NetCup Performance Issue

itsdeadjim

@labze said:
Dedicated cores aren't guaranteed performance. People need to understand this before getting out the pitchforks.

Even a server only loaded with only 50% usage can see over a 50% reduction in performance. Some CPUs do better, some do worse. And no, CPU pinning won't change this fact.

Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

I dont know how many times this has be written before someone will open a new thread complaining for false ad (not talking for OP by the way)

labze

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

emgh

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

Yeah, begs the question what the point of the word even is to be fair. Only way to know really is to try. Unless you know exactly how the provider pack their nodes, as well as how strictly they enforce their FUP.

lirrr

i dont care about perf drop
i only care about if cpu steal > 3%

emgh

@lirrr said:
i dont care about perf drop
i only care about if cpu steal > 3%

i dont care about perf drop
i only care about yabs

yea i m a let user how did u know

Moopah

@emgh said:

@lirrr said:
i dont care about perf drop
i only care about if cpu steal > 3%

i dont care about perf drop
i only care about yabs

yea i m a let user how did u know

Wait are you me??

labze

@emgh said:

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

Yeah, begs the question what the point of the word even is to be fair. Only way to know really is to try. Unless you know exactly how the provider pack their nodes, as well as how strictly they enforce their FUP.

My understanding of dedicated cores is that there are no FUP. You can utilize your cores 100% all the time. But if everyone does that on a node, then you'll have a bad time dedicated or not. Netcup is likely suffering the same issue as many other providers seem to. These mining people buy large amount of VPS and deplete all available resources and then go complain that the performance sucks. I am not exactly sure what kind of workload this is, as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

lirrr

@emgh said:

@lirrr said:
i dont care about perf drop
i only care about if cpu steal > 3%

i dont care about perf drop
i only care about yabs

yea i m a let user how did u know

nowhere I saw people request yab except LET and LES (or nodeseek if you count that Chinese version of yab )

lirrr

@Moopah said:

@emgh said:

@lirrr said:
i dont care about perf drop
i only care about if cpu steal > 3%

i dont care about perf drop
i only care about yabs

yea i m a let user how did u know

Wait are you me??

u r chicken and I love cheap and fast chicken
i guess im u

Calin

@labze said: as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

20x VPS with crypto mining

emgh

@labze said:

@emgh said:

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

Yeah, begs the question what the point of the word even is to be fair. Only way to know really is to try. Unless you know exactly how the provider pack their nodes, as well as how strictly they enforce their FUP.

My understanding of dedicated cores is that there are no FUP. You can utilize your cores 100% all the time. But if everyone does that on a node, then you'll have a bad time dedicated or not. Netcup is likely suffering the same issue as many other providers seem to. These mining people buy large amount of VPS and deplete all available resources and then go complain that the performance sucks. I am not exactly sure what kind of workload this is, as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

Yeah, but that makes the word a bit meaningless. Like right now, when this mining thing is booming, providers with shared cores might actually on average perform BETTER than those with "dedicated cores" since they're a lot less likely to enforce a CPU limit on abusers, since it's then not "dedicated" anymore.

JabJab

@labze said: I am not exactly sure what kind of workload this is, as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

@CherryServers does your dedicated (/pinned/exclusive/expensive) VDS suffers same problem?

labze

@Calin said:

@labze said: as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

20x VPS with crypto mining

There seems to be several different mining operations going on. CPU Usage was low, around 30% in my scenario but the overall load exceeded 100. I do not know what specifically caused this load to happen as IO/CPU/Network was fine, but something caused load spikes.

totally_not_banned

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

In some ways it probably will, in others it probably won't as the scheduler obviously treats the whole thing as one big server, so what's optimal for one VM might not be for another. Besides the scheduler has quite a couple different policies and knobs to tune it, so just assuming it automagically makes the best decision by default seems a bit brave.

totally_not_banned

@emgh said:

@labze said:

@emgh said:

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

Yeah, begs the question what the point of the word even is to be fair. Only way to know really is to try. Unless you know exactly how the provider pack their nodes, as well as how strictly they enforce their FUP.

