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.
Comments
Unless you are forcefully and rather violently clearing out ram (either via custom image, guest agent or some non kosher way), linux will eat your ram.
From a business to another business, don’t oversell. Instead price your resource adequately.
Your customers will love you long time.
as written before, you won't see the 'returned' or deduplicated RAM as a lower number in the reserved column. afaik that is not how it works. but. if you try other hypervisors do let us know of your findings, will ya?
I've had too many stability problems from this to advise it, even with the math perfectly laid out and extremely predictable business customers
This has caused Windows guests to blue screen to no end for me, mostly useful when running many instances of similar Linux guests
Edit: to be clear, Linux instability happens too.. just less often. It seems to provide the best returns
I will admit it's been nearly five years since I gave it an honest try though
I believe you'll want the qemu guest agent and KSM support present in the VMs to get the full effect. Which, if any positive number of the VMs undo, spells possible disaster for everyone. They've changed your formula, unknown to you.
Overselling CPU time is easier than memory time, one is a pool of resources and one is just an additional interrupt.
To share time, one just sleeps - to share memory, you have to play three dimensional Tetris - taking and granting, likely paging to disk somewhere in between
This is why SSDNodes has to be so aggressive and tends to perform poorly.
Unload the balloon module on one of their VMs and within a week you'll be rebooted with it back. Unless you tell it to always return /bin/true when loading the module of course
😂 agreed
@yoursunny Maybe start a hall of shame list of hosts that do this crap? Hopefully it stays a short list.
it's only crap if done wrong. if the provider achieves a good balance and can reach a higher density this might allow for cheaper pricing - and in the end that's what we are after, no?
That's part of the issue, the guests can effectively opt out of this balancing act and cause mayhem by creating unexpected pressure
It just takes unloading the balloon module, and it's easy to block from loading automatically with modprobe.d
yes. that's why ballooning is only one part of the whole thing and also why this is nothing to set and forget but to balance carefully depending on the use case. also why you must have enough swap on the host node to be able and cover for it just in case.
don't get me wrong, I fully agree that there is quite some margin for error and if a provider thinks he can easily cram more people onto a box by just flipping a switch, that's most likely not going to end well.
however the whole topic here wasn't about that, but why the KVM process shows the full amount as reserved memory and not decreasing...
Just here to help people realize one must be careful with accelerant and ignition
It became relevant immediately, part or not
Given this is Proxmox, I seriously hope this isn't for end users. It's fine for certain uses but I'd trust my business on straight kvm and libvirt (plus lxc or whatever else) before I added that
It's a personal distrust of management layers though
If you sell a VPS with x GB RAM then the customer pays for and has the right to have x GB of RAM to be available at any time and always, period.
Because it's a bad mix of cowardice and scam. If you want to optimize your profit by playing with non existing RAM then say so. Just sell and say "1 GB RAM fix, 4 GB RAM if and when you are lucky. No guarantees whatsoever". But of course that doesn't sell well. What you are actually meaning to ask is "how can I scam my customers?".
That's not what all this is about. It's about me trying to get the most out of the resources my node has. Example node has 256GB RAM running 20 virtual machines each with 24GB each but none of them are currently using more than 5GB why should I be ok with the real usage being <100GB the likelihood of all of them actually using 100% RAM at the same time is very slim.
I don't see it being a scam as the server would never be overloaded to the point of a customer not being able to use 100% of RAM if needed unlike what currently seems to be acceptable with CPU power where you should only expect to be able to use 30-50% of the power 24/7 according to other users on this forum.
On the topic of other providers doing this over provisioning I can see Contabo does this too
root@vmixxxxx:~# dmesg | grep -i balloonroot@vmixxxxx:~# lsmod | grep balloonroot@vmixxxxx:~# cat /boot/config-`uname -r` |grep -i balloonCONFIG_MEMORY_BALLOON=yCONFIG_BALLOON_COMPACTION=yCONFIG_VMWARE_BALLOON=mCONFIG_VIRTIO_BALLOON=yCONFIG_HYPERV_BALLOON=mCONFIG_XEN_BALLOON=yCONFIG_XEN_BALLOON_MEMORY_HOTPLUG=yCONFIG_XEN_BALLOON_MEMORY_HOTPLUG_LIMIT=512They also drop cache hourly
root@vmixxxxxx:/etc# grep -i "drop_caches" * -Rcron.d/sync:12 * * * * root sync; echo 3 > /proc/sys/vm/drop_cachesI need to try and replicate what they are doing maybe as mine only shows
cat /boot/config-`uname -r` |grep -i balloonCONFIG_MEMORY_BALLOON=yCONFIG_BALLOON_COMPACTION=yCONFIG_VMWARE_BALLOON=mCONFIG_VIRTIO_BALLOON=mCONFIG_HYPERV_BALLOON=mCONFIG_XEN_BALLOON=y# CONFIG_XEN_BALLOON_MEMORY_HOTPLUG is not setThink this will help? @Falzo
Xen is the right way for memory balooning.
