Review of an ARM-based cloud server
As a web developer I have the opportunity to work with a lot of different hosting providers. The good, the bad and the ugly. In this post I’m sharing my latest experiment with an ARM-based (as opposed of traditional x86) cloud based hosting provider: Scaleway.
Background: ARM vs x86
ARM and x86 are two different architectures for Computer Processor Unit (CPU)
ARM CPU are used in most mobile devices today. While x86 continues to dominate the compute-intensive usages: workstation and server processor. Using ARM processor for server is like asking the question: Can we use a lot of those cheap and power-efficient processors to replace a big and complex CPU AND deliver the same service?
There are essentially 2 manufacturers for x86: Intel and AMD. But Intel have around 80% market share in total (laptop, desktop and server) and even more when we look only at the server market. Very concentrated.
On the other hand there are currently a lot of designers, manufacturers for the ARM CPU: and among them some some big names: Apple, AppliedMicro, Broadcom, Cavium, Nvidia, Qualcomm, Samsung and others… Quiet a competitive environment.
ARM has been very successful in mobile devices because it is cheap and and power-savy. Those criteria are more important than raw computing power in this area. But with the thriving market of smartphones and other portable devices those last 10 years (Iphone 1 was released in 2007) the ARM architecture has evolved a lot: clock speed increased, multi-cores… Now with the 8th generation (ARMv8) the CPU went from 32 bits to 64-bits, bringing it closer to the current architecture and power of the x86. Not here yet but on it’s coming fast.
On the software side, most of the Linux distribution and other critical projects (for the web) (webserver, browser, database, hypervisor…) are already available and supported on ARM. That’s mean that from a enduser point of vue: it doesn’t matter if the underlying CPU is ARM or x86, all the important software of the web ecosystem can run on it.
So now, one step further can we or can we not safely and efficiently replace x86 server by an ARM servers? If we can we can expect to reduce hosting cost because ARM chip are both cheaper to build and to run (per MIPS).
Scaleway is the first cloud hosting provider, to my knowledge, to offer ARM-based server. And it is cheap. Let’s try to create a Virtual Private Server (VPS) and configure it to run a website.
The test
The admin interface is nice and sleak. No further comment needed on that.
Creating a VPS is straight forward. Most of the VPS and other products are based on x86. ARM is just a recent addons to their offer. There are currently 7 ARM-based VPS: from 4 cores to 64 cores ! (and from €2.99/month to €279.99/mo ). In this test I will review the 4 cores and the 8 cores. All are based on ARMv8 64bits SoC by Cavium ThunderX.
After the account is created and credit card entered, You choose your server and it will be created in less than 2 minutes.
You choose the OS among 9 differents Linux distro officially supported (Ubuntu, Debian, Fedora…) + the community provided OS: some more specific image: wordpress, node.js ; or even Docker!
You will need to register your ssh public key in the admin interface before the server creation (use to grant you ssh access to the VPS). A (optional and paid) public IP can be affected to your VPS.
I choose to install Ubuntu 16.04 + the ordinary stack to powered a Drupal website:
- Apache 2.4
- MariaDB 10.0
- PHP 7.0
- Drupal 8.5
et voilà!
No issues, no difference between the configuration of this LAMP stack and the configuration of a LAMP stack on x86. And performance, admin pages are working fine. But this is just test with 1 user (me), let’s try to do a load test and compare it to another server.
How does it compare to an AWS EC2 instance?
