about measuring overselling

miaomingc

How do you guys measure vps overselling.
I wrote a simple program to run a load on cpu and calculate cpu steal time
https://github.com/miaomingc/cpu-oversell-detector

mrs92

How about disk and network?

miaomingc

@mrs92 said:
How about disk and network?

Disk is cheap for vps provider,no need to oversell.
But they may use low speed hdd + cache instead of sdd drive .Depending on the buffering strategy, it's hard to detect if they do

mrs92

@miaomingc said: Disk is cheap for vps provider,no need to oversell.

I mean disk io.

cybertech

does this work on hetzner?

miaomingc

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

miaomingc

@cybertech said:
does this work on hetzner?

use lscpu to find out which virtualization you use

wadhah

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

emgh

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

wadhah

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

emgh

@wadhah said:

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

lmao a bunch of ridiculous replies from Hetzner there, golden

cybertech

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

please put disclaimer does not work on hetzner because they disabled cpu steal

emgh

@emgh said:

@wadhah said:

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

lmao a bunch of ridiculous replies from Hetzner there, golden

Even to the YouTube question

”WHAT DIFFERENCE WOULD IT MAKE!?!?”

Lmao

wadhah

@emgh said:

@wadhah said:

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

lmao a bunch of ridiculous replies from Hetzner there, golden

It was so hard for me to stay polite in that thread

emgh

@wadhah said:

@emgh said:

@wadhah said:

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

lmao a bunch of ridiculous replies from Hetzner there, golden

It was so hard for me to stay polite in that thread

You did good, I would have lost it 110 %

I think everyone here can comfirm

maverick

@emgh said:

@wadhah said:

@emgh said:

@wadhah said:

@emgh said:

@wadhah said:

@miaomingc said:

@cybertech said:
does this work on hetzner?

Should work on any time-slice-based virtualization,like KVM

No hetzner disabled cpu steal reporting, so this should not work on their vpses.

what what, they did?

Yeah couple of months ago,

https://lowendtalk.com/discussion/200758/has-hetzner-disabled-cpu-steal-time-reporting/p1

Not sure if it's for all their slices though, probably only the oversold ones

lmao a bunch of ridiculous replies from Hetzner there, golden

It was so hard for me to stay polite in that thread

You did good, I would have lost it 110 %

I think everyone here can comfirm

my i m comfirm

Mrsimon

Wow.overselling is a really common issue in the VPS industry.

Mrsimon

@miaomingc said:

@mrs92 said:
How about disk and network?

Disk is cheap for vps provider,no need to oversell.
But they may use low speed hdd + cache instead of sdd drive .Depending on the buffering strategy, it's hard to detect if they do

“Fusion drive”

Motion3549

How long I should run this ? @miaomingc

AlexPads

Hello,

I would like to note that this detector runs your CPU at 100% usage the entire time its running. On our platform if you do this for more than a few hours your VM will become CPU throttled.

VDS services would not be throttled as the CPU is dedicated.

Best,

Motion3549

@AlexPads said:
Hello,

I would like to note that this detector runs your CPU at 100% usage the entire time its running. On our platform if you do this for more than a few hours your VM will become CPU throttled.

VDS services would not be throttled as the CPU is dedicated.

Best,

Good to know about your transparency.

AlexPads

@Motion3549 said:

@AlexPads said:
Hello,

I would like to note that this detector runs your CPU at 100% usage the entire time its running. On our platform if you do this for more than a few hours your VM will become CPU throttled.

VDS services would not be throttled as the CPU is dedicated.

Best,

Good to know about your transparency.

This is already highlighted in our Acceptable Usage Policy. Its very standard practice for Virtual Servers.

Motion3549

@AlexPads said:

@Motion3549 said:

@AlexPads said:
Hello,

I would like to note that this detector runs your CPU at 100% usage the entire time its running. On our platform if you do this for more than a few hours your VM will become CPU throttled.

VDS services would not be throttled as the CPU is dedicated.

Best,

Good to know about your transparency.

This is already highlighted in our Acceptable Usage Policy. Its very standard practice for Virtual Servers.

it would be great if providers made the most frequently asked questions (i.e yabs, looking glass etc) by customers about their services upfront on their sales page.

Dyingcat

Simply run yabs and compare its gb6 multiple with single score.

Some providers limit disk IO so as to make it usable till the end of this century. So disk IO can't tell overselling.

lukast__

@Dyingcat said: Simply run yabs and compare its gb6 multiple with single score.

No, GB6 multicore is basically meaningless with more than 4-6 cores.

Dyingcat

@lukast__ said:

@Dyingcat said: Simply run yabs and compare its gb6 multiple with single score.

No, GB6 multicore is basically meaningless with more than 4-6 cores.

I meant 1 vcore.

Usually you can see gb6 score single =1234, but multicle =1123, the overselling is obvious.

If you do yabs on multicore vps, limit to 1 core. One can turn to chatGPT for help.

lukast__

@Dyingcat said:

@lukast__ said:

@Dyingcat said: Simply run yabs and compare its gb6 multiple with single score.

No, GB6 multicore is basically meaningless with more than 4-6 cores.

I meant 1 vcore.

Usually you can see gb6 score single =1234, but multicle =1123, the overselling is obvious.

If you do yabs on multicore vps, limit to 1 core. One can turn to chatGPT for help.

