All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Atom Dedicated vs a 600HZ vCPU?
JohnMiller92
Member
I was looking at my VPS specs ($7/year), and they are:
model name : Intel(R) Xeon(R) CPU E5620 @ 2.40GHz stepping : 2 microcode : 21 cpu MHz : 600.022 cache size : 12288 KB
From what I understand, that means I'm allocated 600MHz from a "virtual core", which is descendant from a E5620 processor?
I was also looking at the Kimsufi servers, and they have a Intel Atom D425, for $4.99/m. Which runs at 1.8GHZ.
I'm curious (this might sound really dumb), but obviously 1.8GHZ is higher than 600MHz, but is it really going to be faster? I have them compared here.
I'm kinda curious about virtual private servers compared to Atom/mini dedicated ones. Doesn't have to be just this particular D425 vs a partial E5620.
There are cases where you can get more performance buying a VPS than one of these cheap dedicated servers, right?
A lot of questions and am sorry if it might seem like gibberish, just really curious. Thanks for reading =]
Comments
According to the cpubench scores, the atom scores 350ish points on its own. An 8-thread e5620 at 2.4ghz scores around 5k bench. Divide by 8 threads (you're probably assigned part of a thread, not part of a core) and by 4 again (2.4/0.6=4) and you're at 150ish cpu score.
So, in theory, the atom should be faster by about a factor 2-2.5
That's interesting! Did not know hosts could give out parts of a thread. Wow, gives a whole new meaning when I see dedicated atom boxes in comparison to what I have, haha. Much more powerful than I thought
The older Atom D400's are not what you expect it to perform like.
They're the older Atom microarchitecture that performs pretty poorly because of the way how the chip was designed. IIRC, the Atom up until Silvermont used in order execution as opposed to out of order execution which most processors at the time were using.
You're actually much better off getting the Westmere based Xeon over the dedicated atom D425.
Also the Xeon may have CPU instruction sets that the Atom lacks - which could improve performance drastically depending on your typical workload.
Does the host allow CPU spike? If so, Atom shouldn't even be an option.
Not sure tbh. What's "CPU spike" do? Lets the customer go past the 600MHz threshold for xxx period of time?
cpu MHz : 600.022 from proc/cpuinfo
Doesn't mean it's limited at 600 mhz. So your understanding is wrong.
Don’t worry - it’s never in stock
600mhz may be your default share but most, if not all, should allow short CPU bursts.
Think it like speedstep.
Ha, no kidding. Been checking that damn website for the past 3 months on an off. Really want to try out the i3-2130, but don't need that much space at all. I really just need 30GB SSD. Those dedicated cores is what I'm really after too
Thanks! Did not know that, so does this mean when htop shows 100% cpu usage, and if the host has "CPU spiking" enabled, that means you're now affecting other neighbors pretty badly? And if it stays like that, you'll prob get in trouble?
So
cat /proc/cpuinfo, is not really a 100% accurate representation of your maximum CPU power then? I assume it should just be used as a reference to what processor you have, butt he MHz can really be anything. (up to the max rate the CPU runs at)e: If I got that right lol
@JohnMiller92 You can ask your provider what their policy is for CPU bursts. Most providers allow for bursts as long as you don't abuse by hogging the core above the allowed limits. If you have some CPU intensive tasks, it would greatly benefit from a single burstable Xeon core rather than all the cores of that ancient D425.
Also, the speed that you saw is the idle speed. the E5620 can burst to 2.66 GHz if its performing computations.
If you're after the D425 consider signing up for the KS notifier websites. I've had success getting the KS-1 a few times. Most KS-1's come with the Atom N2800 CPU. It's not powerful but better than the D425.
The CPU you get with Kimsufi is probably going to be N2800 anyway. If they don't have the CPU they state on the website, they give you a similar or a better one, and they definitely have more N2800s.
Aww, I see. I just figured it was the max amount allocated cause I was looking at Vultr vs DO vs etc, and their
cat /proc/cpuinfoshows it running at like 2.39GHz. However, am not sure if they were idle. For example here, it shows2399.996 MHz. Probably a different CPU, but mycat /proc/cpuinfostill shows the max (2.4GHz), but running at600MHz.Most likely the host node is set to power saving and what you see is Intel SpeedStep taking a step in de-clocking the CPU to save power.
