One of our customers serves north of 1 mio overall provisioned subscribers on two servers with this software, so it's quite scaleable vertically already. It's easier when UDP is used though, because with TLS things add up quite quickly.
To go big, we've horizontal scaling mechanisms using subscriber partitioning by load-balancing SIP and provisioning requests over multiple pairs of such servers (usually placed in blade-center servers).
The key here is to keep as much CPU heavy things like media relaying end-to-end where possible, because the signaling part is pretty light-weight in SIP. To scale out and keep reliability up while keeping complexity low, we have a shared-nothing approach wherever possible. Works well for us.
To go big, we've horizontal scaling mechanisms using subscriber partitioning by load-balancing SIP and provisioning requests over multiple pairs of such servers (usually placed in blade-center servers).
The key here is to keep as much CPU heavy things like media relaying end-to-end where possible, because the signaling part is pretty light-weight in SIP. To scale out and keep reliability up while keeping complexity low, we have a shared-nothing approach wherever possible. Works well for us.