New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
What would be the typical disk performance on a fully loaded OpenVZ?
What would be the typical disk performance on a fully loaded OpenVZ?
Fully loaded is probably a very relative term. But what would be say dd (with fdatasync) performance on a system with (your definition of) most customers in a non SSD setup. I know a lot dd performance posts are there. But would help to know the full load performance.
My assumption is that mostly fully loaded would be the best time to benchmark. (excluding the possible abuse events).
Is there any other criteria to be looked at when benchmarking for VPS for longer term usage. Looking at quantitative aspects.
Comments
That's too much relative to even try to answer. How much food can people eat and still function normally?
It's not problem just with "fully loaded" relative term but also with "typical disk performance". What's "typical" here at all?
@spirit thats one hell of an analogy
.. i understand its difficult.. but you can always say. how much is enough food for lunch. Then you (i guess) stop eating for that time (atleast as a provider one takes a call on when to stop adding new nodes to the same hardware) I am looking at that point.
Logging is one of the worst problems. If you get 30 VPS's verbosely writing little itty bitty pieces of data one line at a time multiple times per second then that really starts to kill your I/O. If you have the ability to control individual VPS logging I would STRONGLY suggest you try reduce the verbosity of logging. It will help a lot.
No, no... problem is deeper.
Who are those people? Kids? Women? Physical workers? And even if we know this we're not done yet because there's another question. What does it means "normal functioning"? So if I ask you again how much food exactly can people eat and still function normally? In numbers please.
One of my obviously fully loaded VPSs disk I/O throughput speed was 16.6 MB/s. Obviously fully loaded. Or crappy hardware. Or mixture of both. Another tested VPS I/O throughput speed was 9.2 MB/s. Even more fully loaded? Or just worse hardware? Or abusive users? Third fully loaded VPS performed 20.5 MB/s.
What's typical peformance here anyway?
On a scale from Prometeus to ChicagoVPS, how many clients is considered "fully loaded"
Fully loaded is such a relative term, it can't be accurately measured. You should look for the criteria that meets your specific needs. You can do all the math you want, but the math from one host won't equal the math from another host. You can try, but you'll most likely end up disappointed.
disk performance of fully loaded ovz is less than of an empty node
Qftw.
We consider our systems full when the average disk latency gets up to about 40ms on 5-10% average utilization. We've found that these numbers provide us with the best quantity of loading versus client expectations. Sure, we can put more people on the server, but there's not really much need to do so.
@jcaleb - you rock
@spirit - excellent point (not the analogy though :P) ..understood the crux of your point. So what would be the ideal time to performance benchmarking on a hardware. @Damian 's input is in the same direction as what i was seeking. Providers can overload to any point as they wish.
But what would be right point considered as breaking point. dd alone might not be the right paramater. But a CF card speed would definitely be worrying. isnt it?
As a guest OS, i dont have enough right tools to measure/benchmark my VPS. dd seems to be the one option (atleast thats what i know of).
One of the specific points that I don't bother with "the dd test" is that most (all?) of our customers aren't doing 1 gigabyte sequential writes, so a method that tests with 1 gigabyte sequential writes is rather useless for us.
Yes, i'm sure we've lost a few new clients who sign up and do "the dd test" and don't get massive arbitrary numbers, but the remainder of our client-base who use our systems that we've tuned for random i/o operations, which is the true use of the system especially on a VM node, generally like the service.
The difficulty is that there's no good easy/simple/standardized method for testing random i/o, so "the dd test" prevails among this community.
This. dd is more of a '1000 foot' tool than a useful standalone benchmark.
The clueless masses like it when big numbers are crapped out, plus a monkey can type a one-liner . . . so it's become sadly ubiquitous.
Prometeus is a bad example to use because their definition of a fully loaded node would be most providers' definition of a nearly empty node.
I believe that the recent database leak showed that CVPS's definition of "fully loaded" is selling 200-300GB of RAM on some of their 32GB E3 nodes.
HostSlim's definition of fully loaded was overselling everything, including the HD, to the point where the node I was on ran out of drive space a couple of times.
TL;DR there is no answer to the OP's question because A. every OpenVZ provider has a different comfort point in overselling, and B. drive performance also depends on the node specifications, type of drives used, type of users the host attracts (business customers running websites will probably be less disk intensive users than [fill in the disk abuse type blank users], etc.
Maybe he is asking what to look for disk speed on fully loaded node.
Thanks for the mentions.
Here we are looking mostly at CPU problems because our storage is generally SAS raids of 6 or more drives or SSD or SAN.
SSD storage doesnt have the problems random writing creates on a mechanical drive and as such have a lot more IOPS available. On the other hand, we rarely employ large SSD arrays as we mostly use them for 16/32 GB E3s.
This leads to a relatively equal performance for "dd" throughout the spectrum, from fast local storage arrays limited by their mechanical nature, through SAN huge arrays with 32 GB cache and big active-active controllers limited by the FC pipe's BW to superfast SSD drives limited by their small numbers and software raid.
We consider anything over 100 MB/s in dd and stable iopings as adequate, but strive to have 300+ and below 1 ms which is not always possible.
Numbers such as 1 GB/s happen in one regular server we put as an experiment and in the myoffload mysql server, but anything over 100 MB/s should be more than enough on a fully loaded server (most are right now as we are out of stock except iwstack) in almost all usage scenarios (baring some busy databases or fast log writers for which we recommend SSD or myoffload if possible).
What tool do you use to measure that?
Interesting to hear different view points...
Providers might have different view points on what a fully loaded is. But what would be the quick dip test for any buyer to evaluate a VPS. When people say "rock solid hardware", i always wonder what they base things on. Unix benchmark? some other benchmark tools? dd test is the simple test, but as @microlinux said, its a 1000 ft high.
network uptime is fairly straight forward to measure. Support, though subjective could also be measurable(in a sense). but other than dd, do I have anything, which would tell me my VPS providing services, promised to me. Over selling is completely acceptable. Over-over-over selling? Any possible measures? of course your "node neighbours" could be sometimes abusive, you cant predict, but consistently abusive?
When Iperweb and @iniz started offering mega RAM offers, they were transparent enough in telling carefully oversold. In fact, some of these VPS, do provide me really good results in dd or in unixbench (consistently)
@all
All these posts just sound like... "what do you mean fully loaded?, there is no such thing as fully loaded"
That's because the concept doesn't make sense in this context.
Maybe their hardware is as solid as a rock?
10Mb/s had this today.