And why should customers be OK with you betraying them?
The context is not "how to make Harmony feeling fine", the context is business, it is that you sell something you potentially don't have. When a customer buys a VPS from you he is under no obligation whatsoever to use his RAM in any particular way or pattern.
If I use say 1 GB of my 4 GB RAM all the time for 2 years and suddenly for no particular reason I feel like using the full 4 GB RAM they must be available and immediately so (e.g. without swapping something out so that I can have my 4 GB). Period.
And btw, it's none of your business in the first place whether and how I use my RAM. Imagine a landlord asking the price for 100 m² but providing only 30 m² explaining that you'll get a larger part of what you pay for when he thinks you need it!
If you don't like that then make it perfectly clear right in the offer/product description that a given product does not have the announced full RAM available and that e.g. your "4 GB mem. VPS" actually is a 1 GB mem. VPS.
Well I can assure you that I wont be overselling to the point where anyone can complain about not being able to pull 100% of their RAM. Do you have an issue with providers selling disk space that is thin provisioned? You don't have that space reserved on the disk so it's not dedicated it grows and shrinks as your data does but it's up to the provider not to oversell it too much. If I am operating multiple nodes it's an easy live migration to swap a customer to a new node with less RAM I just don't see RAM reaching 100% unless I do a really bad job at managing them.
You are also way more likely to get a VPS that offers x TB at 1Gbps and be fighting with another user for that 1Gbps speed than fighting for RAM on an oversold node IMO.
Are you going to add "Fair Share" next to your memory/ram specification, in clear human readable language and not hidden in tos? It the answer is yes, then to have the right to do what you are aiming to do.
If not, you will be false advertising, because by default, the assumption is, if I am paying for it, it is mine.
You see, low-end talk is a consumer forum and not a business one. So when most people talks, they will always take consumer side and not the business, unless consumer is completely wrong/off.
Road has 25mph speed limit. That's the max you can use. But, it is shared. This is common. This, during rush hour, you will expect slower speed, due to traffic.
If I am renting a parking spot on the other hand, I expect that spot to be available, whenever I want to. I can park a truck there or my hamsters bicycle. This is my spot.
Your question is more suitable for wht and you will not only be looked at as "that guy" and be trolled by many.
It is a good technical question, but a horrible question to ask in a forum where your customers shop.
Thanks for the heads up but I got some good help here. I just expected more from a budget providing community that I would have thought likes to squeeze as much as possible from their resources (within reason). Maybe it's something providers like to hide.
Swapping means making it available and it takes time. If I buy a VPS with x GB RAM I expect x GB RAM to be available as opposed to "made available".
Note: SSD are orders of magnitude slower than even cheap RAM.
I think it's OK to sell VPS with some fixed (always available) plus some extra RAM that is not always available but is swapped - but then you must say that and clearly state "x GB RAM plus y GB swapped RAM".
But that's not what you want. You want to sell a cheap mix of real RAM and swapped RAM but just call it "RAM" which I consider fraud.
I think you still didn't get the message.
But now that he did, the good guys that make all sort of benchmark scripts could implement a test for this sort of thing right? Like @MasonR 's famous YABS.
If you are selling dedicated Ram at a budget price, yes it is a good deal. You are not. It is not a good deal to get 1 GB of usable ram and 4GB of burst memory for $2/month. So yes, the community expects good deals. Resource that they can use. Not the idea of "good deal".
Then again, we have seen providers selling 9x i7-7700 vCores. Granted, they are no longer in business and that should be your hint.
Summer brings all types of entertaining stuff in this forum.
A decent amount of providers in this forum, that are popular and fan favorite, don't oversell memory.
