VSP 360 · Volume placementBlock storage · 1,056 active volumes · 16 Oct 2026, 17:20–18:20 CEST

Five busy volumes run 4 to 14 times slower than their neighbours. Four of them share one pool on Array A while Array C sits idle.

Most active volumes answer in about 0.12 ms. The exceptions are Array A 00:00:06, 00:00:01, 00:00:1C and 00:00:1D, all in Pool A0, which carries 13,800 IOPS and peaked at 1.42 ms this hour, and the scan host's volume on Array B, which is slow because of its iSCSI path rather than its pool. None of them crosses 2 ms on average.

0.12ms
Typical volume
Median response time of 1,056 active volumes
5volumes
Slow for their load
At least 100 IOPS and 3× their array's median
13,823IOPS
Array A Pool A0
One pool, about 91% of the array's I/O
~100IOPS
Array C, same site
137 TiB free in its Pool 0
  1. Array A Pool A0 is the misplaced pool, more than any single volume. It holds every busy Array A volume and peaked at 1.42 ms, while Pool A1 on its own RAID5 group ran 1,353 IOPS at 0.2–0.6 ms and Array C at the same site ran about 100 IOPS.
  2. 00:00:06 is the first volume to move. It reads 203 KiB blocks, mostly sequentially, at 3,200 IOPS and 1.0 ms. Its latency comes from its I/O size, and its bandwidth competes with the small random I/O of its pool neighbours.
  3. The scan host's volume on Array B is on the wrong path, not the wrong pool. Its pool answers in 0.15 ms, but the volume averages 1.57 ms over 10 GbE iSCSI.
  4. Two NVMe/TCP namespaces read at 7–10 ms with only 3–4 IOPS. That points at the NVMe/TCP network or NVMe host 1, not placement.
Figure 1 · IOPS against latency, one dot per volume

Volumes that are slow for their load stand out above the cloud at the right; slow volumes at the left are idle

Both axes are logarithmic. The dashes along each axis are the same volumes projected onto that axis, so their density shows each distribution on its own. Diagonal lines join points of equal queue depth (IOPS × response time). Hour averages, response time weighted by IOPS. Array C volumes are from 16:15–17:15, its latest complete hour.
Detail

What to move, and where

× median compares each volume with the median response time of active volumes on its own array (Array A 0.124 ms, Array B 0.110 ms). Read size is the average read I/O size.
Also misplaced, in the other direction

83 volumes of 1 TiB or more did no I/O on the NVMe arrays (A and B) this hour