Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!


Detect overselling on OpenVZ
New on LowEndTalk? Please Register and read our Community Rules.

Detect overselling on OpenVZ

HerrMaulwurfHerrMaulwurf Member
edited September 2011 in General

Hello, I'm currently testing several providers which offer openVZ based VPS. Is it possible to test/check whether the hardware nodes are oversold? Especially RAM and CPU.

Best regards
HerrMaulwurf

«1

Comments

  • About RAM, maybe this helps

    http://www.lowendbox.com/blog/how-to-tell-your-openvz-vps-is-swapping/

    About CPU, do some benchmarks and aim for consistent or similar results.

  • kiloservekiloserve Member
    edited September 2011

    HerrMaulwurf said: whether the hardware nodes are oversold? Especially RAM and CPU.

    Everybody who sells VPS services "oversells"... 100% of VPS providers oversell.

    The question is whether or not it is overloaded.

    An overloaded server will give you inconsistent benchmarks or very low benchmarks all the time.

    As for benchmarks, here are a few general rules of thumb that we have come up with to gauge our own VPS servers vs other offerings:

    1) For disk I/O tests
    < 40 MB/s is below average
    40 - 80 MB/s is about average
    80 MB/s -120 MB/s is very nice
    120 MB/s + is outstanding
    2) For Unixbench CPU benchmarks (single CPU)
    < 150 is poor
    150-300 is ok for a budget VM
    300-400 is average
    400-500 is very good
    500+ is outstanding for a budget VM
  • KuJoeKuJoe Member, Host Rep

    @kiloserve False.

  • kiloservekiloserve Member
    edited September 2011

    @KuJoe Please explain. Are you saying that your company does not oversell on VPS nodes?

  • KuJoeKuJoe Member, Host Rep
    edited September 2011

    Is it that hard to believe that some companies don't oversell?

    Yes, I know the argument that they are selling resources somebody else oversold to them, but that's not a valid argument since they are selling the resources that they paid for. Some resources that people say are oversold are usually not guaranteed resources such as CPU, disk I/O, network ports, etc... if it's a shared resource and not dedicated it cannot be labeled as oversold. The 3 common things that are "dedicated" when it comes to VPSs are RAM, Disk Space, and Bandwidth.

    RAM does not have to be oversold to make a profit, but you can easily get by with underselling RAM.
    Disk Space is harder to undersell, but still easy to do if you build your own servers AFTER planning.
    Bandwidth is commonly oversold, but considering how many users we have that use 100% of their bandwidth each month, we refuse to oversell it.

  • kiloservekiloserve Member
    edited September 2011

    KuJoe said: Some resources that people say are oversold are usually not guaranteed resources such as CPU, disk I/O, network ports, etc... if it's a shared resource and not dedicated it cannot be labeled as oversold. The 3 common things that are "dedicated" when it comes to VPSs are RAM, Disk Space, and Bandwidth.

    Ah, you put a distinction on disk i/o, CPU. Just a difference in definition. I can see why you say you do not oversell using your definition.

    To us, all subsystems of the server are included, this means disk and CPU i/o as well and cannot be separated. It is like saying "We have been in business for 20 years after you add in 15 years".

    For example, if you have a server with 100GB RAM and a single 7200 RPM drive, you can fit 99x 1GB clients in and it is "technically" not overselling. But the single 7200 RPM drive will be vastly oversold in our opinion and lead to junk performance.

    We tell our clients that our servers are always "oversold" but never "overloaded".

    We don't like to put a distinction on Disk I/O or CPU and count them separately, we consider the server as a whole, and that is why our definition is different.

    Your definition is different than ours but also correct as you have specified.

  • Thanks for all the hints guys :)

    Disk I/O moves between 25MB/s and 60MB/s and is quite unstable. But I don't know whether this is due to (heavy) overselling or just normal... Unixbench score is 636, but I checked only once.

  • kiloservekiloserve Member
    edited September 2011

    HerrMaulwurf said: Unixbench score is 636, but I checked only once.

    Unixbench score of 636 is quite high for a single CPU ( relative to the 25-60MB/s disk I/O), is that just 1 CPU or the summary final total with all the CPU's?

    The fluctuation between 25 and 60 MB/s is quite a bit. If it generally stays towards the 60MB/s side you should be fine.

    Now you have a baseline, check it in a couple of days and see if it drops more. And if it does, you are probably overloaded on disk I/O.

    If you don't really do much disk activity, 25 MB/s may even work for you.

  • The VPS is running on a Xenon based hardware node, but has just one core enabled which is locked at 2000MHz. 636 is the System Benchmar Index Score which I guess is the final score. I also filled the RAM and checked whether the VPS swaps or not. There was only little swap usage, so RAM is not oversold I guess.

    I checked I/O again and I've got only 17,2MB/s at the moment. I'll test it several times again and I'll maybe ask the hoster if he can tweak the performance a bit.