Actually, no, I don't think this works either, because if there's CPU steal it's also impacting the single core score, and given that GB6 multi-core works like this:

Geekbench 6 uses a “shared task” model for multi-threading, rather than the “separate task” model used in earlier versions of Geekbench. The “shared task” approach better models how most applications use multiple cores. The "separate task" approach used in Geekbench 5 parallelizes workloads by treating each thread as separate. Each thread processes a separate independent task. This approach scales well as there is very little thread-to-thread communication, and the available work scales with the number of threads. For example, a four-core system will have four copies, while a 64-core system will have 64 copies. The "shared task" approach parallelizes workloads by having each thread processes part of a larger shared task. Given the increased inter-thread communication required to coordinate the work between threads, this approach may not scale as well as the "separate task" approach.

I would say that if the multi-core score is lower than the single core score, that's solely because of thread scheduling and inter-thread communication/coordination that isn't necessary for the single core benchmark.

Motion3549

@lukast__ said:

@Dyingcat said:

@lukast__ said:

@Dyingcat said: Simply run yabs and compare its gb6 multiple with single score.

No, GB6 multicore is basically meaningless with more than 4-6 cores.

I meant 1 vcore.

Usually you can see gb6 score single =1234, but multicle =1123, the overselling is obvious.

If you do yabs on multicore vps, limit to 1 core. One can turn to chatGPT for help.

Actually, no, I don't think this works either, because if there's CPU steal it's also impacting the single core score, and given that GB6 multi-core works like this:

Geekbench 6 uses a “shared task” model for multi-threading, rather than the “separate task” model used in earlier versions of Geekbench. The “shared task” approach better models how most applications use multiple cores. The "separate task" approach used in Geekbench 5 parallelizes workloads by treating each thread as separate. Each thread processes a separate independent task. This approach scales well as there is very little thread-to-thread communication, and the available work scales with the number of threads. For example, a four-core system will have four copies, while a 64-core system will have 64 copies. The "shared task" approach parallelizes workloads by having each thread processes part of a larger shared task. Given the increased inter-thread communication required to coordinate the work between threads, this approach may not scale as well as the "separate task" approach.

I would say that if the multi-core score is lower than the single core score, that's solely because of thread scheduling and inter-thread communication/coordination that isn't necessary for the single core benchmark.

Why your cpu logger in archive mode? It's a great tools.

Rubben

i measure overselling with cpu steal

1 contabo = 50% cpu steal

If cpu steal is consistently above 0.3 contabo, I consider it oversold
at 0.5 contabo, i open a ticket
above 1 contabo i chargeback yesterday

lukast__

@Motion3549 said:

@lukast__ said:

@Dyingcat said:

@lukast__ said:

@Dyingcat said: Simply run yabs and compare its gb6 multiple with single score.

No, GB6 multicore is basically meaningless with more than 4-6 cores.

I meant 1 vcore.

Usually you can see gb6 score single =1234, but multicle =1123, the overselling is obvious.

If you do yabs on multicore vps, limit to 1 core. One can turn to chatGPT for help.

Actually, no, I don't think this works either, because if there's CPU steal it's also impacting the single core score, and given that GB6 multi-core works like this:

Geekbench 6 uses a “shared task” model for multi-threading, rather than the “separate task” model used in earlier versions of Geekbench. The “shared task” approach better models how most applications use multiple cores. The "separate task" approach used in Geekbench 5 parallelizes workloads by treating each thread as separate. Each thread processes a separate independent task. This approach scales well as there is very little thread-to-thread communication, and the available work scales with the number of threads. For example, a four-core system will have four copies, while a 64-core system will have 64 copies. The "shared task" approach parallelizes workloads by having each thread processes part of a larger shared task. Given the increased inter-thread communication required to coordinate the work between threads, this approach may not scale as well as the "separate task" approach.

I would say that if the multi-core score is lower than the single core score, that's solely because of thread scheduling and inter-thread communication/coordination that isn't necessary for the single core benchmark.

Why your cpu logger in archive mode? It's a great tools.

Interesting, just archived it around an hour ago, but I archived it because of two reasons:

• There were a few design flaws, one of which can't be fixed easily without a server (namely that showing all data at once wasn't a good idea because that generates very big HTML files), and the other one is that it's currently doing a single read for each data element which, because of the syscall overhead, is very inefficient and quite slow if one recorded data for multiple months
• The second and main reason is that I'm currently developing a monitoring tool (demo) that records much more types of data (also disk read/write, network usage) and has the advantage that the individual servers do not need to store anything (which also means that no data is lost when the disk/RAID of a server fails), and one doesn't need to ssh to each server to view the statistics, and because of that new tool I do not use cpu_logger anymore.

Of course cpu_logger also has advantages compared to my new tool, such as that it's easier to set up if one only has a few servers, and if one wants, one can of course continue to use it; I just don't plan to maintain it anymore.

hyperblast

@AlexPads said:
Hello,

I would like to note that this detector runs your CPU at 100% usage the entire time its running. On our platform if you do this for more than a few hours your VM will become CPU throttled.

VDS services would not be throttled as the CPU is dedicated.

Best,

sir, does yabs throttle vps from you as well? my vps moved from a 7940HX to a 7945HX and gb6-value declined