Not_Oles' Cyber Weekend Dream Deal!

Not_Oles

Friday is almost over here in Sonora! ?️ I've been trying all day to dream up a Cyber Weekend deal. What do you guys think of this one?

Not_Oles' Cyber Weekend Dream Deal!

Buy three months. When you renew for the second three month period, you pay only for two months. Thus, month number six is free.

Fine print: Not recurring. If you cancel during the first three months you are charged the per diem rate. If you cancel during the second three month period you are charged the full monthly per diem rate during months four and five. Thus, no refunds during month six. Deal expires 11:59 PM UTC on Cyber Monday based on when Paypal payment is made. No service level agreement. Just for fun. Not for business use. Grumpy, incompetent administrator. Limited availability. Delivery might take awhile. GNU/Linux command line peeps encouraged. *BSD and Plan 9 users welcomed. Window$ users welcomed. Apple users welcomed too. Additional free month for first person who successfully runs KVM inside an unprivileged LXC container.

The Plans

4 x Ryzen 3700X, 4 GB ECC RAM, 50 GB NVME RAID 1, 1 IPv4, 1 Gbps, $10/mo

8 x Ryzen 3700X, 8 GB ECC RAM, 100 GB NVME RAID 1, 1 IPv4, 1 Gbps, $20/mo

More information

Hetzner AX51 in Helsinki

LXC, but KVM if you have a good reason.

Yabs and additional information in previous thread.

You get access to the Proxmox control panel. You can see the load averages and performance graphs for the host node. You can start, stop, reboot, backup, and restore your VPS. You can see the load averages and performance graphs for your VPS.

IPv4 works fine. IPv6 into the host node also works fine. No IPv6 in the VPSes until I figure out what to put in /etc/network/interfaces and whatever else is needed. A few reboots and network bumps are guaranteed.

Check my 1337 front end skillz.

How to Order and Pay

To order, please post here, PM me, or use the email on my profile.

Please send Paypal payments to the email on my profile. Please read the fine print section, above, before paying.

What People are Saying

"the fastest VPS I ever used"

"I've been using this one from him. Highly recommend! His support is really amazing!"

"premium stuff"

"resplendent as the Sonoran Sun"

"god bless not_oles ^^"

seenu

I guess i know the answer but for the sake....

are cores dedicated?

CamoYoshi

A generous deal from an already generous guy. Thank you sir!

Not_Oles

@seenu said:
I guess i know the answer but for the sake....

are cores dedicated?

Thanks for your question! If I understand correctly, the processor in the AX51 has 1 socket, 8 cores, and each core has 2 hyperthreads:

root@hels /etc/network # nproc
16                        # 1 x 8 x 2 = 16
root@hels /etc/network # 

So the 4 and the 8 in the above Plans are 4 or 8 of the 16 hyperthreads of independent execution.

Where, as here, there are less than 16 users, each user has, in effect, at least one fully dedicated core.

The advantage is, when the machine is relatively quiet, and it is quiet most of the time, a user with nproc = 4 can have as much as 25% of the entire processor capacity. A user with nproc = 8 can have as much as 50%. If the machine is super busy, then the kernel has to decide how to allocate the processor capacity. But, if, as here, there are less than 16 users, each user ought to have at least 1/16 the full processor capacity all the time. In this scenario, most of the time, each user can have much more than 1/16 because, almost all the time, not all users are doing things simultaneously. Pinning one thread as fully dedicated per each of 16 users would forbid any user from ever using more than 1/16 no matter how quiet the machine was at the time.

There are people here on LES who know much more about all this than me. I have tried this scheme on multiple machines with multiple users for months. It seems to work really well! But I certainly am open to suggestions.

I read an interesting paper about all this: Biting Off Safely More Than You Can Chew: Predictive Analytics for Resource Over-Commit in IaaS Cloud.

It's good to ask your question because people throw the word "core" around ambiguously. Really, if I understand correctly, there typically are sockets, cores, and threads. Each core has two threads. But, people often refer to a thread as a core. So one has to be careful to understand how the terms are being used.

Greetings from Mexico!

seenu

Very nice explanation, as always.

my doubt is...can you really make any profit/breakeven if you sell it for only 16 people?

or, since not everyone uses all cores all the times, it is okay with little over-committing?

---

on another note...i am still thinking to take ax41 myself or grab a netcup Rs spl2 from another guy or take a slice from you.

i want to use it to host my production sites - which are not cpu intensive
i want to run gitlab runner on it which dependabot - which check everyday for all updates npm, composer packages in all of my repos and then runs a CI to see whether they are compatible or brings any breaking changes.... which is very cpu intensive as per my knowledge.

actually i became crazy with ryzen because a guy at LET offered 1core ryzen vps for free for a month, which i subscribed without any expectation but just to run my gitlab-runner (i have such fetish for gitlab runner and dependabot lol)

and i just blown away with ryzen performance
my existing ssdnodes 24core vps will take 8mins to finish a job and single core ryzen is able to finish same in 1.5mins ... that is like mind blowing for me.

and from that moment....my love for ryzen has gone up
then i started thinking about a dedi for myself ... reason : if 24core vps is not able to deliver something a 1core one, it means.... they are not true cores and they are highly overselling OR ryzen is such a powerful one.

sorry for hijacking your thread :P

AC_Fan

@seenu Nobody goes for SSDNodes for the CPU, it's highly oversold. But, they do have a decent amount of RAM, disk and bandwidth for their APAC locations, so if somebody's use case primarily requires those (for eg. Nginx cache, private VPN etc.), it can be a decent solution. CPU intensive tasks are best done with German providers.

seenu

@AC_Fan i didn't went with SSDNodes for CPU intensive work but currently most of my production sites are with them only

only after utilizing a ryzen core, i realized how much oversold their servers are...

