| 24 Aug 2023 |
 Zhaofeng Li | If you have deep visibility/metadata into the store itself, the store itself can even be the chunk cache | 19:42:04 |
 raitobezarius | this I am planning too flokli but it will be when Raito Corp will be done and we will do 400Gbps distribution of the software universe | 19:42:21 |
| raitobezarius | it's surprisingly hard to get OCP to give you a free 400Gbps switch | 19:42:34 |
| Zhaofeng Li | In reply to @raitobezarius:matrix.org
please give me 5Gbps saturating cache Oh nice, I have an appointment to install 5G Google Fiber next month, but my roommate is thinking about bailing out because of the cost | 19:42:38 |
| raitobezarius | In reply to @zhaofeng:zhaofeng.li
Oh nice, I have an appointment to install 5G Google Fiber next month, but my roommate is thinking about bailing out because of the cost well I pay it 30EUR/mo | 19:42:51 |
 flokli | In reply to @zhaofeng:zhaofeng.li
If you have deep visibility/metadata into the store itself, the store itself can even be the chunk cache This is already the case actually - I don't realize the actual files on disk currently, you can look at them through fuse | 19:43:00 |
| raitobezarius | but I support you towards 5G | 19:43:08 |
| flokli | Rust in kernel VFS when | 19:43:10 |
| raitobezarius | In reply to @flokli:matrix.org
Rust in kernel VFS when :DDDDDDDDDDDDDDDDDD | 19:43:21 |
| Zhaofeng Li | In reply to @flokli:matrix.org
This is already the case actually - I don't realize the actual files on disk currently, you can look at them through fuse Very cool, didn't know that! | 19:44:00 |
| raitobezarius | https://cs.tvl.fyi/depot/-/blob/tvix/store/src/fuse/mod.rs | 19:44:23 |
| raitobezarius | if I'm not wrong | 19:44:45 |
| flokli | It's all a bit wip, there's no seek support, because I want to do this in concert with verified streaming | 19:45:12 |
| flokli | But it should happen all in the next months | 19:45:30 |
| flokli | But it should already be good enough for builds | 19:46:51 |
| flokli | I'll talk about it during nixcom | 19:47:24 |
| flokli | * I'll talk about it during nixcon | 19:47:32 |
 BMG | In reply to @flokli:matrix.org
I'll talk about it during nixcon You giving a scheduled talk? | 20:10:11 |
| raitobezarius | a smol talk about a certain reimplementation of Nix | 20:10:26 |
| BMG | Can always chat about it anyway over a few 🍻 | 20:10:37 |
| flokli | Smalltalk | 20:10:37 |
| raitobezarius | something something about putting the oxide inside the nix | 20:10:45 |
| raitobezarius | ok let's push this schedule | 20:10:56 |
| raitobezarius | * ok let's push this (NixCon) schedule | 20:11:00 |
 @elvishjerricco:matrix.org | Ok I redid my dumb nar cache and zfs dataset. I made a spreadsheet of all the nar file sizes and found that nars <= 512KiB account for 1.2% of the data, but 80% of the files. So I rsync'd the data to a new dataset so that I could set recordsize=1M and special_small_blocks=512K. Now there's 7GiB on the special optane vdev and the rest of the 371GiB is on the HDDs, and it's fast as hell | 23:50:10 |
| raitobezarius | send us statistics | 23:50:22 |
| @elvishjerricco:matrix.org | what are you interested in? | 23:50:44 |
| raitobezarius | if you run a heavy duty transfer, IOPS/throughput? | 23:51:16 |
| raitobezarius | at the ZFS level | 23:51:21 |
| raitobezarius | but also at the application level | 23:51:28 |
