Dedicated Server vs. VPS
I have been thinking a lot recently about whether it makes more sense to have one large dedicated server with Proxmox and pfsense, etc. or if it's easier to just have VPS's across providers. What do you think?
Dedicated Server
Pros:
- Full Control
- Ability to run pfsense, control routing, etc.
- Ability to setup resources the way you want (vCPU, RAM, etc.)
- All VM's sit behind a FW
- Private Networking without VPN
- Most cost effective
Cons:
- You have to manage the Hypervisor - Updates, etc.
- You have to manage the networking.
- If the server goes down, everything does.
VPS
Pros:
- No hypervisor to manage
- Networking is general easier.
- Not dependent on one provider.
Cons:
- Need VPN for private network
- Increased Cost
- Less control
Comments
Depends what you're hosting I guess, and how much of an issue it is if it "all" goes down in one go.
There's too many "it depends" in this. If you want to have edge nodes close to end users on different continents, well consolidating onto a single dedi works against that. When I run the $, a bunch of VPS are still cheaper for me.
this. If you think there is a hassle managing multiple VPSs, there are a lot of automation scripts to help.
However, a large number of providers would throw in BGP if you get a dedi. Not many VPS providers do.
The all seeing eye sees everything...
I am in same boat for many days and thinking a lot about it.
and one reason i still staying with VPSes is thinking the maintenance part of dedi.
and other points:
biggest point for me is...you get dedicated resources and you can utilize them however you want and when i calculate, i am paying more for my VPSes than a single AX41 and resources wise AX41 provides more.
overselling/overcommitting has become a big issue... in many VPSes i see most of the ram is under buffer... i guess that means ballooning is going on.
ThemeForest Profile
It would depend on what your ideal situation is to determine if you need a virtual or dedicated server. Often times, VPS will be cheaper, but you take some compromises... such as (but not limited to):
Flip side to the dedicated aspect, sure you control the resources, networking, etc but if the dedicated in mention is that of a colocation (eg, you own the hardware), you can often get "beast" machines for a fraction of their rental price tag. To represent this, take this example specification:
If I had to price this machine for rental, we have a pretty simple formula for that: (128 * 0.70) = (89.6 + (transit_bandwidth / cost)) + profit... so, that machine in Toronto, we could lease for at least $89.60 monthly without transit, adding in let's say $150 for 150TB on 10GbE. That means it's a $239.60/mo machine.
Flip that to a colocation situation, you eat up-front costs to have lower MRC. To build that exact box described, the hardware (2xE5-2650v2, NIC, 4x2TB HDD) in an HP DL380P Generation 8 format would run around $350 for the machine with the NIC, let's assume it's $58.56 for 2TB drives (each), that means we're at $234.24 for four brand new 2TB Seagate drives, and the memory can be had for around $100 or so. In total, we're at $234.24 + $100 + $58.56 = $392.80 in upfront cost (this excludes shipping). You can then normally find colocation for 1u devices at around $50-100/month (eg, Dallas I see commonly $50/month with power, transit for 1U).
That being said, if you have the funds to build and colocate, I suggest it. Over the long run, you make the ROI fairly quick (depending on config, this can be from 1mo to 6mo to 1yr).
For me, I just use dedicated to host machines for VPS containers. For me, it does not make sense to run one application on a single dedicated server.
Vladimir S. - IT Consultant, Entrepreneur, Web Developer, and Disability Advocate.
[email protected] | www.ivlad.net
My pipe dream is to get a fiber connection to the closest dark fiber junction box.
Not gonna happen soon unless my neighbors want to chip in for a WISP.
The all seeing eye sees everything...
Neither. Get a lot of suitably spec'd cheap KVM, install proxmox on them all, enjoy best of both worlds.
Get the best deal on your next VPS or Shared/Reseller hosting from RacknerdTracker.com - The original aff garden.
Where do you look to see these offers? Thanks!
I hope everyone gets the servers they want!
Disagree with the notion that VPS is more expensive than dedicated.
