Skip to content
Register

Vultr 4k random write is terrible

VPS Hosting Locked by uma 5 replies 174 views
This thread is locked — new replies are not accepted. (Closed: resolved)
uma

Member OP

99.99% or bust

13 #1

Collected results across 8 active instances. All tests run with YABS v2024-06, fio 3.35, iperf3 3.16. The graphs don't lie.

  • Contabo — 1423 / 823 MB/s / 1.2 GB/s / 940 Mbps
  • OVHcloud — 1567 / 412 MB/s / 890 MB/s / 920 Mbps
  • Vultr — 1389 / 91 MB/s / 340 MB/s / 880 Mbps
  • Hetzner — 1498 / 756 MB/s / 1.1 GB/s / 960 Mbps
  • InterServer — 1523 / 634 MB/s / 980 MB/s / 910 Mbps
  • Hostinger — 1445 / 701 MB/s / 1.0 GB/s / 930 Mbps
  • KnownHost — 1612 / 589 MB/s / 940 MB/s / 950 Mbps
  • RackNerd — 1401 / 445 MB/s / 720 MB/s / 890 Mbps

Vultr stands out. 91 MB/s on 4k random write is concerning. Status page shows no incidents. Alert fatigue from false positives is better than missing t

436 days. reboot is surrender.
svendeal

Member

Deal Sniper

#2
uma said:
The graphs don't lie.

Back in my day we ran YABS equivalents on VAXstations with real SCSI drives and you could feel the seek heads move

Vultr is overselling that disk so hard the platters are probably crying... someone else's computer strikes again

systemd probably caused this somehow

https://github.com/masonr/yet-another-bench-script

world record: 4min Arch install
garykwh

Member

Inference on a Budget

#3

Lol 91 MB/s for 4k random? Ngl my LLM inference scratch disk does better on a usb stick

VRAM math says thats like... 0.3% of a single A100 HBM channel?

But fr tho Vultr what are you doing

CUDA cores are my love language
pieter_rtm

Member

#4

As said, 412 MB/s 4k is within spec for shared NVMe.

Specs for comparison:

  • 2 vCPU AMD EPYC 7443
  • 4 GB RAM
  • 80 GB NVMe, no RAID overhead
  • 1 Gbps unmetered

Dry humor: at least we beat Vultr by 4.5x.

Containers before it was cool
FlowSana

Member

SRE

#5
  1. Vultr's result indicates either severe oversubscription or degraded storage pool health.
    • fio 3.35 with --direct=1 and libaio engine is the correct methodology here.
    • A healthy NVMe pool should sustain 300+ MB/s for 4k random writes per tenant.
    • I have observed similar patterns on Ceph clusters with 85%+ full OSD sets.

Practical recommendation: Contact Vultr support with ticket reference and request pool migration. Escalate if first-line denies storage issues.

It's always DNS. Always.
thu_miller

Member

4 #6

The graphs don't lie. But Vultr might. 91 MB/s is not NVMe. Its barely SATA SSD. Maybe IDE. I had a server once. 2009. Faster disk...

Thread closed

Replies are closed on this thread. Closed: resolved.