My understanding of dedicated cores is that there are no FUP. You can utilize your cores 100% all the time. But if everyone does that on a node, then you'll have a bad time dedicated or not. Netcup is likely suffering the same issue as many other providers seem to. These mining people buy large amount of VPS and deplete all available resources and then go complain that the performance sucks. I am not exactly sure what kind of workload this is, as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

Yeah, but that makes the word a bit meaningless. Like right now, when this mining thing is booming, providers with shared cores might actually on average perform BETTER than those with "dedicated cores" since they're a lot less likely to enforce a CPU limit on abusers, since it's then not "dedicated" anymore.

I'd be curious how many providers take the dedicated thing somewhat serious to begin with. Like pinning only pairs of threads from the same core to VMs (or disabling HT - yeah probably not going to happen but theoretically). Sure if the node CPU throttles when used beyond a certain percentage (due to power limits, thermal throttle or whatever) VMs could still be affected by load outside their containers but if that's taken out of the equasion (together with IO, busses and such) i'd be really curious how that would perform.

My guess would be that most providers don't really put much thought into this at all and just make a calculation like "OK, so i have 32 threads, that means i can sell 16 2x dedicated vCore VPS" with no pinning, special config or resources set aside for the hypervisor at all.

Moopah

@totally_not_banned said:

@emgh said:

@labze said:

@emgh said:

@labze said:

@emgh said:

@labze said: Virtualization, RAID, Networking, CPU scheduling and so forth, all of this requires CPU usage and even a significant amount on large nodes.

By cores, do you mean physical cores or logical cores? I guess that would change to which extent other guests compete for resources, although it will have an impact anyway, of course.

It's both. The CPU scheduler picks whatever is most optimal.

Yeah, begs the question what the point of the word even is to be fair. Only way to know really is to try. Unless you know exactly how the provider pack their nodes, as well as how strictly they enforce their FUP.

My understanding of dedicated cores is that there are no FUP. You can utilize your cores 100% all the time. But if everyone does that on a node, then you'll have a bad time dedicated or not. Netcup is likely suffering the same issue as many other providers seem to. These mining people buy large amount of VPS and deplete all available resources and then go complain that the performance sucks. I am not exactly sure what kind of workload this is, as I've had a client bring a node to its knees with these activites, even throttling the CPU to 15% and IO to 15MB/s / 2500 IOPS caused load of above 100 on a 7950XD server. It frankly is abusive behaviour on any shared system, dedicated or not.

Yeah, but that makes the word a bit meaningless. Like right now, when this mining thing is booming, providers with shared cores might actually on average perform BETTER than those with "dedicated cores" since they're a lot less likely to enforce a CPU limit on abusers, since it's then not "dedicated" anymore.

I'd be curious how many providers take the dedicated thing somewhat serious to begin with. Like pinning only pairs of threads from the same core to VMs (or disabling HT - yeah probably not going to happen but theoretically). Sure if the node CPU throttles when used beyond a certain percentage (due to power limits, thermal throttle or whatever) VMs could still be affected by load outside their containers but if that's taken out of the equasion (together with IO, busses and such) i'd be really curious how that would perform.

My guess would be that most providers don't really put much thought into this at all.

I specifically pin threads within the same NUMA node or CCX group when creating VMs to mitigate some potential poor process scheduling

itsdeadjim

@Moopah said: I specifically pin threads within the same NUMA node or CCX group when creating VMs to mitigate some potential poor process scheduling

I think most of the times the real benefit with pinning is avoiding cache invalidation.

totally_not_banned

@itsdeadjim said:

@Moopah said: I specifically pin threads within the same NUMA node or CCX group when creating VMs to mitigate some potential poor process scheduling

I think most of the times the real benefit with pinning is avoiding cache invalidation.

Good point. RAM is slow and cache is precious.

dev_vps

@itsdeadjim said:
The only issue I see here is that they dont ban miners.

Other than that, where is that @online7237 dumbhead?

encashing his rewards check as soon as bank opens, just to make sure check is actually dedicated to his account or a shared resource among multiple customers

totally_not_banned

@dev_vps said:

@itsdeadjim said:
The only issue I see here is that they dont ban miners.

