All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Quick benchmark of Hetzner Cloud CPX21
Moving on from the deadpool that is hotlineservers, I decided to go set up camp for the AX41 since the space of LEPs that offer dedicated 5950x threads are now down to.... 0?
I went with the more isolated Helsinki node that may take up to 4-5 days to set up and put the 5€ shavings into a hetzner cloud. I've always heard good thinks about hetzner, now I'm starting to understand why. TBH I put hetzner off my radar many times because I always saw it as a place many LEPs go to kickstart their business in Europe, so I assumed it would be difficult for users like myself to navigate, it wasn't difficult- though my existing experiences with other LEPs definitely helped.
I'm ballparking, but I remember 3600x VPS geekbenchs would be around the same range. Sometimes they would be lower (down to 800-1000) if throttled or oversold. In fact cpuinfo shows it's max frequency at 4.3GHz, so at least something similar or equivalent. I miss 5950x geekbenches >1400 though.
Disk Speeds looks good except at 4k block size. What was suprising for me is how bad I expected the network speed to be, but it's pretty good for my purposes since it's not a heavy priority for me.
# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
# Yet-Another-Bench-Script #
# v2021-06-05 #
# https://github.com/masonr/yet-another-bench-script #
# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
Wed 30 Jun 2021 08:29:12 PM CEST
Basic System Information:
---------------------------------
Processor : AMD EPYC Processor
CPU cores : 3 @ 2445.404 MHz
AES-NI : ✔ Enabled
VM-x/AMD-V : ❌ Disabled
RAM : 3.7 GiB
Swap : 0.0 KiB
Disk : 75.0 GiB
fio Disk Speed Tests (Mixed R/W 50/50):
---------------------------------
Block Size | 4k (IOPS) | 64k (IOPS)
------ | --- ---- | ---- ----
Read | 81.83 MB/s (20.4k) | 1.07 GB/s (16.8k)
Write | 82.05 MB/s (20.5k) | 1.08 GB/s (16.9k)
Total | 163.88 MB/s (40.9k) | 2.16 GB/s (33.8k)
| |
Block Size | 512k (IOPS) | 1m (IOPS)
------ | --- ---- | ---- ----
Read | 1.50 GB/s (2.9k) | 1.65 GB/s (1.6k)
Write | 1.58 GB/s (3.0k) | 1.76 GB/s (1.7k)
Total | 3.08 GB/s (6.0k) | 3.41 GB/s (3.3k)
iperf3 Network Speed Tests (IPv4):
---------------------------------
Provider | Location (Link) | Send Speed | Recv Speed
| | |
Clouvider | London, UK (10G) | 2.64 Gbits/sec | 4.28 Gbits/sec
Online.net | Paris, FR (10G) | busy | 2.19 Gbits/sec
WorldStream | The Netherlands (10G) | 3.00 Gbits/sec | 2.21 Gbits/sec
Biznet | Jakarta, Indonesia (1G) | 340 Mbits/sec | 136 Mbits/sec
Clouvider | NYC, NY, US (10G) | 1.67 Gbits/sec | 994 Mbits/sec
Velocity Online | Tallahassee, FL, US (10G) | 755 Mbits/sec | 836 Mbits/sec
Clouvider | Los Angeles, CA, US (10G) | 786 Mbits/sec | 456 Mbits/sec
Iveloz Telecom | Sao Paulo, BR (2G) | 165 Mbits/sec | 543 Mbits/sec
iperf3 Network Speed Tests (IPv6):
---------------------------------
Provider | Location (Link) | Send Speed | Recv Speed
| | |
Clouvider | London, UK (10G) | 2.76 Gbits/sec | 3.24 Gbits/sec
Online.net | Paris, FR (10G) | busy | 2.77 Gbits/sec
WorldStream | The Netherlands (10G) | 6.18 Gbits/sec | busy
Clouvider | NYC, NY, US (10G) | 1.42 Gbits/sec | 994 Mbits/sec
Clouvider | Los Angeles, CA, US (10G) | 957 Mbits/sec | 597 Mbits/sec
Geekbench 5 Benchmark Test:
---------------------------------
Test | Value
|
Single Core | 1126
Multi Core | 3203
Full Test | https://browser.geekbench.com/v5/cpu/8627189
cpufrequtils 008: cpufreq-info (C) Dominik Brodowski 2004-2009
Report errors and bugs to [email protected], please.
analyzing CPU 0:
no or unknown cpufreq driver is active on this CPU
maximum transition latency: 4294.55 ms.
analyzing CPU 1:
no or unknown cpufreq driver is active on this CPU
maximum transition latency: 4294.55 ms.
analyzing CPU 2:
no or unknown cpufreq driver is active on this CPU
maximum transition latency: 4294.55 ms.
