so I don't have the answer but at some at some point I think
[17:17.320 --> 17:23.720]  is it worth to have a micro optimization for I mean having something like a couple of percent
[17:23.720 --> 17:29.960]  of improvement but we lose all the reproducibility we lose the way to check that the computation is
[17:29.960 --> 17:34.600]  correct so I don't know if it's I mean this is a question for a collective question for the
[17:34.600 --> 17:38.760]  researcher in general so there is a question
[17:57.800 --> 18:03.160]  so the question is is it enough or not to have all this machinery to have reproducible science
[18:03.880 --> 18:11.560]  so the answer is I don't know because I mean if to have the answer is that everybody should
[18:11.560 --> 18:20.200]  run this to be be sure that he's in case or not so I don't know and the I don't remember what I want
[18:20.200 --> 18:21.480]  to say about that
[18:22.040 --> 18:40.600]  I mean I don't show here but there is a paper published with this method and yeah there is more
[18:40.600 --> 18:49.320]  reproducibility than than the other but at some point the the the the software is just one part
[18:49.320 --> 18:55.800]  of the big picture of the reproducibility issue in in in science and in fact is a is a is a is a
[18:55.800 --> 19:04.360]  collective practice in fact because you for example I'm trying to reproduce the paper from
[19:04.360 --> 19:11.480]  August this August and some data are missing the script are missing the the package that's
[19:11.480 --> 19:18.840]  been used are missing so so I cannot reproduce and this is not geeks it's just because
[19:19.320 --> 19:23.320]  publisher didn't good job thanks everybody