Cavium ThunderX based Scaleway ARMv8 Cloud Servers Go for 2.99 Euros per Month and Up

Orange Pi Development Boards

Scaleway launched 32-bit ARM server hosting services in 2015 for 10 Euros per month, before dropping the price to 2.99 Euros per month half-year later, and now the company has just launched a new offering with 64-bit ARM servers powered by Cavium ThunderX processor going for 2.99 to 11.99 Euros per month depending on configuration.

Click to Enlarge

The processors are equipped with DDR4 ECC memory, and all three services included unlimited transfer, so you don’t need to pay for any bandwidth fee. While the price is shown per month, you’ll be billed by the hour (0.006 Euro/h for ARM64-2GB), so if you are using those for development it may even cost less per month, as you can turn them off when not working.

All server are located in a Paris data center in France, and runs Ubuntu 16.04, but more operating systems and “InstantApps” will be added to the selection. More servers will soon be available in their Netherlands datacenters (AMS1).

You can add a new ARMv8 server in Scaleway dashboard to get started with the new servers. You’ll find a few more details on Scaleway Virtual Cloud Servers page.

22
Leave a Reply

avatar
22 Comment threads
0 Thread replies
12 Followers
 
Most reacted comment
Hottest comment thread
12 Comment authors
Pablonutjob2GanjaBearJeroenPhilipp Blum Recent comment authors
  Subscribe  
newest oldest most voted
Notify of
blu
Guest
blu

I was using their C1 baremetal servers back in the day, and they were fine, but I moved on to ARMv8. I’m glad they’ve upgraded to ThunderX – they have a returning customer.

tkaiser
Guest
tkaiser

Any information available from which kernels we can choose from on Cavium?

JotaMG
Guest
JotaMG

The price for a similar x86_64 setup is the same 2.99 euros.
I wonder what is the best choice, performance wise??

Simos
Guest
Simos

@cnxsoft

They are running the mainline 4.9.23 kernel. Here is the config, https://github.com/scaleway/kernel-tools/blob/master/aarch64/4.9.23-std/.config
As before, it is a kernel version that deviates from what you get with Ubuntu 16.04 (not 4.4, nor 4.8).

CampGareth
Guest
CampGareth

@JotaMG
You can pretty much guarantee it’ll be the x86 servers that win on performance. ARM has advantages, cost and efficiency, so it should be cheaper to buy and run for Scaleway. You’re an end user though, you shouldn’t care about power consumption or much of anything except performance and price. If lower operating costs for Scaleway don’t translate into lower prices for you there’s no benefit to ARM. What we’d need is someone like Amazon picking this up since they’ve got the volume and a long-term view.

Member

what does the baremetal mean? A Cavium ThunderX CPU has only 2 cores?

Philipp Blum
Guest
Philipp Blum

@fkpwolf
It isn’t bare metal. The C1 instances are bare metal (32 bit). The Cavium ones are virtual servers because the Cavium could have 48 cores. There are also Cavium CPU with 8 Cores available but I don’t think they using it for the server. The density of core per rack unit is not high enough.

JotaMG
Guest
JotaMG

@CampGareth
Thanks, that’s what I suspect, prices for Arm should be lower than for x86.

Jeroen
Guest
Jeroen

What do they mean with 1 Flexible public IPv4?

GanjaBear
Guest
GanjaBear

Let’s see some benchmark numbers, shall we?

nutjob2
Guest
nutjob2

@JotaMG
Don’t jump to conclusions too quickly. Some x86 “cores” are actually threaded cores, not physical cores. Some providers also further split cores to make them all perform the same on machines with different levels of CPU performance. I’m not sure what these people do but the only way to really know is to benchmark. You also have to benchmark the things that matter for the software you’re running, since each system will have differing memory, disk and network performance, each one of which may be the determining factor for your setup.

blu
Guest
blu

Re performance of the new servers, here’s something posted on the Scaleway forums by user Bob:

I did some sysbench CPU test runs (10000 prime numbers). It’s a very basic test, but here are some ballpark figures:

x86-64 “Starter” (VC1), Atom Avoton C2750, 10000 primes in 9s on one core
ARMv7 “BareMetal” (C1), 10000 primes in 258.5s on one core
ARMv8 “Starter” (ARM64), 10000 primes in 6s on one core
x86-64 “Intensive”, not tested

This is a very parallelisable task, so e.g. 2 cores took half the time of 1 core. I suspect one of the operations used was particularly bad for the ARMv7 core, since the result is very poor.

Of course, you must consider your specific use case – this is a very simple benchmark. You can test yourself at the low €/hour rate Scaleway charges…

https://community.online.net/t/x86-64-armv7-vs-new-armv8-cpu-performance/4503/2

tkaiser
Guest
tkaiser

@blu
When do people stop using such moronic anti benchmarks like ‘sysbench’? This thing runs in L1/L2 cache only (great if your application does not and is affected by memory throughput) and calculating prime numbers is also nothing that relates to any real workload!

This sysbench anti benchmark finishes stuff magnitudes faster when running on some platforms (that’s why ARMv7 looks that bad above) since it’s not a hardware test but one for compiler optimizations. And when you use sysbench on exactly the same hardware numbers with Ubuntu Xenial are 30% better than those made with Debian Jessie when using distro packages since… GCC 4.9 vs. 5.4 (yes, that’s all — really).

And if you want to lower your scores just use sysbench 0.5 instead of the widely used 0.41 everywhere: performance now drops by 10 percent. So if you upgrade the benchmark it reports your hardware got slower!

TL;DR: sysbench is crap for various reasons to measure CPU performance. It’s great to learn about compiler optimizations though. And never trust in any sysbench numbers on the net unless sysbench version and GCC version is known! Otherwise numbers are not comparable.

tkaiser
Guest
tkaiser

– “openssl speed rsa2048 -multi ”
– 7z b
– minerd –benchmark

and you’re done to get a rough estimate about CPU horsepower in short time (focussing on important stuff like integer, float, memory performance so you already have an idea how/whether the numbers spitten out by these tools will correlate with your application!). Using distro packages you know then what get if you’re lazy. Firing up the compiler with appropriate settings and you might end up with numbers twice as good (for free, just by letting the compiler do its job optimizing for the platform in question)

But on such servers IO bandwidth/latency is also important so another round with fio (try out several mixed read/write loads) and ioping is necessary.

Using sysbench’s cpu test is only fooling yourself (unfortunatly that’s done everywhere, same with dd and hdparm for ‘storage benchmarks’)

blu
Guest
blu

@tkaiser
I’m of the opinion that no synthetic benchmark is definitive, and also that no synthetic benchmark is absolutely useless – they all demonstrate *some* aspect of the performance of the system, where the latter often includes the compiler, as that is often insurmountable – real sw does get largely affected by the quality of the compiler and its settings. I’m far from the thought that a prime-number generator is indicative of the overall performance of the system, or, say, a particular task other than the one being measured ; )

Pablo
Guest
Pablo

I did a very basic speed comparison with a notebook (an apple macbook pro with an old I5) running my java code. ARMv7 “BareMetal” (C1) has around 70% of the performance of my macbook. HOWEVER this is not a real speed test. Perhaps the JVM on ARM is more efficient, perhaps their SSD are better than my SSD. Perhaps Ubuntu 16 is faster than Sierra, etc, etc. Both machines are too different and my “test” is too cheap to use it as a serious comparison.

3 euros for 70% of the speed of my little apple is good enough for me and my server though. I suggest you test it. You can buy some hours and run some tests.