On the face of it sure - on a per core / per unit of resources.
...but that all goes out the window onces sharing kicks in. VPS model allows for significant overcommit...while simultaneously allowing for individual users to occasionally draw not just on their "fair share" but temporarily go beyond that. A bit like a uber car will be on the road more hours per day than your owned car sitting in the garage all day...yet everyone still gets to where they need to go. If you're measuring that by how many cars/cores I've got then maybe that's the wrong question...cost per journey is where it's at.
Unless usage case requires sustained use or is very vulnerable to noisy neighbours I don't see any dedicated model beating that.
Wouldn't necessarily host critical stuff on a VPS, but still I think it's one of the best options in terms of price.
Seller might abuse this by overselling...but in principle sharing is the way. Especially given timezones, idling etc.
On VPS:
High CPU, high IO, high load => suspend.
I want to encode 1-minute of push-up videos, once a week. It would be a few minutes of high CPU on Ryzens.
However, I'm too scared to do that on a "fair share" VPS. I can't afford a dedicated machine so I have to run the encode process on Oracle Cloud "1/8 OCPU" that takes an hour per minute of content.
I want minutely VPS. Major cloud providers offer hourly billing, but I need less than an hour.
Accepting submissions for IPv6 less than /64 Hall of Incompetence.
Aside from the unaffordable price of my VPSes and the grumpy, incompetent, clueless administrator, is there some reason why your few minutes per week couldn't be run on my server?
Maybe it's actually a large part of your fun to do it on Oracle's fraction of a CPU?
Greetings from Mexico! ?
I hope everyone gets the servers they want!
I have a different idea.
These non-time-sensitive, compute-heavy jobs, such as video encoding or machine learning, are more suitable in a batch system.
Customers post their jobs to a queue: program (Docker container), command, input files, hardware requirements (CPU instruction set, CPU speed, RAM, SSD space for intermediate and output files, GPU requirements, etc), time limit (job will be terminated but customer still pays if time limit is exceeded)
Providers bid on pending job: "I will run this job for xx Satoshis"
This allows customers to run a compute job at minimum cost, while providers can monetize spare resources on their servers.
Compare to AWS "spot instances", the job will run to finish once accepted, and will not be terminated in the middle.
Accepting submissions for IPv6 less than /64 Hall of Incompetence.
The clearing houses would take as commission 6% of the value of all compute transactions. Imagine being the guy who sets this up at the precise moment when the time is ripe!
These days it is a royal PITA to have to choose between dedicated and cloud. Dedicated is too expensive and cloud is too complex and too slow (at levels which cost less than dedicated).
@aaronstuder is asking in this thread a most significant question! The reason why I am selling excellent VPSes is hopefully to provide an escape route for a few people trapped in the Grand Canyon between expensive dedicated on one side and complex, slow cloud on the other side.
@yoursunny does have a better idea than mine. So, okay, let's see how it goes! I will be the clearing house for a few test jobs unless @yoursunny or someone else wants to do it. I will bid some price for use of a few cores, a little RAM, and some disk space for a limited time on a few jobs.
So now I am waiting for (1) @yoursunny or someone else to step up to serve as clearing house or else I will serve as clearing house, and (2) people to submit requests for bids.
Greetings! ??
I hope everyone gets the servers they want!
SLURM is widely used in supercomputing to manage job queues. You have to figure out the bidding and micropayments side of things.
There is a vulnerability when the same person own the clearinghouse and a provider:
Another issue is, it's difficult to ascertain to correctness of computation. Deterministic computation can be validated via occasional recomputing, after which dishonest providers would be discredited; SETI@Home and the likes are sending the same job to multiple computers and compare the outputs.
Indeterministic computation is more difficult to validate. What stops a provider from running the computation on a less powerful CPU core than what the customer requested (e.g. encode the video at 82% quality instead of 85%), and upload a slightly inferior output file and falsified logs? Customer might be able to detect and complain, but it's hard to mechanically prove.
Accepting submissions for IPv6 less than /64 Hall of Incompetence.
