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
bad news for reward seekers or miners
Glad I keep my old rs series and haven’t upgrade them…
The answer is yes, I know it from a trusted source haha.
It might not be that unrealistic after all. If the drop actually comes from some EPYC CPU downclocking after reaching an overall load of about 50% there might not be a lot of steal because resources that don't exist in the first place can't be stolen. At least that would be the logical thing in such a situation. I have no idea how steal is actually calculated in reality though.
This is because virtualisation meanwhile is so much more than just splitting some ressources.
Think of this simplified example:
There is a host system with 100 cores. From this 50 VMs are sold with 4 cores. Everything is equally set in the config of the hypervisor these guests.
Now, as long as the overall cpu usage is below a certain level, lets use 80% as example, you could start a benchmark on a single machine and will reach easily a high single core score as well as multi core. Simply because there is enough room on the host to dedicate these ressources to you while you are running the benchmark.
If however more than 50% of the guests are running something on 100% all the time obviously the hypervisor will try to balance and your VM will only get a certain share. As the system is overbooked by 100% in terms of cores, you will max get 50% of what you'd expect and will see a bad single core score as well as steal, because the hypervisor will tell the guest, that it requests more than what can be delivered.
Now let's bring in the option of limiting the ressources of a guest behind the scenes slighty differently. With KVM for instance you can give a VM 4 cores but limit the max usage to 2 actual cores.
Now same as above, the hypervisor runs at 80% only and you start a benchmark. You will still get the high single core score, as you are allowed to use it. The multi core probably will be a bit worse, but most likely you wouldn't even care or notice. Scaling for single->multi never is linear anyway.
If again all VMs are set up the same way and again more than 50% are going all in, they would neverthelesse be limited to a usage of 2 cores max and leave room for the other guests. Even if nearly all VMs run at 100% you'd still have your own two cores as it essentially is not overbooked.
In a benchmark you now will see a not so good single core speed, as you probably can't burst anymore and multi core will also reflect this. However, there will be not much steal, as the hypervisor is able to handle everything still within the limit set.
Please keep in mind this is a simplified example to just express that there is many more options to setup/balance/optimise a system. It is VIRTUAL after all. Hell, you can show whatever CPU string you want the client believe they are on.
Notice something
Given supposed dedicated resources there always should be enough room to give you the resources you booked.
At least in the usual KVM setup i kind of doubt this. Individual Qemu processes know nothing about each other, let alone have some kind of communication system. Any kind of balancing is down to the task scheduler.
Why complicate things even further though?
No offense but as far as i can see your post is essential a lot of text and none of it explains how steal is calculated on a non-overbooked system when the CPU throttles for whatever reason.
I agree with you that in a non-overbooked host, steal is theoretically unrelated to CPU throttling alone.
But practically it goes like: CPU throttles because there is high load, and since there is high load, scheduler may take longer to perform scheduling operations, especially in situations where they are all waiting shared resources, like memory. This results in cpu_steal% inside VMs, even the hypervisor has its own core.
Your cores have been hexadecupled.
Yeah, i guess in the end steal means something like timeslices that the system unexpectedly wasn't able to provide regardless of the underlying cause. I probably don't know enough about how multitasking is actually implemented to envision what that looks like in the field though. The scheduler interrupting a process in the middle of a MOV operation and not switching back to it just because the memory bus didn't come through fast enough seems kinda weird but then i probably don't have the most realistic idea how that looks like in general.
Netcup, probably: ”Your cores have been crippled.”
GB5 score for €0.84/month VPS
I’m running close to 3000 servers on Hetzner Cloud which are all running blockchain nodes and I’ve never had any issues. I suggest them. They certainly don’t have any supply issues.
Hetzner says they don’t allow crypto related stuff but I’m running these kind of nodes for the last 6 years and they never told me I can’t. Maybe because I’m running thousands of them. They even gave me a 3% discount on my invoices for paying over €25K/year and with bank transfer.
Lmao, this info is probably enough for them to find your account now.
Nobody tag them!!
They certainly would’ve noticed someone running 3000 nodes for the last 6 years by now and maybe ban the account if they didn’t allow. 😄
Or because your nodes are just being nodes and not trying to mine coins/gain rewards by maxing out everything 24/7. I don't think a mere crypto node would pose much of a problem to just about any provider.
They’re are earning rewards since they’re staking but you’re right about the load. They’re not maxing out. They’re not mining.
I know other people running thousands of similar nodes in Hetzner Cloud and Contabo as well.
Contabo is also offering crypto staking node services and they’re running over 15,000 Horizen nodes for instance.
Maybe not maxing out, but I'd imagine Hetzner is more okay with high sustained load by some users compared to smaller low-end hosts, as they have such a big amount of very powerful hypervisors that it only really becomes a problem if the "high resource users" reach a certain % of total users, a % that this guy's 3 000 servers likely dosen't come close to.
They can probably load balance away lots of issues.
Yes, but there's tons of complains on Contabo's performance.
Indeed. Personally I’ve never used Contabo and I would never use them unless I absolutely have to. For now I’m fine with Hetzner.
Yeah, hosting on Contabo is just sad