All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Hetzner auctions vs Netcup Root-Server
Hi guys,
you can read on a regular basis people recommending Hetzner's dedis from the auction and I get the bright side of those servers as they are quite cheap (start at 23€) for their performance.
However, I just saw the latest Netcup Root-Server portfolio and there was the RS4000 G8 with 32GB Ram, 10 ded. cores and almost 1TB Hdd for 25€.
I am wondering now: Why should I go with the Hetzner server that offers way less cores (Ram and storage are identical usually) instead of the Netcup RS?
I would expect far better performance from the RS compared to the Hetzner server given that it is 10 vs. 4 cores.
I'm afraid I am missing something over here. You guys might enlighten me.
Comments
Hetzner has 100% of the port speed dedicated I think.
IIRC the bandwith is way more restricted over at Hetzner than Netcup (30 vs. 80TB). I would assume if someone really needs a dedicated 1Gbps port you push way more than 30TB a month.
You are comparing VPS with Dedicated server.
some cpu features like vmx is not available in netcup "u need to pay more". like what Yura said vps ≠ dedicated.
Bullshit, Not a single provider, who sells Dedicated cores, sells you Dedicated cores.
They sell you threads, of a CPU, but they call it cores.
The difference between Threads is, that thanks to Hyper Threading, 2 Threads are using 1 Core. If someone would sell you dedicated cores, we would need to sell you actually 2 threads = 1 core.
Misleading bullshit.
That's not an issue, as long the machine is not under high load, if that's the case, you get kinda fucked, since 2 Thread will access the same core, so basically you performance goes down a lot.
Hetzner cost you 21.85EUR for i7 with 9k benchmark + 32gig + 2x3TB Disks, that beats the shit out of the VPS.
A VPS performance may differ from time to time, a dedicated gives you all resources 24/7, so you cannot compare these. One troll or abuser on the node and all your shit gets fucked up.
I've found in practice the 8-core Hetzner cloud server is faster than the 4-core i7-3770 by the amount you'd expect from treating the 8 cores as real cores and scaling the frequency. It ends up around 1.3x faster iirc. Yes they are actually vcores so the obvious inference is that the server hardware (at the time of my tests) had underutilized CPU and you could really go to town on them. Once there is more cpu contention of course performance will drop.
I haven't tried the netcup rootservers but everyone says they're great. They are modern and have SSD and ECC memory. Overall though I mostly prefer being on baremetal hardware to be completely separated from other users.
If I were getting a server of that type now, I'd probably get one of the Hetzner auction E3-1245v2 which costs a bit more than the i7's of similar speed, but have ECC.
the only hard part for me is that , in server bidding , if you want SSD, these servers come with at least 40-ish euros ...
so if you need more disk IO than CPU , netcup would be good place to go.
Could also be a provider would disable hyper-threading, and thus 1 core == 1 core.
If you take haproxy as an example, it's actually recommended to disable hyper-threading on the CPU to get the most of out the CPU - 20 threads causes worse performance than 10 cores.
So just because a CPU supports hyper-threading doesn't mean a provider actually uses it.
If you are going to keep it long enough to not mind the setup fee, you can get an EX-41 or EX-41S SSD for 34 euro in Finland. They are better than any of the cheap auction servers.
https://www.hetzner.com/dedicated-rootserver/matrix-ex
One issue that I found weird is that Netcup is very selective about which CPU features they pass-through to their VMs, I had to beg support to enable AES-NI (which they did) but no-go for AVX. AES-NI is super important for me because I do full-disk encryption so that was nice of them. Compared to my €34/m Hetzner server performance is still lacking a lot though I pay a hell of a lot less for the netcup server.
Pick something that fits your requirements.
Hetzner Cloud enables AVX if that matters. FDE on a virtual server of any sort sounds dubious. If you're serious about that, use a dedi.
Do you have a Root-Server or a vServer at netcup? I've had a Root-Server since April, and both AES and AVX were enabled from the beginning -- I didn't need to ask.
I also have a vServer: AES is enabled (I didn't need to ask) but not AVX.
I think that netcup's practices have evolved over time, but it sounds to me like you have a vServer and not a Root-Server.
Pick something that fits your requirements.
Indeed.
I talked with them, they know that they actually just sell threads.
Dr.Server sold also that stuff, but he sold dedicated threads as he wrote that into the offers.
Which was correct.
That's not entirely correct. A thread actually represents a logical processor core, therefore the term is not used fallaciously, merely it is a question of definition. Usually, you're aware of what you're buying and you can easily determine whether you're getting a physical core or a logical core.
In terms of performance, hyper-threading theoretically doubles performance, as one logical core has it's own pipeline and it's own instruction set, however, certain cases can clearly deficit the performance on hyper-threading, that being data dependencies or cache misses.