Comments
I've already deployed and started using Hetzner AX41, so I'm unable to run a benchmark at this time, but since I have use for a Hetzner CPX51 to temporarily speed things up...
Interesting and meek results compared to CPX21, based on this I decided to run another CPX21 GB5 benchmark.
CPX51 has 530% more cores than CPX21, but performs only 248% better in multicore processing (I ran CPX51 GB5 a second time with the same result) and absymally in single core.
None of this is conclusive of course since this is just a simple adhoc test. A more robust method would be to sample every few hours for a few days, which I won't do since I'm only seeking to gain some understanding of the value I'm getting to make some decisions on how to configure my setup.
CPX41
Loaded up another CPX41, looks like a sweet spot.
just saying, don't overinterpret these number or rely to heavily on it being like that all the time... as you said yourself:
most likely even multiple test over a longer timespan won't guarantue anything in the long run.
it's all shared, so depending heavily on the node you're placed on and the neighbors there. obviously that can change every other day, as it's hourly paid cloud and people will come and go on these nodes eventually.
CPU clock is intentionally choked.
These instances aren't for compiling code for sure.
@Falzo The providers can't put some controls so these numbers are stable like -+10-15% over time or too difficult?
Yeap, there's little point in doing anything scientific for your reasons stated. I do want a loose sense of which product tier is consistently more bang for buck since that can steer how I might temporarily spin up and down VPSes for extra boost.
@lowending what's your use case for powerful CPU? you probably already know about it but if your use case involves jobs that can be safely killed and restarted automatically, maybe look at https://cloud.google.com/preemptible-vms?
Another fan of the "real" cloud services, I see
However, most of the time, Hetzner cloud will end up being cheaper for them.
This, but also I am highly skeptical of adopting commercial systems/frameworks that once you start developing on, you basically lose all bargaining power because you've integrated into a type of workflow that they use, that makes moving out super difficult and expensive.
Some features like serverless is extremely interesting though.
I am pretty sure they can and do. but probably not in that small range. I'd rather expect something in the range 50-100% per core or so.
I just ran another GB5 on the 2 CPX41s, after 8+ hours of cpu loads of 75% constantly, multicore scores at 25% less than before. Again, pinch of salt. May just be more active neighbours instead of provider oversight.
random bench on a CPX31 just now (in production since months):
it is what it is. much bang for the buck, but not guarantued nor dedicated 🤷♂️
Is the NVME or CEPH better on Hetzner Cloud
> @lowending said: fio Disk Speed Tests (Mixed R/W 50/50):
There are tools like Terraform that can help in making this a bit easier and almost all the basic stuff is available on all 3 clouds. Preemptible VMs are available in GCP and AWS and Azure because they all need to get money out of spare machines/VMs. So I think the major work is adapting your app to work with machines that live 24 hours max and can shutdown safely in 30 seconds or less. if it is made that flexible, you can use it on different clouds without too much effort in the future.
Yes that's like putting puzzle pieces together to do something useful (Step Functions / Google Workflows).
Haha. I just like to look at different solutions available. HC is really nice and great pricing. They are slowly adding new features like the firewall feature that was added recently which is nice that it is in active development. Wish they add a US location some time down the line (and ability to use private networking between Helsinki and US city).
Ceph is a little more safer than local disk so if you don't want local disk speed you can use it (not a substitute for a working offsite backup).
I've already pretty much done/been doing what you describe using docker, which is why these types of products don't attract me against the risk of lock-in. Docker has been a dream for this, even if not quite perfect. I haven't implemented any of the quality of life automations though, not sure if it's worth the hassle.
I did drool over the idea of running cross different clouds, but unfortunately private networking is a crucial feature to reduce metered traffic and latency between VPSes. So while it was good in theory, when I implemented it in real life it became more "expensive" than it was worth. That being said, the idea of automated replication across another infrastructure to run within that private network in the event of a disaster could be done- something to think about for the future.