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
Use YABS, or if you know a bit about fio, use it instead..
special thanx to JSG, ... boost contabo nvme issue
my conclusion, .. contabo is contabo as usual .... there is no significant difference than before
if someone has "high level contacts" with the provider, may be we can get definitive answer to this key question
for the provider -
sometimes it is better not the answer the question, as the answer may cause more damage.
Ask the yabs "experts" ...
Let me see how far I can get. Yes, my contact certainly has that info or can get at it, but still, how many vCores they put on a HWT is something many (most?) providers try to avoid talking about.
And no, sadly some AUP like "fair share is x %" is not really an answer but more of a figure they calculate and which based much on actual real use/load of the system, or in other words, on the fact that actually quite few users use their fair share. At best I'd take it as a crude and vague indicator.
But I'm with you, if I had to decide about it, I would demand all providers to clearly tell in all offers how many VPS vCores there are on a HWT. But don't hold your breath ...
@all
question is -
You will thus find out the max theoretical IO supported by the server storage intern or extern storage.
NOW ! you can test your WM and can see if "is there any global I/O limit" considering :
a. max IO from soft RAID
b. max IO from connection of external storage ( one external storage can easy give you 50 GB/s , but if you have only 10 Gbps connection you lost the most )
c. how many VMs are on the server and how many servers are on external storage
Most lowend providers use NVMe internally with direct access, below is an example of what to expect, max. IO on 4k
You pretend to know so much, yet ...
Nope. I just reserve the right to no comment on obvious BS and emotionally motivated hit attempts.
Interesting. So mirroring a drive is BS? Strange.
Mirroring a drive - any drive - means significantly enhanced safety of data. customers data, well noted.
And striping a drive isn't BS either. As you obviously don't know it, let me help you: striping very significantly decreases accesses to drives which is particularly desirable with VMs and also increases drive life time.
Tech porn yada yada. Your super-duper 100 Gb/s Infiniband Roce v2 remote storage is quite slow as compared to other providers solutions.
Well, how many VM vCores per HWT do you put? Is a vCore of say your currently offered €3/mo VM 1/3rd of a HWT? 1/4? Tell us!
Hey look! @dedicatserver_ro is pissing in someone else's garden again!
If @jsg weren't so full of shit then he wouldn't have a chance to attack him.
I see, so a drive connected via PCIe magically can't break? And also a drive connected via PCIe magically does make copies of the data stored on it?
It seems, you are the one who missed some classes ...
Oops!
when you have no arguments ...
Wikipedia is simple enough for you, I hope
Btw, no matter how you try to deflect by repeated personal attacks, albeit barrel bursts, your VPSs drive performance isn't exactly great.
Oh and, how about answering my question?
how many VM vCores per HWT do you put? Is a vCore of say your currently offered €3/mo VM 1/3rd of a HWT? 1/4? Tell us!
... says Mr. clueless whose new hero isn't exactly winning his self-inflicted war ...
He just does what he thought he's good at. That he was thinking wrong might be his other specialty.
RDMA over Converged Ethernet (RoCE) is a network protocol that allows remote direct memory access (RDMA) over an Ethernet network. It does this by encapsulating an IB transport packet over Ethernet. There are two RoCE versions, RoCE v1 and RoCE v2. RoCE v1 is an Ethernet link layer protocol and hence allows communication between any two hosts in the same Ethernet broadcast domain. RoCE v2 is an internet layer protocol which means that RoCE v2 packets can be routed.
....
RoCE versus InfiniBand
Stop the BS already. You know perfectly well that 'IB' stands for 'Infiniband' and actually links to it.
I don't see confirmation of your assertion in your screenshot. A screenshot of a fully loaded node showing the number of real cores & HWT along with the number of VM cores would be more useful.
But if the number you tell ('2') is true, kudos, that's decent.
FC, FC-NVMe[33][34]
TCP, NVMe/TCP[35]
Ethernet, RoCE v1/v2 (RDMA over converged Ethernet)[36]
InfiniBand, NVMe over InfiniBand or NVMe/IB[37]
The standard for NVMe over Fabrics was published by NVM Express, Inc. in 2016.[38][39]
The following software implements the NVMe-oF protocol:
Linux NVMe-oF initiator and target.[40] RoCE transport was supported initially, and with Linux kernel 5.x, native support for TCP was added.[41]
Storage Performance Development Kit (SPDK) NVMe-oF initiator and target drivers.[42] Both RoCE and TCP transports are supported.[43][44]
Starwind NVMe-oF initiator and target for Microsoft Windows, supporting both RoCE and TCP transports.[45]
Infiniband Roce v2 remote storage - It does not exists, but you have no way of knowing):
a. Infiniband remote storage exist
b. RoCe v2 remote storage exist
You don't seem to understand what you're reading
Says the person making a million clock_gettime() syscalls and then is amazed that his benchmark isn't representative of real results, because he doesn't know about context switch and KPTI overheads.
