!kFJOpVCFYFzxqjpJxm:nixos.org

Nix HPC

72 Members
Nix for High Perfomance Computing clusters18 Servers

You have reached the beginning of time (for this room).


SenderMessageTime
25 Mar 2023
@rbutani:matrix.orgrbutani joined the room.23:55:19
27 Mar 2023
@ryantalo:matrix.orgRyan Swart joined the room.09:41:42
29 Mar 2023
@ss:someonex.netSomeoneSerge (UTC+U[-12,12]) joined the room.19:39:45
@ss:someonex.netSomeoneSerge (UTC+U[-12,12]) Hi! Just wanted to mention that torchWithRocm is going to require allowUnfree = true when https://github.com/NixOS/nixpkgs/pull/222273 gets merged. I think this needs to be addressed upstream in openai/triton (they unconditionally download and vendor a copy of ptxas from cuda), but I haven't the capacity to handle this now 19:54:53
31 Mar 2023
@ryantalo:matrix.orgRyan SwartKind of a meta question - I am trying to run static nix on an HPC cluster and am running into some roadblocks (with a custom path and disabling the sandbox as much as possible I am still running into permission denied errors - can't chroot or run bubblewrap); given that I have the ability to run arbitrary binaries - is there a way to disable all "special" path modification etc in such a way that common HPC access limitations can allow nix to still build derivations?02:28:46
@ryantalo:matrix.orgRyan SwartI have looked into both static-nix and nix-portable, but neither seem to be a drop-in solution for my case02:29:36
2 Apr 2023
@jboy:utwente.ioJohn ✒️ryantalo: have you tried something like this? https://www.jboy.space/blog/nix-on-hpc.html11:56:18
1 Apr 2023
@ss:someonex.netSomeoneSerge (UTC+U[-12,12]) Anyone tried running running Nix from Singularity and binding /nix/store? The cluster at my uni won't enable user namespaces, so no unchroot and no bubblewrap for me 23:44:29
@ss:someonex.netSomeoneSerge (UTC+U[-12,12]) singularity run --bind /abc:/efg something.sif works, but I want to use singularity-tools.buildImage and that puts e.g. the shell into /nix/store, which then is hidden by the --bind =' 23:45:35
@ss:someonex.netSomeoneSerge (UTC+U[-12,12]) And singularity run --overlay is, again, disabled on the cluster xD 23:45:55
@ss:someonex.netSomeoneSerge (UTC+U[-12,12])
In reply to @ryantalo:matrix.org
Kind of a meta question - I am trying to run static nix on an HPC cluster and am running into some roadblocks (with a custom path and disabling the sandbox as much as possible I am still running into permission denied errors - can't chroot or run bubblewrap); given that I have the ability to run arbitrary binaries - is there a way to disable all "special" path modification etc in such a way that common HPC access limitations can allow nix to still build derivations?

disable all "special" path modification etc

Hi, which path modifications are you referring to?

23:46:27
2 Apr 2023
@ryantalo:matrix.orgRyan SwartThat describes using chroot - which I don't have permissions to do; nix seems to get to a point of reading/writing source files, and takes forever to work through them (like 5 mins+, log example here: https://gist.github.com/ryanswrt/3ebf02bd10e25319f153014698a14c63) and then eventually dies13:40:48
3 Apr 2023
@ryantalo:matrix.orgRyan Swart It ends up dying with error: setting up a private mount namespace: Operation not permitted 00:50:04
@jb:vk3.wtf@jb:vk3.wtf
In reply to @ss:someonex.net
singularity run --bind /abc:/efg something.sif works, but I want to use singularity-tools.buildImage and that puts e.g. the shell into /nix/store, which then is hidden by the --bind ='
i use singularity on a hpc to bind in a nix store, but not in conjunction with a singularity image as since i have a nix store in the container i just use the nix in that store
01:19:37
@ss:someonex.netSomeoneSerge (UTC+U[-12,12])
In reply to @jb:vk3.wtf
i use singularity on a hpc to bind in a nix store, but not in conjunction with a singularity image as since i have a nix store in the container i just use the nix in that store
Is it a prepoppulated store, or a writable image?
11:15:52
@peddie:matrix.org@peddie:matrix.org joined the room.17:29:11
@jb:vk3.wtf@jb:vk3.wtfI prepopulate a nix store in project/scratch area with a copy of nix then bind mount it in the container so it's writable20:53:52
@jb:vk3.wtf@jb:vk3.wtfThe singularity image I use is completely empty, really just using singularity to bind mount as there's no namespaces20:54:42
4 Apr 2023
@tpw_rules:matrix.orgtpw_rulesi've done hax with singularity too. i need to document them sometime03:43:20
@ryantalo:matrix.orgRyan SwartHave you benched the performance profile of nix-in-singularity? How close is it to native?09:28:23
@jb:vk3.wtf@jb:vk3.wtfno i haven't but i expect it to be close to native as it's just a bunch of binds handled by the kernel11:39:46
@ss:someonex.netSomeoneSerge (UTC+U[-12,12])

