Scaleway ARMv8 Servers

M_Ordinateur

Hello everyone!

Scaleway (Online) has just announced ARMv8 servers with Cavium ThunderX SoCs.

The cheapest one is the same price as the C1, 2.99€/month. The bandwidth seems unlimited at 200Mbit/s.

Here are the prices they quote on the announcement email:

• ARM64-2GB - 2x ARMv8 cores - 2GB Memory - 50GB SSD - €2.99/mo
• ARM64-4GB - 4x ARMv8 cores - 4GB Memory - 100GB SSD - €5.99/mo
• ARM64-8GB - 8x ARMv8 cores - 8GB Memory - 200GB SSD - €11.99/mo

What do you think about it?

WSS

As much as I really respect ARM getting a second chance at existing- they're still dogshit slow.

rm_

2 cores of this should be a bit faster than 2 cores of C2750. It would be very interesting to see a geekbench of both. (and they are not)

However Scaleway STILL doesn't have a proper IPv6 deployment setup (only one IPv6 address, which is dynamic, i.e. changes as you start and stop the server).

vfuse

At least it's 64bit now, wondering what performance is like since it's not stating that these are dedicated cores.

time4vps

Anyone has bench results?

WSS

@rm_ said:
However Scaleway STILL doesn't have a proper IPv6 deployment setup (only one IPv6 address, which is dynamic, i.e. changes as you start and stop the server).

Really? I thought that was just an issue with the C1 I briefly spun up.

rm_

vfuse said: not stating that these are dedicated cores

Those ThunderX are pretty crazy up to 48-core CPUs http://www.cavium.com/ThunderX_ARM_Processors.html
so the oversell ratio for CPU should be lower (if any) than at the x86 VPS counterpart.

vimalware

Now they should just add mountable HDD block storage, in addition to SSD.

sin

I'm booting one up right now, will try to do some benches

M_Ordinateur

I have just booted one, tell me which benchmark you'd like me to run

jgillich

Here is a benchmark of the ThunderX: https://www.servethehome.com/exclusive-first-cavium-thunderx-dual-48-core-96-core-total-arm-benchmarks/

Like @WSS said, single core performance is extremely bad. The dedicated C1 made sense, but this? A single Xeon core runs circles around 8 ThunderX cores.

ehab

@ M_Ordinateur

http://geekbench.com/

brueggus

rm_ said: only one IPv6 address, which is dynamic, i.e. changes as you start and stop the server

M_Ordinateur

@ehab said:
@ M_Ordinateur

http://geekbench.com/

I get
-bash: ./geekbench_x86_64: cannot execute binary file: Exec format error. I'm not sure Geekbench 4 supports ARM

WSS

Did you seriously just try to run an obvious x86 binary under ARM?

rm_

jgillich said: A single Xeon core runs circles around 8 ThunderX cores.

From the scaleway perspective you need to compare this vs Atom C2750 which their x86 2.99 EUR offer is based on. There are only 2 benchmarks in that URL that also have a similar Atom (C2758), but those seem to be hard computational ones, i.e. not a typical VPS/Web scenario. IOW it might be not as bad, or can be even better in certain cases.

vfuse

I think this should work for ARM: https://github.com/kdlucas/byte-unixbench/tree/master/UnixBench

sin

I'm running that serverscope.io thing, it's running fio right now and then will do unixbench after...here's the upload speeds:

Testing upload speeds
  1. iperf.fr (Rouen):                            ......................... 366.87 Mbit/s
  2. Ikoula (Reims):                              ......................... 331.76 Mbit/s
  3. LaFibre.info (Douai):                        ......................... 361.91 Mbit/s
  4. MyTheValentinus (Roubaix):                   ......................... 240.50 Mbit/s
  5. media-upload.net (Roubaix):                  ......................... 61.75 Mbit/s
  6. ePlay TV (Roubaix):                          ......................... 27.79 Mbit/s
  7. Magic-VPN (Roubaix):                         ......................... 124.44 Mbit/s
  8. Matthews Tech (Gravelines):                  ......................... 130.99 Mbit/s
  9. CloudConnX (Eastbourne):                     ......................... 49.75 Mbit/s
  10. Universite Catholique de Louvain (Louvain-La-Neuve):......................... 214.70 Mbit/s
  11. Combell (Brussels):                         ......................... 65.10 Mbit/s
  12. Orange Belgium (Brussels):                  ......................... 78.54 Mbit/s
  13. Riffle Media BVBA (Brussels):               ......................... 62.80 Mbit/s
  14. Proximus (Brussels):                        ......................... 375.14 Mbit/s
  15. SKSGroup (Brussels):                        ......................... 120.67 Mbit/s
Average upload speed is 174.18 Mbit/s
Running dd as follows:
  dd if=/dev/zero of=benchmark bs=64K count=32K conv=fdatasync
32768+0 records in
32768+0 records out
2147483648 bytes (2.1 GB) copied, 12.0565 s, 178 MB/s

sin

unixbench (via serverscope) failed for me but here's the rest of the results:

https://serverscope.io/trials/YazY

jgillich

Here are some benchmarks: https://news.ycombinator.com/item?id=14211611

Seems like a pretty big jump from the old C1.

sin

@jgillich said:
Here are some benchmarks: https://news.ycombinator.com/item?id=14211611

Seems like a pretty big jump from the old C1.

Yeah it feels a hell of a lot snappier to me then the old C1, not to mention I'm getting much better network and disk speeds then I was getting on the C1.

default

At this speed, by the time they solve IPv6 problems, the market will have IPv8. How about solving that IPv6 addressing first, and then come up with new offers?!

WSS

@default said:
At this speed, by the time they solve IPv6 problems, the market will have IPv8. How about solving that IPv6 addressing first, and then come up with new offers?!

We're talking about Scaleway, not ColoCrossing.

M_Ordinateur

@sin said:
unixbench (via serverscope) failed for me but here's the rest of the results:

https://serverscope.io/trials/YazY

Here's a ServerScope with Unixbench: https://serverscope.io/trials/46x9

407.8 on one CPU

rm_

Found more feedback on IPv6 at their forum: https://community.online.net/t/horrible-ipv6-setup-on-vc1s-c2/2111/12

Really, whoever worked on this, they did the lousiest job possible (it's impossible to do any worse even if you tried), but probably still got paid and marked the "IPv6" checkbox as "done". After that there are still no improvements or changes for over a year, despite a staff member promising in April 2016 that there will be.

sin

rm_ said: After that there is still no improvements or changes for over a year, despite a staff member promising in April 2016 there will be

They seem to be half-assing a lot of things.

jgillich

M_Ordinateur said: 407.8 on one CPU

Vs 744 for the C2 Atom: https://paste.ee/p/MmKRJ

So yea, unless ARM is a requirement, I don't see a usecase for these. Still great to see ARM getting foothold in the server market to create some competition for Intel.

jeromeza

How would these compare to the Kidechaire in terms of CPU?

I need a cheapie box to run my mail on (personal, nothing business critical)

Kidechaire CPU: VIA® Nano® U2250 1.6 GHz, x64, VT

oneilonline

I wouldn't recommend any enterprise or production level be run on an ARM CPU, it's too slow.

jgillich

jeromeza said: How would these compare to the Kidechaire in terms of CPU?

Seems like the Nano has a similar Unixbench score to one ThunderX core. I'd still go for the Atom based VPS, not only is it faster, but it also has broader software support.

rm_

jeromeza said: How would these compare to the Kidechaire in terms of CPU?

Well since benchmarks show after all that the x86 countepart is almost 2x faster for the same price, there seems to be no reason to pick ARM at this point. Maybe in the future when they figure this out and will lower prices of the ARM plans.

Still I would not recommend buying Scaleway at all until they get IPv6 right. If anything, you could go with the OVH VPS-SSD, same 1xIPv6, but at least it's a static one.