Well I've got a gigabit fiber line here at home. Would definitely host CPU/storage/mem heavy backend on it, but def nothing public facing lol.
One of the major mid tiers announced that recently. Forgot who..ovh I think.
I doubt a provider would suspend you for spiking the cpu once a week though?
Good to know. It's Google Compute Engine offering per-second billing.
Their egress traffic is pretty expensive though.
Virmach would definitely suspend if you exceed 33% for more than 10 minutes.
Clueless Administrator LLC would probably cut some slacks.
The point is, why should a customer pay for "fair share" resources for the whole week/month, if what they really need is dedicated resources for 30 minutes every month?
Also, if several customers decide to burst to 4 full cores at the same time, everyone's job slows down. A batch system could solve this problem.
Accepting submissions for IPv6 less than /64 Hall of Incompetence.
FWIW, the ballers can idle dedis these days
Shitty Comcast don't even offer fiber in my neighborhood even though the junction box is within one block from where my house is.
To be fair, they might not own that fiber.
The all seeing eye sees everything...
Thanks @yoursunny for mentioning SLURM! I had not heard of it. I skimmed through the SLURM FAQ and the Quick Start Administrator Guide. I am guessing I could compile SLURM. During my quick visit to the SLURM website I didn't catch any hints for the bidding and micropayments parts, which also would be interesting.
Friendly greetings from the city and from the desert! ??️
I hope everyone gets the servers they want!
To be honest people in that troubled scenario where they need dedicated resources but can't afford a Bare metal most likely don't know about VM's with dedicated resources. Probably they want to pay $2 a month and run Facebook on that machine. If you are pulling that much resources then you are making money of it. If is not the case then you are wasting time on shit and giggles. That being said... there is **almost no reason nowadays to have VM's **unless you want to do hourly billing , spin up , process, destroy. Dedicated servers are way way way too affordable in 2020, you can get an E3 12xx V2/V3 all to yourself , full cores, bare metal, for 40ish /mo on a decent proven DC, even less in some providers. You can even split that E3 into many VM's and get most of it and reduce your expenses. Don't believe the Ryzen / AMD wagon , old E3's dedicated servers will still outperform any cheapo Ryzen VM you are paying. Remember dedicated cores on most providers is not true dedicated cores, it means dedicated thread, and that's not a true core. So while you pay for an expensive VM with 4 cores, in reality you are paying for 4 threads, dedicated or not, still half a core. Can't really compare that to having a full processor (E3 for example) to you, full cores, full threads, nothing shared. IMO if you have the knowledge, just go with a dedicated server.
Also in some cases premium shared hosting might outperform most cheap VM's you can buy around.
By the way Ryzen CPU's are NOT even SERVER GRADE. EPYC is what you really want. I do have lot's of respect for providers that innovated the market with Ryzen but only those that were the pioneer. Not all ryzen CPU's are worthy of such inclusion , I see people around buying low end Ryzen's and putting it on desktop grade boards, thats not how this works. Literally cheapo providers kind of ruined the vibes of Ryzen. Ryzen != Server grade.
I spend less than $20 a month for 5-8 VPS.
I must be cheap.
The all seeing eye sees everything...
I'm not making money. F-1 and J-1 visa holders are not allowed to make money.
I build websites and apps for giggles and LOLz.
That's true. I built my own virtual machine management software before Vagrant becomes popular.
A thread is worse than a full core. If the vCPU is moving between physical cores, you are losing performance in CPU cache misses.
Shared hosting is useless because my apps don't run on HTTP.
Truth. Real dedicated server uses Cascade Lake and EPYC. However, they require parallelizable workloads because single core frequency is worse than Ryzen.
YABS single core score:
I spend less than $80 a year for 6-9 VPS.
I must be broke.
Accepting submissions for IPv6 less than /64 Hall of Incompetence.
Virtual Machine Server
>
>
True, but there are some Intel server CPUs which are potentially faster than Ryzen - E-2278 YABS 1417 single core (and the 2276 is clocked at 3.8ghz).