@jbedo do you also disable the sandbox?

I suppose what I'm seeing is nix trying to set up mount namespaces for the builds, but idk:

Singularity> nix-shell -p gh
error: executing shell '/nix/store/k4j76vg58c6j90s8l46kx3843vysx732-bash-interactive-5.2-p15/bin/bash': No such file or directory
Singularity> nix shell nixpkgs#gh
error: setting up a private mount namespace: Operation not permitted
Singularity>
17:00:50
6 Apr 2023
@ryantalo:matrix.orgRyan Swart Ok, managed to get things building + running using the singularity method jbedo described; alas inode limits are killing me now. It seems that nix build nixpkgs#hello is essentially evaluating every nix package for some reason; why would /nix/store/tmp-1759374-1/x/pkgs/development/python-modules/qimage2ndarray be getting built for a simple c binary otherwise... 02:12:15
@jb:vk3.wtf@jb:vk3.wtf
In reply to @ss:someonex.net

@jbedo do you also disable the sandbox?

I suppose what I'm seeing is nix trying to set up mount namespaces for the builds, but idk:

Singularity> nix-shell -p gh
error: executing shell '/nix/store/k4j76vg58c6j90s8l46kx3843vysx732-bash-interactive-5.2-p15/bin/bash': No such file or directory
Singularity> nix shell nixpkgs#gh
error: setting up a private mount namespace: Operation not permitted
Singularity>
yeah i disable sandbox and a few other things that don't work on luster (WAL & luster.lov ACL)
02:47:49
@jb:vk3.wtf@jb:vk3.wtf
In reply to @ryantalo:matrix.org
Ok, managed to get things building + running using the singularity method jbedo described; alas inode limits are killing me now. It seems that nix build nixpkgs#hello is essentially evaluating every nix package for some reason; why would /nix/store/tmp-1759374-1/x/pkgs/development/python-modules/qimage2ndarray be getting built for a simple c binary otherwise...
it shouldn't be building anything you don't need, but it does need to unpack nixpkgs which might be your problem since there's quite a lot of files
02:49:23
7 Apr 2023
@aaronmondal:matrix.orgAaron Siddhartha Mondal joined the room.03:56:57
24 Apr 2023
@emilytrau:matrix.org@emilytrau:matrix.org changed their display name from emilytrau (she/her) to emilytrau.08:37:39
1 May 2023
@twesterhout:matrix.org@twesterhout:matrix.org joined the room.12:46:05
2 May 2023
@twesterhout:matrix.org@twesterhout:matrix.org Hello, has anybody managed to run MPI-aware applications via nix-portable or Nixie on an HPC cluster where Nix isn't installed? Single-node executables work fine, but I'm struggling to do anything with multi-node stuff. Currently trying to go through Singularity with pkgs.singularity-tools, but no luck so far. So if anybody has an example, that'd very much appreciated. 07:33:55
18 May 2023
@yuu:matrix.org@yuu:matrix.org joined the room.14:04:45

Show newer messages


Back to Room ListRoom Version: 9