From f84fec9501a41016b767ad21228a81fecaf4ec96 Mon Sep 17 00:00:00 2001 From: Zoltan Csizmadia Date: Sat, 5 Sep 2026 20:18:34 -0500 Subject: [PATCH] docs: correct how Direction B got fast - no WSL update was involved I wrote up the TCP-relay improvement as though a WSL update had delivered it. It did not, and the real explanation is more useful. Nothing was updated recently on this machine. Kernel 6.18.33.1 was built on 5 June and the WSL package staged on 26 June; the slow-relay numbers were recorded on 27-29 August, on kernel 6.6; the WSL VM restarted on 1 September and has run 6.18 since. A running WSL2 VM keeps the kernel it booted with - updating the package on disk does not restart it. So the newer kernel sat installed and unused for two months while the VM carried on with the old one, and the improvement appeared only when the VM next restarted. The consequence is what makes it worth recording: you cannot tell which behaviour a machine will show from its installed WSL version alone, because it depends on whether the VM has restarted since the update. That argues for probing the transport at run time rather than gating on a version number, and it means other machines may be sitting on the same latent improvement unaware. Refs #110 --- bench/RESULTS.md | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+) diff --git a/bench/RESULTS.md b/bench/RESULTS.md index bd3ef80..74abca6 100644 --- a/bench/RESULTS.md +++ b/bench/RESULTS.md @@ -265,6 +265,30 @@ That is one machine and one WSL version, so the hvsocket path stays and stays recommended; but the claim that Direction B *requires* it no longer holds here, and it is worth re-checking before treating hvsocket setup as mandatory. +#### It was not a recent update, and that matters + +The obvious reading — "a WSL update fixed it" — is wrong, and the real story is +more useful. Nobody updated anything here: + +| | | +|---|---| +| kernel 6.18.33.1 built | 5 Jun 2026 | +| WSL package 2.7.8.0 staged on this machine | 26 Jun 2026 | +| slow-relay numbers above recorded, on **kernel 6.6** | 27-29 Aug 2026 | +| WSL VM restarted (now running 6.18) | 1 Sep 2026 | +| TCP round-trip measured at 0.224 ms | 5 Sep 2026 | + +A running WSL2 VM keeps the kernel it booted with; updating the package on disk +does not restart it. So the newer kernel sat installed and unused for two months +while the VM carried on with 6.6, and the improvement only appeared when the VM +next restarted. + +The practical consequence: **you cannot tell which behaviour you will get from +the installed WSL version alone.** It depends on whether the VM has restarted +since the update. That is a strong argument for probing the transport at run +time rather than gating on a version number, and it means other machines may be +sitting on the same latent improvement without knowing. + ### The mutual handshake costs nothing measurable The handshake gained a third leg when both ends started authenticating each