and i don't know (TIL) SSDNodes is famous for APAC

AC_Fan

@seenu said:
@AC_Fan i didn't went with SSDNodes for CPU intensive work but currently most of my production sites are with them only

only after utilizing a ryzen core, i realized how much oversold their servers are...

and i don't know (TIL) SSDNodes is famous for APAC

Yeah, I wouldn't go for SSDNodes if I care about an individual VM, cluster of cattles only on those type of providers. And they aren't famous as such, just decent value for APAC, if they can fit your needs.

yoursunny

@Not_Oles said:
The advantage is, when the machine is relatively quiet, and it is quiet most of the time, a user with nproc = 4 can have as much as 25% of the entire processor capacity. A user with nproc = 8 can have as much as 50%. If the machine is super busy, then the kernel has to decide how to allocate the processor capacity. But, if, as here, there are less than 16 users, each user ought to have at least 1/16 the full processor capacity all the time. In this scenario, most of the time, each user can have much more than 1/16 because, almost all the time, not all users are doing things simultaneously. Pinning one thread as fully dedicated per each of 16 users would forbid any user from ever using more than 1/16 no matter how quiet the machine was at the time.

This looks like the perfect box for encoding push-up videos 24x7.
If only I'm strong enough to record 8 hours of push-ups every day...
Chest is very sore after 69 push-ups yesterday.

Not_Oles

@seenu said: can you really make any profit/breakeven if you sell it for only 16 people?

Well, you can take a look at the numbers. Very approximately as follows: the server costs $60 per month. Sixteen users at $10 per user (or eight users at $20 per user) is $160. So I make $100 per month. Profit actually is less because there aren't sixteen users, the server is a bit more than $60, there are Paypal fees, etc.

it is okay with little over-committing?

Where is the line between over-committing and over-committing? Did you get a chance to look at the paper to which I linked above? If there is a high 90+ percent chance that you get your entire 4 or 8 cpus every time you ask for them, what is the difference between that and dedicated?

I am not doing mathematical modeling. And I do not have a big enough sample. But I do know that my setup has worked well for months with a bunch of great users on multiple servers from multiple providers. I have had exactly zero complaints from users about cpu performance. One of my users really should make a complaint just for fun! Why not!

i am still thinking to take ax41 myself or grab a netcup Rs spl2 from another guy or take a slice from you.

How could you be satisfied with less than a fully dedicated, bare metal server? I think you should go for bare metal. Just my two cents!

very cpu intensive as per my knowledge.

How intensive is it? More than running yabs? More than compiling? It seems (if I understand you correctly) that one Ryzen core is able to finish your once-a-day job in 1.5 minutes. So, do you only need the intensive cpu for 1.5 of the 1,440 minutes in a day?

sorry for hijacking your thread

No worries! It's fine! Have a great day! Greetings from Sonora!

seenu

@Not_Oles said: How could you be satisfied with less than a fully dedicated, bare metal server? I think you should go for bare metal. Just my two cents!

I guess so

Did you get a chance to look at the paper to which I linked above?

i didn't completely because it is asking for signup

@Not_Oles said: once-a-day job in 1.5 minutes

it is not once a day actually, for each repo once a day then for each npm/composer update, it runs another CI job

on an avg, i see more than 100 jobs a day...still plenty of idle time for cores though.

Not_Oles

@yoursunny said: 69 push-ups yesterday

I will be impressed when you get sixteen users to sit on your back while you are doing the push-ups.

Thank you again for your previous comment about NUMA. I looked into NUMA a little bit, and it seems not too relevant to this cpu because, here, there is only one socket.

On Debian 10 plus Proxmox, when

• I am logged in to this server via ssh and
• without changing anything about IPv4, I add an IPv6 portion to vmbr0 in the host node's /etc/network/interfaces and then
• execute systemctl restart networking

what I am seeing is that the existing ssh connection is preserved but all LXC and KVM containers lose their external WAN IPv4 networking until they are rebooted.

I guess I see that the bridge vmbr0 is not involved in my external ssh connection to the host node but is involved in the external WAN connections of the containers and VMs all of which use the bridge vmbr0.

Why can't the containers and VMs continue to use the IPv4 part of the bridge, the configuration of which was unchanged? Thanks!

Greetings from Mexico! ????

Not_Oles

@seenu said: i didn't completely because it is asking for signup

If you look around a bit, maybe you can find it or find out how to find it.

100 jobs a day...still plenty of idle time for cores though.

So, about one hour and fifty minutes (about two hours) out of twenty-four hours. Hmm. Fair share with 16 users might be 24/16 = 1.5. Could you be needing just a little less than fair share?

yoursunny

@Not_Oles said:

@yoursunny said: 69 push-ups yesterday

I will be impressed when you get sixteen users to sit on your back while you are doing the push-ups.

I had one 5-year-old user sitting on my back. However, my father wasn't pleased when he saw that photo.

Thank you again for your previous comment about NUMA. I looked into NUMA a little bit, and it seems not too relevant to this cpu because, here, there is only one socket.

EPYC has a feature of dividing a single processor into two or four NUMA nodes. It would offer better isolation for some containers.
However, if you are letting all the container scale to multiple cores, you won't need this configuration.

what I am seeing is that the existing ssh connection is preserved but all LXC and KVM containers lose their external WAN IPv4 networking until they are rebooted.

I guess I see that the bridge vmbr0 is not involved in my external ssh connection to the host node but is involved in the external WAN connections of the containers and VMs all of which use the bridge vmbr0.

I don't know anything about bridges. Users will forgive the grumpy, incompetent administrator.

My setup for computer networking class students are much worse than these, but I only make changes when they are sitting in the classroom so that nobody is affected.