Well, yes and no. I just yesterday worked on a blake2b implementation and tests ( proper working, endianess independence, and performance with roughly similar CPUs to the ones here) and have it fresh in my memory.
Clock is overrated. Have a look at the memory speed in the intel comparison provided by OP (in his "here" link)!
This translates to what I also observed again and again: If you do lots and lost of computation on a small data space clockrate is on your side. But most of what runs on servers is the opposite, it's pumping data with some limited work on those data and that's where memory speed and cache size strongly dominate the game.
Add to that SSE 4.x, AVX[2], AES NI, etc. and even an old Xeon VPS fraction will brutally beat an Atom on most server jobs.
If I were asked for a rule of thumb I'd say that (even an weak cpu) dedi is attractive for (a) steadiness (as in "always the same performance") and (b) somewhat better security but a VPS based on a Xeon (or Epyc) will almost always easily offer MUCH better RELEVANT performance.
That's btw. one of the reasons why I designed vpsbench the way I did. It tests RELEVANT performance not some worthless instruction mix and it intentionally works hard on memory and caches - because that's what you need in 97+% of servers.
A lot of people don't understand that one clock cycle does not necessarily match up one machine cycle on a given processor. The Atom was designed as a low cost CPU, and therefore there's going to be much less circuity on the die to support instruction set extensions and CPU designs that a Xeon or a Core 2 machine at the time.
There's a reason why Xeon's are much more expensive and perform better compared to an Atom processor.
E5620 is an older architecture (less performance per GHz than a current cpu) and as a low order approximation I'd say just compare the passmark scores and scale by frequency. That said, if you want a dedicated server or dedicated cores, you are probably better off buying that instead of expecting decent compute performance from a $7/year vps. If you max out your 600 mhz share of the E5620 for long periods that is likely to be considered abusive. Ultra cheap vps are best for very lightweight services.
I haven't tried them myself but several people here have recommended https://www.netcup.de/vserver/ which at 7 euro/month (annual payment) comes with 2 dedicated E5 "cores" (most likely vcores, performance will depend on overall cpu utilization on the node but at worst should be about like 1x full core). That's in a DE location which might or might not work for you.
$7/y and 84€/y is a pretty large difference
Forget passmark scores. They might or might not be helpful in determining a processors usefulness for a desktop but for typical server usage they are quite meaningless and even misleading.
Just 2 hints: memory access is about 100 times slower than cache access and thread switching still is worth A LOT of cycles. So, on a server there are seriously important and even decisive factors that are hardly visible if at all in the usual CPU comparisons.
Trust me I didn't write vpsbench for the fun of it. I wrote it because I wanted a benchmark for SERVERS and their typical jobs (plus something I could run on a VPS without worries about angry neighbours or providers). And I also had damn good reasons to write both single core and multi core processor and memory tests.
As for the 5620 yes that's an old processor but it still runs circles around an atom in server applications, particularly in multithreaded ones.
I find for my own cpu-intensive server workloads, passmark is a good predictor of application runtime. That said I'm sure it varies with the application.
Yes, fast dedicated cores are simply not a $7/y product. If slow dedicated cores suffice, a Scaleway C1 (4 core ARM server) is 6€/y per core with no internet (internal network only) or 9€/y per core with a dedicated ipv4. Each core is about 1/10th the speed of a current E3 or maybe 1/5th of an E5-5620 core. It's actually a reasonable deal for pure computation though I'd still go for something faster if possible.
Edit: oops, per @FlamesRunner, I had the IP and no-IP Scaleway prices switched. Fixed now.
Bit of an exaggeration there dont you think? obviously they are not meaningless, they have meaning. I think the multicore score is given too much importance on the passmark website, but the single core speed is there.
If you think they really are meaningless, I'd like to see some real world benchmarks that show that.
Pardon me but I'm not interested in fighting belief systems. I have laid out some important factors and unlike probably most here I have actually done quite some research on the matter and also have written a server benchmark as well as plenty of tests (always including performance tests) for crypto code which have been run on diverse machines. What I say here is based on that concrete experience.