You just threw TNT on the fire..
I'm not here to be tested - and certainly not by you who again and again offers tech mumbo jombo to distract from a simple fact: your disks aren't great performers.
Cute attempt but unfortunately (for you) you confused two similar sounding but actually very different calls, gettimeofday() in fio which is relatively expensive, which is why fio offers a mode massively reducing those calls, and clock_gettime() which is a very cheap/fast call and on modern systems pretty much just an ASM instruction plus the time info is obtained from memory shared between the kernel and user space, which, sorry to disappoint you does not involve crossing the kernel boundary ("a syscall").
Not that I asked for it but thanks for confirming again what I said: you do not even know what you talk about.
Not since KPTI was implemented:
https://en.wikipedia.org/wiki/Kernel_page-table_isolation
I completely understand that urge to prove someone wrong, but it's a repeat of that HS saga once again, simply because you want to be right, not because you want to understand the truth.
In addition, you're sidestepping the main issue too, which is why are you making so many
clock_gettime()
API in a CPU benchmark.As wrt the matter: KPTI or not (my benchmarks are done on FreeBSD btw) fact is that gettimeofday() is massively more expensive and slower than clock_gettime().
For a start, you (and a few others) have given in to their urge to prove someone (me) wrong. That's how this thing began.
Plus: easy to prove false: I made a small error (I think even in this very thread) and I "confessed" it right away and even apologized.
You (and some others) however clearly and evidently are all about arguing against me. Just take this very post of yours. Fact is that gettimeofday() is massively more expensive and slower than clock_gettime(). Yet you only focused on finding some detail, any detail, no matter what, that seemed useful for showing that I was wrong. That is the only goal of this sharade. Maybe your next argument will be that on some exotic architecture there is little difference between those 2 calls; if that exists you'll find it and try to use it against me.
You can not win this.stupid game. Simple reason: Chances are you continue to fail - and if you don't, just assumed you actually could show me weak point, I'd simply thank you and improve my benchmark.
Pardon me, but your attempts are boring. Give it a rest.
Edit, following your edit:
BS. The clock cycles used are tiny and meaningless compared to the cycles used by the tests (e.g. writing to the disk). Once more you talk about something that seems to be a good weapon but actually you simply don't understand and know what you are talking about.
... and therefore are useless to 99% of people.
Sure you can, back your claims up with evidence rather than making long posts that bullshit around the matter.
If he can not win he will say BS.. classic..
The key question that I was referring is this one
is there a throttle for max I/O that applies to ALL the customers ?
So far no reply from @contabo_m
@dev_vps that's on you to judge. Contabo is extremely unlikely to admit to anything.
Well, they are the one putting this quote on their website
“This NVMe is three times faster than any other NVMe tested (at other providers)”
That does put the onus on them to justify that quote with some benchmark numbers
For a layman customer - does it mean writing 20gb data file (using their VPS with the latest NVMe) will take 1/3 of the time taken for any competitor provider VPS with NVMe storage?
Well, I can say my penis is three times longer than any other penis measured. But I also don't have to show you my penis.
Yeah right, because FreeBSD magically make slow processors and memory fast and fast processors and memory slow and it also magically changes slow disks to fast ones and vice versa.
OK, I've had with you assholes.
(a) You do not set the rulers for me. (b) guess whose opinion I care more about. That of significant provider or that of an obtrusive idiot? (c) I'm not selling something. I wrote a benchmark software in my time and with my work that I use for benchmarking and making reviews in my time and with my work (and lots of it) which I provide for free to our community - and you assholes have but demands and vile as a thank you?
Nobody forces you to read my reviews, just like nobody forces me to read certain threads of certain providers whom I find obtrusive. So I simply don't read them, case closed. You could so the same, but no, instead of saying "I don't like your reviews but thanks for the effort" you fervently try to bash me, my software and my work and to abuse anything I say trying to find something, anything, you feel could be useful in your little private war against someone who has asked nothing from you but halfway decent behaviour.
I'll be generous and not ask what you did for this community or what qualifications you have ...