Re: [mosquitto-dev] Advice needed (long, sorry)
Thanks guys for your advice,
- we will start working on AWS from our 2nd phase instead of the 3rd one.
- we will return to a multi-instance mosquitto infrastructure behind a load balancer and try to make our best to force the client to connet to the desired broker using an initial http request.
- we hope that the load balancer itself does not become the bottleneck now.
- we continue tuning linux kernel to scale better.
Thank you for all. Be sure that we will document everything better to be able to give you our results.
Nice to meet you "virtually" :-)
El 12/02/2015 19:36, "Manuel Domínguez Dorado" <manolodd@xxxxxxxxx
Hi, attached you have sysctl.conf and mosquitto.conf files that we used for that.
Changes to /etc/rc.d/rc.localdefrt=`ip route | grep "^default" | head -1`
ip route change $defrt initcwnd 10
And changes to /etc/security/limits.conf* soft nofile 100000
* hard nofile 100000
And changes to /etc/pam.d/sshd
session required pam_limits.so
And changes to /etc/ssh/sshd_config
Sources we read to tune the aforementioned paramenters:
And... that's all. We configured dozens VM (20-30... I don't remember, but more or less), each one running a java software that creates hundreds of threads each one of them creating a paho client and connecting to the broker over TLS. No subscriptions and no publishing. When we started publishing everything went wrong. Everything on our servers using gigabit LAN.
Sorry (and sorry also Stefano), we did the test without any other planification as we did lots of tests of different types at that moment. As we realized that we were far of 100K connections we tried other strategies. So we have no a very formalized and structured documentation, nor charts, neither a dataset... :-( I remember that we achieved 23.2xx connections.
Hope it helps.