Skip to content

Mini-guide: detecting and documenting provider overselling

General Discussion by bravomartha 7 replies 366 views
13 #1

Mini-guide my foot I have been dealing with oversold hosts since 2019 when I first got my kitten Mr whiskers and he was sitting on my lap when the first OVHcloud VPS went down and I thought it was the cat but no it was the server and I had eggs for breakfast that day it was sunny out and I remember because I had to open the blinds to check the router lights and the eggs were over easy which is ironic because the server was over sold and that is what this guide should actually cover so here is my method first you run fio on the disk and if the numbers bounce all over that is neighbor noise second you check CPU steal with top and if it is over 5 percent for more than ten minutes that is bad third you ping the gateway every second for an hour and look for jitter patterns and write it all down with timestamps and post it and then people tell you to read the TOS but you already did and they do not care

#2

Hehe mano bravomartha is legend kkkkk but he is right, my VPS it crashed when neighbor run ffmpeg and I not even know until I check `iostat` caramba! More fast to detect = more fast to document, nossa!

#3
bravomartha said:
Check CPU steal with top

Back when we ran bare metal, "steal" was not in our vocabulary. Kids these days rent slices of mystery boxes and wonder why performance varies. Your three-step method is sound. I would add: establish baseline on day one, before neighbors arrive. Mark my words: providers will soon offer "dedicated vCPU" that is just longer time slices on the same oversold silicon. The marketing always outpaces the engineering.

IPv4, IRC, and irssi — fight me
#4

Pro tip: Oracle Cloud free tier makes an excellent control for comparison testing. Same region, same shape, known-noisy vs suspected-noisy.

1. Launch two instances: one Oracle (baseline), one target provider
2. Run identical fio profile: `fio --name=randread --ioengine=libaio --iodepth=32 --rw=randread --bs=4k --direct=1 --size=4G --runtime=300 --numjobs=4`
3. Collect `iostat -x 1 300` on both simultaneously
4. Compare 95th percentile latency, not average

Heads up: Oracle's IOPS are capped but consistent. Deviation above Oracle = likely oversold. Deviation below = run.

licensing is a suggestion
#5

Docker is nothing but just cgroups and namespaces and your benchmark container adds scheduler noise þþþ systemd is the real problem here, not the provider, containers at home are a disease

#6

What baseline tool you using, harold

SPF, DKIM, DMARC — holy trinity ✉️
#7

I had the same ffmpeg neighbor crash

#8

Mr whiskers was the real victim here

boot anything, anywhere, anytime

Post a reply

You need an account to reply. Log in or register to join the conversation.

Post reply Preview Save draft