Folks who are successful, such as BuyVM are transparent, their product page will tell you what is dedicated and what is shared.
There are only "handful" of providers who oversells their Ram and disk and usually this community and other internet community will tell you that, "you get what you pay for". But at the same time, those aren't your small one man show but folks with seed money and massive financial backing, with dedicated team and automation, that would mitigate issues, before they become an issue.
Price your product right or market "shared memory". Or better yet, do what intel did, make your own "fancy term" for SSD/nvme swap and call it whatever you want to.
There is a reason why this community and many other looks down on OpenVZ providers.
Might not be possible, unless you have something running for an extended period of time. Not couple of seconds or minutes. Maybe @jsg can implement something like this with his benchmark script, that runs randomly at random burst.
Yes, that ... plus some more factors, but ...
My benchmark already has the needed capabilities but I of course respect that another benchmark author was addressed. Maybe that author finds a way, I don't know, my bash knowledge is rather limited because I don't use it. [polite smile]
@jsg
That would be great to have! I simply mentioned MasonR as his script can be used freely by anyone whenever they get a new machine. I know you don't publish your benchmark software which is understood but limits the ability for anyone to run it on their own machines. Perhaps you could provide a small distributed utility that anyone can run that answers the question: Dedicated RAM? Yes/No? even just binaries for it would be fine.
The way I see it, more people doing the same thing is a plus, not a bad thing. Community would then have options.
A distributed utility would be more useful because then any of SSDNodes victims can test it on their servers and anyone trying to use a new provider, can run it too on their own because not all providers are going to ask you to perform a benchmark or provide the required credits/resources.
I do make my software available for free, just not the source code (anymore). Unfortunately though I have been, to put it very mildly, questioned why I (dare to) publish my software via Yandex and not github.
But again: I did make it available.
@jsg if distributing it is the issue, what site to put it on, that can be fixed easily right? Even Fosshub accepts Freeware I think: https://www.fosshub.com/signup.html
Freeware has been there from the start it's not something new. FOSS is awesome of course but sometimes you gotta do what you gotta do. Up to the author.
Thanks but I try to keep the number of accounts I have within reasonable limits and don't like to create a new account for providing my work for free when I have a perfectly fine and working account already (with yandex) just because some people don't say "thank you" but complain about it not being [on whatever they consider adequate e.g. github]
As for "open source" I did provide the source of v.1 ... and virtually nobody downloaded it. It was just another case of demanding (me) to conform to some kind of la mode du jour ("we want source code!!!" and "we need to be able to check the source to trust it!!!") which turned out to be just virtue signalling. So I stopped playing their stupid games.
As I said it's available on yandex (for free). Feel free to PM me if you want the most current version. I have no problem at all to provide my benchmark software to people who are really interested (as opposed to virtue signalling, uttering demands, and blabla). In fact I have even created special versions (e.g. musl lib) for some really interested people.
tbh no idea... haven't played around with specifics on that one.
that's still going to be a tough call as it is about 100% overcomitting. (keep in mind that the ARC cache is definitely not avail).
I would most likely not try that, but if, then set the min per VM to 12GB and have 200GB swap available.
watch the memory on the host node (and not the reserved per instance). if it reaches 90-95% usage that's totally fine. you want your memory to be used, even if it's just filecache inside the guests.
also remember ballooning does not reclaim memory because it can, but only if need be.
and as long as the system does not start swapping in and out like crazy there simply is no need and everything should be fine and the hypervisor does exactly it's job with all the available features...
swap can grow over time still, as old pages go out, but that's also fine. swap isn't an enemy. but yeah, just my 2cents. interesting topic after all.
Some of you in this topic are really acting like Contabo isn't a thing.
Hello, McFly?!? Of course we expect it to be dedicated. Most providers use a predictable and scalable pricing and resource structure to limit how much the CPU gets shared. Total node RAM in a ratio with the CPU cores.
So in your example of 20 clients with 24GB of RAM, they get 1/20th of the CPU. Every additional VM dilutes their CPU resources. They'll notice that much more and easier than the odd times RAM is maxed.
Who the fuck wants 24GB RAM VM's with 2 or less cores? Seriously, who and why?
A quick tip: when basically everyone in a community that really tries to do as much as possible within reason recommends against doing something, then whatever you are fixing to do is most likely not within reason.