    Thanks for your help!
    HerrMaulwurf

  • KuJoeKuJoe Member, Host Rep

    Keep in mind that the dd test is not very accurate on a populated server because if you run dd at the same time somebody else it (which is extremely common I've noticed) your results will be extremely lower than normal. While it is a good tool to give you an idea of what you're getting, in reality it is only effective using an average of multiple tests spanned across different time frames.

    @kiloserve While I can understand your definition I only see those as separate when the company offering VPSs does not specify an amount. If we weren't a LEB provider I would definitely specify those types of resources.

  • ramnetramnet Member, Host Rep

    "dd test is not very accurate"

    that's all that needs to be said.

    still boggles the mind how many ignorant people think dd means anything. I can manipulate a dd test to make a powerful server looks like junk and a junk server look like the fastest server in the world.

    take a look at a little tool called ioping if you want a realistic metric. or just watch %wa in top for a bit.

  • Which is why I keep suggesting a list of commands which we can use as a scale to compare boxes. :)

    I can manipulate a dd test

    So can anyone. It's just a copy and paste.

  • drmike said: So can anyone. It's just a copy and paste.

    HAHAHAHA!!!! pwnt n_n

  • I kid you not. I look at some of these reports that folks make, lo0ok at the numbers and go "That's not right..."

    Been meaning to bring that up actually.

  • I cringe every time someone, especially providers who should know better, suggests using DD as an I/O benchmark.

  • yomeroyomero Member
    edited September 2011

    OMFG a bot!
    I am not ready to get married with that pretty girls u_u

    EDIT: LEA deleted the bot u_u so, this isn't funny anymore u_u

  • suggests using DD as an I/O benchmark.

    You use what tools that you can use. It's considered a benchmark even though it's not the best.

    Of course the poorly used "bandwidth" sets me off.

    Thanked by 1Microlinux
  • KuJoeKuJoe Member, Host Rep

    I hate when people use Cachefly and only Cachefly as a benchmark. For some reason Cachefly is extremely slow for our servers but there are dozens of other tests out there that will max out our connection speeds... but since they aren't Cachefly they don't count I guess. :(

  • You use what tools that you can use. It's considered a benchmark even though it's not the best.

    For sequential performance . . . the real beef is in random I/O performance.

  • if you've got a one line test that does random IO using tools that come in most distros as standard, let's hear it. :P

  • I do not, I/O is not an easy thing to test, but I'm not willing to accept inconclusive results just because they were easy to obtain. Objective performance testing is a not a one line operation.

  • re: I/O. One of the reasons why looking at bus speeds is important to me. :)

  • drmike said: re: I/O. One of the reasons why looking at bus speeds is important to me. :)

    Any specific arguments to that? IMHO that doesn't matters very much (if you mean FSB) because is mainly associated with memory bandwidth, and the memory is several times faster than any I/O subsystem

  • yomero said: Any specific arguments to that? IMHO that doesn't matters very much (if you mean FSB) because is mainly associated with memory bandwidth, and the memory is several times faster than any I/O subsystem

    .. not sure, but PCI/PCIe bus speed may matter while setting a RAID array with the help of an PCI/PCIe card.

  • The time for initiating a transaction through the PCI/PCIe bus is measured in nanoseconds. The time for moving the disk head from one location to another is measured in milliseconds. Guess which matters more for your random i/o scenario ;-)

  • Indic said: .. not sure, but PCI/PCIe bus speed may matter while setting a RAID array with the help of an PCI/PCIe card.

    Yes, the PCI express bus, and better if is a 2.0 card, and better if is a 4x card, but not the CPU bus.

  • kiloservekiloserve Member
    edited September 2011

    Microlinux said: I cringe every time someone, especially providers who should know better, suggests using DD as an I/O benchmark.

    Hello Microlinux

    I am a "provider" and I recommend DD as an easy benchmark to do.

    Can you or any of the above posters who say DD is a "bad" benchmark explain why it's bad?

    We all understand it's not the end all, be all but in no way is it false. Even if it's only sequential, It is fairly representative of the system's disk I/O.

    If a user posts a DD score of "200 MB/s+" , he has great I/O. When would this be false?

    If a user post a DD score of "10 MB/s", his I/O sucks. When would this be false?

  • I have 2 VPSs that get around 50MB/s on the dd test, but one is slow due to io wait and the other is much faster.

  • kiloservekiloserve Member
    edited September 2011

    dmmcintyre3 said: I have 2 VPSs that get around 50MB/s on the dd test, but one is slow due to io wait and the other is much faster.

    Hello Dr Mcintyre

    Did you run the DD during the high I/O wait with fdatasync? If you run DD during high I/O wait, the speed will plummet since DD cannot bypass I/O waits.

    Especially on high I/O wait servers, DD should come up with very bad results.

  • I think the problem is more that people are relying on a single dd result and think that it's representative of the server. Basic scientific method, people!

Sign In or Register to comment.