Différences

Ci-dessous, les différences entre deux révisions de la page.

--- proxmox:zfs-perf [2023/07/30 12:22] – root
+++ proxmox:zfs-perf [2025/11/15 04:26] (Version actuelle) – root
@@ Ligne 8: / Ligne 8: @@
 #ajouter un disque log et cache
-zfs add -f test-zfs log /dev/sdxx cache /dev/sdxxx
+zpool add -f test-zfs log /dev/sdxx cache /dev/sdxxx
 </code>
+### Trouvé sur le net ###
+Symptom:
+Copying anything—even an ISO—caused I/O delay to spike to 40–90%, the VM froze, and the whole node choked.
+Even with only one VM and plenty of CPU/RAM.
+I know, I know…
+“You’re not supposed to use consumer SSDs in production.”
+Totally agree.
+But sometimes a client chooses the budget they choose and the job is to make it work as safely as possible.
+Anyway…
+✔️ The Root Cause
+ZFS synchronous writes + consumer SSDs = absolute misery.
+Consumer SATA SSDs have:
+    slow fsync latency
+    tiny SLC caches
+    no power-loss protection
+    awful random write performance once the cache fills
+    controllers that can stall under ZFS write patterns
+Even with a 2-disk mirror, copying a file would hit the end of the SLC cache → SSD latency would jump → ZFS TXG flushes stalled → Proxmox I/O delay went crazy.
+I also spun up another test box at home using Intel DC enterprise SSDs and none of these issues showed up — so the hardware difference was the smoking gun.
+✔️ The Fix
+These ZFS dataset settings instantly stabilized the system:
+zfs set sync=disabled rpool/data
+zfs set atime=off rpool/data
+zfs set recordsize=64K rpool/data
+What each does (short version):
+    sync=disabled → stops ZFS from forcing every tiny write to hit the SSD immediately.
+    (Yes, slight risk during an unexpected power loss. We have a UPS and BDR.)
+    atime=off → stops ZFS from doing metadata writes for every read.
+    recordsize=64K → better block size for VM workloads.
+After these changes:
+    ISO copies completed instantly
+    I/O delay dropped from 90% → 1–5%
+    Windows VM became responsive
+    No more host freezing
+Night and day.