Nucleon is a dynamic TCP load balancer written in Rust. It has the ability to insert and remove backend servers on the fly. To do that it leverages Redis Pub/Sub mechanism. Adding or removing a server to a cluster is as easy as publishing a message to Redis.
All you need to build it is Rust 1.3.
Just go in the repository and issue:
$ cargo build --release
Nucleon can be used with or without a Redis database. When ran without Redis it is not possible to add or remove load balanced servers without restarting the process.
Usage:
nucleon [OPTIONS] [SERVER ...]
Dynamic TCP load balancer
positional arguments:
server Servers to load balance
optional arguments:
-h,--help show this help message and exit
-b,--bind BIND Bind the load balancer to address:port (127.0.0.1:8000)
-r,--redis REDIS URL of Redis database (redis://localhost)
--no-redis Disable updates of backend through Redis
-l,--log LOG Log level [debug, info, warn, error] (info)
Imagine you have two web servers to load balance, and a local Redis. Run the load balancer with:
nucleon --bind 0.0.0.0:8000 10.0.0.1:80 10.0.0.2:80
Now imagine that you want to scale up your infrastructure by spawning a new web
server at 10.0.0.3. Just send a message to the Redis channel backend_add:
redis:6379> PUBLISH backend_add 10.0.0.3:80
(integer) 1
Your will see in logs:
INFO - Load balancing server V4(10.0.0.1:80)
INFO - Load balancing server V4(10.0.0.2:80)
INFO - Now listening on 0.0.0.0:8000
INFO - Subscribed to Redis channels 'backend_add' and 'backend_remove'
INFO - Added new server 10.0.0.3:80
If you decide that you do not need server 2 any longer:
redis:6379> PUBLISH backend_remove 10.0.0.2:80
(integer) 1
Surprisingly well. A quick comparison with HA Proxy in TCP mode with a single backend containing a single server using iperf results in:
| Connections | HA Proxy | Nucleon |
|---|---|---|
| 1 | 15.1 Gbits/s | 15.7 Gbits/s |
| 10 | 13.5 Gbits/s | 11.3 Gbits/s |
| 100 | 8.9 Gbits/s | 10.5 Gbits/s |
Keep in mind that this is a really simple test, far from what real life traffic looks like. A real benchmark should compare short lived connections with long running one, etc.
MIT