I choose to compare the Scaleway servers with a very popular server EC2 of AWS. The comparison included 4 real life servers: 2 of each types:
2 ARM virtual servers: based on Cavium ThunderX
- 4 ARMv8 : Scaleway ARM64–2GB: cores: 4 ARMv8 ; memory: 2GB ; SSD disk: 50 GB — $3.51/mo $0.0071/hr (€2.99/mo €0.006/hr)
- 8 ARMv8: Scaleway ARM64–8GB: cores: 8 ARMv8 ; memory: 8GB ; SSD disk: 200 GB — $14.09/mo $0.028/hr (€11.99/mo €0.024/hr)
2 x86 virtual servers: based on Intel Xeon
- t2.small: AWS EC2 t2.small: core: 1 vCPU ; memory: 2GB ; SSD disk: 8GB $19.8/mo $0.0264/hr
- t2.medium: AWS EC2 t2.medium: core: 2 vCPU ; memory: 4GB ; SSD disk: 8GB $39.6/mo $0.0528/hr
Those servers are in the low to middle-end for each company. Bear in mind that the price differ a lot, but the goal is just to give an idea of what value you can get for your money.
Note: cost at the time of publication, Scaleway Euro price converted to USD. For AWS I used Paris datacenter pricing, to have a better comparison with Scaleway (based in Paris too).
Simple benchmark: sysbench
sysbench is a lightweight benchmarking tool for linux. It supports testing CPU (multithreading), memory, file I/O, and other stuff. It is available as a package in all popular linux distribution. (But in that case I’vE downloaded the binary package directly from the vendor website to use the latest version: (1.0.14 on arm, 1.0.15 on x86).
The results:
Summary
Commands used:
- CPU:
sysbench cpu --threads=8 run - File I/O: random read:
sysbench fileio --threads=16 --file-test-mode=rndrd preparesysbench fileio --threads=16 --file-test-mode=rndrd run
sysbench fileio --threads=16 --file-test-mode=rndrd cleanup - Memory:
sysbench memory --threads=16 run
Summary
Those 2 Multi-core ARM server seem to perform clearly better CPU-wise and memory-wise for those indiviudal tasks. It is not a surprising results: 8 cores is doing more operation than 2. It is interesting to notice that I/O are still much better for the AWS servers. Although I think it is more related to the fact that SSD disk powering those servers are more efficient.
global benchmark: ApacheBench
95% of the servers I have installed are used to host a website. They combine a HTTP web server (eg: Apache), a database (MySQL), a PHP framework (or CMS) and some custom PHP code: this is a LAMP stack. In this scenario the global performance of the entire stack is important. It is the global performance that will defined the response time of the server and the capacity to handle heavy traffic.
So in this section I’m using ApacheBench (ab): with this tool we can send to the server a lot of HTTP requests in a short period of time, simulating big traffic on the website, and measure the average server response time. This tool is not perfect because it won’t represent exactly “real life” traffic but it is ok to compare different servers like we are doing here:
We are requesting with ab the “List of content” admin page in Drupal. In this test website I have created 50 test pieces of content to populate the page.
ab -n100 -c4 -v1 -C 'SESSYYYYYyyyy=ZZZZZzzzz' http://<SERVER_IP>/d8/admin/contentCommand sent
In this commands:
- n is the total number of request to send (different scenario, see below)
- c is the concurrency: the nb of request simultanously. (different scenario, see below)
- C ‘SESSYYY=zzz’: is the session cookies to have access to the page.
The results:
Summary
EC2 instances perform slightly better. We can clearly see that the 2 cores of t2.medium are performing better than the 1 core of the t1.small. But for the 8 cores ARM server is not performing better than the 4-cores ARM server. It is possible that in its current configuration the LAMP stack is not allowing up to 8 HTTP requests to be handle simultaneously… This need to be investigate more, server configuration need to be fine-tuned.
Conclusion
It seems to me that those ARM-based server by Scaleway are working very well to power websites. In fact it doesn’t make any difference for me to create and configure those servers as opposed as a traditional x86. And as a developer/devops this is important for me: this is encouraging technology adoption. Of course, this was just a short test and we will need more time to make sure it doesn’t come with drawbacks. And I’d like to see other cloud provider entering the market.
But, and this is the main point, if it turnout that those processors are significantly cheaper to produce and to run (because less electricity consumption) for the same MIPS (million instructions per second) then it will mean that hosting cost will be reduced. And this is good for all!
Note: The content was initially published on my personal blog.
