Online.net is a French hosting company offering the usual range of services from simple web hosting for 1 € / mo to dedicated servers, but with a twist I haven't seen before: their offer of dedicated servers is incredibly affordable! See this example: for 6 € / mo you can get yourself a dedicated VIA Nano x86 CPU powered server completely for your own usage. These CPU's performance is comparable to those of Intel Atom, so the prices are extremely attractive.
They are also testing a line of dedicated machines based on a quad-core ARM CPU which should be even more efficient, and which they have offered for completely free public testing on-demand in 15-minutes intervals! I have made use of this opportunity and ran some benchmarks on the provided "C1" test configuration. I have used the OpenSSL's "speed" benchmark, the bonnie++ disk benchmark and my own pretty much ad-hoc pysysbench.
The hardware is Marvell ARMADA XP and the ARM machines run Linux (of course). You can inspect their dmesg and cpuinfo data yourself. While I was testing the machine, I also tried to inspect and change the CPU's frequency, but it appears that the cpufreq driver is not available for this specific product (not that it matters directly for the machine's end-users since they are not paying for the electricity, but let's not be wasteful).
Since I've recently benchmarked a modern Atom Z3775 CPU, I'm including the single-core OpenSSL results from that benchmark for comparison here:
|CPU||Xeon E3-1230 v3||Xeon E5405||Atom Z3775||ARMADA XP|
|Norm. freq.||3.3 GHz||2 GHz||1.46 GHz||1.3 GHz|
|Turbo freq.||3.7 GHz||2 GHz||2.39 GHz||1.3 GHz|
|RC4||865 MB/s||213 MB/s||212 MB/s||107 MB/s|
|SHA256||275 MB/s||60 MB/s||43 MB/s||50 MB/s|
|AES-128-CBC||164 MB/s||106 MB/s||94 MB/s||54 MB/s|
|RSA2048||987 s, 31090 v||73 s, 2906 v||36 s, 1303 v||29 s, 935 v|
The results are a bit underwhelming. This is a 32-bit ARM CPU with 2 GB of RAM, but it's operating on a fairly slow frequency of 1.3 GHz and the results reflect this. However, it seems like the CPU's efficiency per MHz is very close to that of the latest Intel Atoms. It looks like if the ARM could be doubled in frequency, it could even surpass the Atom.
Here are the 4-core OpenSSL results:
|Benchmark||ARMADA XP 4-core|
|RSA2048||115 s, 3734 v|
The test setup boots and uses a network block device (NBD, think iSCSI-lite) for storage, so the bonnie++ benchmark results are fairly low:
Version 1.97 ------Sequential Output------ --Sequential Input- --Random- Concurrency 1 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks-- Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP /sec %CP c1-10-1-10-51 4G 111 98 76118 89 35709 28 574 99 85026 22 1589 92 Latency 80619us 355ms 348ms 15835us 13792us 41985us Version 1.97 ------Sequential Create------ --------Random Create-------- c1-10-1-10-51 -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete-- files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP 16 7813 80 +++++ +++ 9626 62 7939 78 +++++ +++ 10896 71 Latency 413us 1794us 1959us 455us 63us 491us
The PySysBench results are here:
Running 8 benchmarks on 4 threads for 10 seconds each (NCPU=4) Running benchmark: Warmup-Ignore ... score: 197.0 average: 49.2 Running benchmark: Hash-SHA256 ... score: 196.7 average: 49.2 Running benchmark: Hash-SHA512 ... score: 94.1 average: 23.5 Running benchmark: Zlib-Compress ... score: 13.8 average: 3.4 Running benchmark: Zlib-DeCompress ... score: 120.7 average: 30.2 Running benchmark: Socket-OneSock ... score: 71.9 average: 18.0 Running benchmark: Socket-Syscalls ... score: 90.8 average: 22.7 Running benchmark: Socket-OnePipe ... score: 77.9 average: 19.5 Total score: 665.8
You can compare these results with those of other systems here.
In conclusion - do you need this?
Running virtual private servers "in the cloud" has the benefit of hardware isolation which is important on at least two levels: firstly, the virtual machines can be moved around in case of base hardware failures, so that the client experiences very low downtimes. It is not so clear that large-scale physical hosting operations can be as agile. Secondly, if the storage is attached directly to provided physical systems (which is NOT the case with the test servers, but might be the case in other offers by Online.net), it looks like it's non-redundant and requires some special handling with regards to backing it up. OTOH, this might be a benefit if you specifically do not want to share storage with other users (for whatever reasons - security is probably not among them since the hardware is still under remote control). But I'm excited a service such as this exists.