If you're buying a virtual machine with a dedicated CPU core, they're high likely speaking of what you call threads, however, that does not mean that performance per-se is degraded to a level you notice. (otherwise it wouldn't be sold as such).
Source? I've worked with a fair amount of haproxy installations for load-balancing, usually a bunch of machines in front of a cluster of workers, which have set an explicit affinity for the CPU map to threads, with one thread left out for other processes running on the LB's.
Haven't used netcup, but hetzner has very poor overall connectivity, longest pings.
Maybe that's true in theory but practically it's nothing like that. Running your CPU intensive computation with a number threads equal to the number of logical cores is not going to get you a 100% gain vs running with threads equal to the number of physical cores. In some cases you might even degrade the overall performance.
That is correct. I also conceded that in the next sentence.
Simply experience making haproxy scale to roughly 18 gigabit in/out at same time (18 gigabit from upstream towards consumers) per box and at the same time having to do SSL termination.
With E5-2640v4 CPUs we would get better performance out of the box with HT disabled compared to HT enabled.
We'd use 4x10g NICs, assign RX/TX queues to individual cores for each NIC and pinning haproxy to specific cores as well.
hyper-threading doesn't double your performance, you'd get maximum 1.7-1.8x out of it in certain workloads - and you'll still end up with a lot of interrupts for that to happen, which isn't exactly amazing in a setup doing hundreds of thousands of connections.
maybe when haproxy has better thread support (it's still super slow compared to using nproc) - since it will allow better utilization of local CPU cache for threads running on threads (threadception?), sharing much of the work done by haproxy processes currently.
Still something that has to be tested when the level of performance for haproxy threads become somewhat similar to running individual processes.
Netcup: VPS 1000 G7SEa1 - huh i guess its technically a vps but it's KVM, not sure where the difference is to the RS series.
Hetzner: Dedicated Xeon E3-1245 V2
That's a vServer.
Both vServers and Root-Servers are KVM. The main practical difference is that Root-Servers have dedicated cores (i.e., threads).
I think that it used to be the case that AES wasn't enabled by default on vServers, but netcup changed this policy at some point over the past year.
@All,
I'm not a customer of Hetzner but have been giving them more than side glances recently. Is Hetzner connectivity that bad? What's everyone's experience with it, within Europe and to the States?
They are fixing their network, but slowly. At least, it took a half year till they fixed incorrect routes to the Middle East (via London).
Example.
Go here http://lg.emix.net.ae/lg/
Enter hetzner.de and select Pings, you will see pings around 136 ms
Enter clouvider.co.uk, and you will see 120 ms
Despite the close geographical location, Hetzner has highest pings.
Try out OVH IP in Germany (DE1) 54.37.202.200, you will see the same 120 ms
Certainly not the best but also not the worst. In europe they are fine at least. You might see a couple more ms latency here or there but nothing huge. Outside of europe it's a bit hit and miss i guess. On one hand there is @willie who seems to experience a not exactly great connection but then i've seen people in the US and canada with stable < 120ms latency. Not sure about bandwidth but the occational 20-30MB download didn't result in any complaints. @Gulf has a nice example there and if you care about +/-16ms to places like .ae is up to you.
According to the traceroute, bandwidth goes the Germany through the Amsterdam. Therefore we see these +20ms (to the NL and back).
Previously it was London, now amsix.
Probably arabs have misconfigured their network, but other provders like ovh.de, @clouvider do not suffer from this issue.
Really strange.
I consistently get decent latency but not so good cross-Atlantic speed from Hetzner. My multi-GB transfers typically run 3-4 MB/sec to the US or Canada, but often 50+ MB/s or even the full 100MB/s (1 gbit port basically maxed out) in Europe.
>
>
Where is it going back? Seems the network you are tracing from simply peers at AMSIX and Hetzner uses that route? At least geographically this makes sense as Amsterdam is closer to Germany than London. My guess is that this is simply cheaper for them though.
For example,
OVH Route: Dubai -> Roubaix, France -> Frankfurt = 120 ms
Hetzner: Dubai -> Amsterdam -> Frankfurt -> Nuremberg = 140 ms
Looks like it goes through Germany to the Netherlands, then returns back to Germany.
Maybe but who knows how the Emirates network gets it's traffic to Amsterdam. All we (and Hetzner) know/see is that they peer there. If we were to assume the shortest possible connection it would likely pass through Belgium. Not arguing that the route is suboptimal (given the traffic highly likely comes via france even if the Emirates Network peers in Amsterdam) but as i said at least geographically Amsterdam isn't to bad and who knows if Hetzner even has a chance to peer with them in France. Roubaix doesn't exactly seem like an IX.
If you think that's bad check the current routing to scaleway. It goes through Singapore while a few days ago it went through Marseilles. Same provider, Etisalat.