So if anyone is interested in understanding the matter better I think I've provided some useful information and hints. Those who prefer to simply believe in this or that test or truth (or "truth") are welcome to do so and I will not try to make them change their mind. My interest here is to contribute for those who are interested and not to evangelize or convince anyone.
Edit: I'm quite amazed by some peoples attitude who really think that their demand to prove what I say to 'convince' them might have any weight to them. But to show some good will:
Running a simple crypto hash over a data space easily fitting even within a small processor cache the difference between an AMD Ryzen and an old N450 (iirc the model) Atom was about 3:1 (1 GB/s vs. about 350 MB/s). Running another crypto algorithm (a PHC finalist) that does not fit in a small processor cache but is largely based on a very similar ARX construct show a difference of about 15:1 for the same two processors.
Considering that running hashes over not small amounts of data like a 1 or more MB web page as well as encrypting them (https) is a very common task for a server nowadays what I just showed is a VERY considerable difference and very clearly favouring a "large"/server processor and an Atom and that difference is to a considerable degree due to caches and memory speed rather than the processor speed. And that was a single core test. Doing the same thing multithreaded (as servers typically do) would shift the balance even considerably more in favour of a Xeon or Zen processor and would bring the Atom performance down to a ridiculous level.
And now feel free to believe whatever you please.
@willie
Think you may have reversed the pricing w/IPv4 and wo/IPv4.
I'd be interested in seeing some test results with actual numbers in them. My own tests indicate that passmark and frequency scaling works pretty well, but as mentioned, that is for CPU-intensive applications (e.g. data analysis tasks that run at 100% CPU for multiple hours on all available threads). A web app that needs to compute for a few milliseconds per visit and is constantly switching in and out of idle might not measure the same way.
I do generally favor dedicated servers over low-end VPS if you're doing anything where sustained performance matters. There are fewer variables with a dedi, no concerns about having noisy neighbors and/or being one, etc. For burst CPU, hourly DO and Hetzner instances work pretty well IME. I can't get too amped up about CPU availability at the $7/y level. If that financial impact is significant to whatever you're doing, either it's not serious or you're in a weird situation that I'd want to hear about.
https://checkservers.ovh/ reports that KS-03 (Atom N2800 with 4gb ram and 2TB disk, 8 euro/m) is available in France, if that's of any interest.
Uhm, I just DID give concrete numbers from tests I very recently ran. And trust me, I have no interest in painting one processor bad and another one good.
Let me open up a little bit: Some of my recent work is to do with KDFs that is with algorithms which (a) intentionally mess around with memory and preferably lots of it and (b) are designed to minimize differences between architectures and models. The reason is that you want to create a certain relatively high load on the client but have that load very SIMILAR and within a relatively tight frame over all processors and models. The reason is that on one hand the client is supposed to do some hard work related to login attempts but on the other hand to not have large differences because such an algorithm is worthless if it takes 20 ms on a fast Xeon but 20 s on a weak processor (say a netbook). Now, if you bring GPUs and ASICS into the game (against which these algorithms should protect) you'll see what I mean. You want the algorithm to take a certain amount of time within a relatively tight frame of time, say 3 s on a netbook but still no less that 0.3 s on a fast Xeon. Such a legitimate user with a weak netbook needs only a bearable amount of time to login while at the same time a hacker farm with hundreds of FPGAs still has no chance to bruteforce their way in (because no matter how powerful their farm is the algorithm doesn't give them a really major advantage).
That's why I have to do a lot of performance testing on a lot of processors (incl. non x86). Doing that I experienced a lot of surprises and gained a pretty good understanding of what makes crypto loads (which are quite similar to typical server loads) fast or slow and what processor speed is really worth.
I think you are right insofar as the use case you describe matches a LOT better with passmark and similar (largely a funny mix of integer and floating point computations and some more stuff thrown in). As you correctly mention a typical server load is quite different however and that is what I clearly related to. My assertion was not that passmark or processor frequency is worthless in general but that it has little value for SERVER use cases (which probably most here are interested in).