Impressive! One thing that caught my eye was:

> In the process, we also optimized our XML parser, released now as Fast XML, a high-performance, memory efficient Expat-based Erlang and Elixir XML parser

The shout out to Elixir is interesting. I glanced through the linked GitHub code and as far as I can tell nothing is written in Elixir, although it's still a trivially true statement since Elixir can easily use Erlang libraries. I guess Elixir is a big enough force in the Erlang community now that they want to highlight that their software works fine for Elixir users as well!

Actually, Elixir support comes from API changes to be more Elixir friendly (for example order of parameter to be "pipe" friendly). This is also about submitting the component on hex.pm package manager to be easily integrated in Elixir mix tool. And finally, this is about writing some part in Elixir. For some projects we are even starting writing code directly in Elixir.

No surprise here.

Whatsapp is based in Ejabberd.

Whatsapp has repeatedly documented how they achieve 2-3M connections per node.

http://www.erlang-factory.com/upload/presentations/558/efsf2...

I believe they use erlang, but not ejabberd.

Whatsapp uses a modified version of ejabber. Ejabber is written in erlang.

http://highscalability.com/blog/2014/2/26/the-whatsapp-archi...

They started with a version of ejabberd of some sort and migrated to more custom components over time IIRC. I can't find the link right now. (Keep in mind that ejabberd is written in Erlang.)

Can someone point me to information on how to tune the linux kernel or OS to handle such large amounts of TCP connections?

I encountered problems with the default kernel configurations during the writing of my bachelor thesis and wasn't able to really read up on what could have been done to solve those problems. For example i observed high cpu usage only on the first core (similar to what the author describes in the article) due to a high amount of network interrupts. I tried to make those interrupts to be handled by all cores, which was neither recommended (as i remember) nor did it really help.

In previous work [1] we followed the Linux kernel documentation [2] and the guidelines published by IBM ITSO [3].

[1] http://link.springer.com/article/10.1007/s00607-014-0394-9

[2] http://www.cyberciti.biz/files/linux-kernel/Documentation/ne...

[3] https://lenovopress.com/redp4285.pdf

I suggest you look for slides from the Whatsapp folks on the kernel and Beam vm tuning they did to get similar numbers. Don't have the link handy but fairly easy to find.

What is more problematic IMHO is generating the load to saturate said server.

IIRC, the Whatsapp folks use FreeBSD.

We manage to load the server properly with Tsung (https://github.com/processone/tsung) but for a more heavy load we may need to add one more Tsung server.

Some info here http://stackoverflow.com/questions/22090229/how-did-whatsapp...

Not a surprise that just moving packets from one socket to another is fast. In 2016s you need to store messages on server, do big file transfers. Presence need to be update every couple seconds. (like whatsapp) You need to have large groups (500+ members) for your tests.

How it will work then?

You will know soon, as we are working on such a higher load benchmark. From existing production service thought we expect very good result, especially as the test server was not fully loaded.

Would be nice if they documented if they used TLS, and if so, how that was configured. Even if the 28kb per user is a typo and they mean 28kB then that would be impressive with TLS enabled.

This benchmark was done without TLS because we wanted to demonstrate the performance of ejabberd and not the performance of OpenSSL itself. TLS is still something you can offload to a load balancer however.

I updated the post to fix typo and mention that test was not over TLS.

Similar to other Erlang implementations like Pheonix I guess http://www.akitaonrails.com/2015/10/29/phoenix-experiment-ho...

Phoenix is Elixir not Erlang. Although they both run on the BEAM VM.