Other than that, where is that @online7237 dumbhead?

encashing his rewards check as soon as bank opens, just to make sure check is actually dedicated to his account or a shared resource among multiple customers

Why stop at shared checks? Shared bank accounts are the future

dev_vps

@labze said:

Dedicated cores aren't guaranteed performance.

basically it just means, customer can use cpu resources as much they want

for most VPS, there is 10%, 20%, 30% average cpu resource usage limit over x number of hours

Razza

I won't argue about dedicated cores not being 100% available at full speed 24/7.

Take Netcup's RS 1000 G11, for example. For €9.81 a month, you get 4 dedicated AMD EPYC™ 9634 cores, 8GB RAM, and 256GB NVMe storage. At that price, it's unrealistic to expect maxed-out cores 24/7 without any throttling. This pricing is comparable to other reputable providers offering shared vCPUs.

Most users don't need constant 100% CPU usage, so this model usually works fine. Even if you do need constant usage, the host node typically has spare CPU capacity.

The problem arises when too many user VMs on the node constantly use 100% CPU, leading to performance drops due to scarce CPU resources.

If you want guaranteed 100% dedicated CPU cores, you will need to raise your budget.

Moopah

@totally_not_banned said:

@dev_vps said:

@itsdeadjim said:
The only issue I see here is that they dont ban miners.

Other than that, where is that @online7237 dumbhead?

encashing his rewards check as soon as bank opens, just to make sure check is actually dedicated to his account or a shared resource among multiple customers

Why stop at shared checks? Shared bank accounts are the future

Shared LET accounts too

totally_not_banned

@Moopah said:

@totally_not_banned said:

@dev_vps said:

@itsdeadjim said:
The only issue I see here is that they dont ban miners.

Other than that, where is that @online7237 dumbhead?

encashing his rewards check as soon as bank opens, just to make sure check is actually dedicated to his account or a shared resource among multiple customers

Why stop at shared checks? Shared bank accounts are the future

Shared LET accounts too

Right on brother! I'd let you shitpost on my account any day.

vpn2024

@Razza said:
I won't argue about dedicated cores not being 100% available at full speed 24/7.

Take Netcup's RS 1000 G11, for example. For €9.81 a month, you get 4 dedicated AMD EPYC™ 9634 cores, 8GB RAM, and 256GB NVMe storage. At that price, it's unrealistic to expect maxed-out cores 24/7 without any throttling. This pricing is comparable to other reputable providers offering shared vCPUs.

Most users don't need constant 100% CPU usage, so this model usually works fine. Even if you do need constant usage, the host node typically has spare CPU capacity.

The problem arises when too many user VMs on the node constantly use 100% CPU, leading to performance drops due to scarce CPU resources.

If you want guaranteed 100% dedicated CPU cores, you will need to raise your budget.

Or just drop the word dedicated from the marketing/pitch, it's not dedicated... much simpler solution.

dev_vps

@vpn2024 said:

@Razza said:
I won't argue about dedicated cores not being 100% available at full speed 24/7.

Take Netcup's RS 1000 G11, for example. For €9.81 a month, you get 4 dedicated AMD EPYC™ 9634 cores, 8GB RAM, and 256GB NVMe storage. At that price, it's unrealistic to expect maxed-out cores 24/7 without any throttling. This pricing is comparable to other reputable providers offering shared vCPUs.

Most users don't need constant 100% CPU usage, so this model usually works fine. Even if you do need constant usage, the host node typically has spare CPU capacity.

The problem arises when too many user VMs on the node constantly use 100% CPU, leading to performance drops due to scarce CPU resources.

If you want guaranteed 100% dedicated CPU cores, you will need to raise your budget.

Or just drop the word dedicated from the marketing/pitch, it's not dedicated...

unmetered usage cores

emgh

Lmao, unmetered cores

Razza

@vpn2024 said:
Or just drop the word dedicated from the marketing/pitch, it's not dedicated... much simpler solution.

That's probably the best option going forward drop the dedicated wording, not sure what to word use to describe shared cores that allow higher utilization than normal shared cores.

emgh

@Razza said: That's probably the best option going forward drop the dedicated wording, not sure what to word use to describe shared cores that allow higher utilization than normal shared cores.

UNMETERED CORES

emgh

SORRY, UNMETERED THREADS!!!!