[Spce-user] virtualizing sip:provider CE

[Christian Rohmann]

On 05/17/2016 04:31 PM, Andreas Granig wrote:
> The number of regs it can do without calls is somewhere around 80
> regs/sec, but if you add some calls, it degrades quite quickly.

Did you further narrow down what the reason for the current performance
limits is? Some profiling of where the bottleneck actually is, so which
component maxes out first, could surely help to scale or improve the
right components (software or hardware).

Especially all those (incoming) registrations, which currently go
persistently into a relational (MySQL) database, are quite expensive for
data (or better called "state") that could be recreated at any time
would the whole cluster ever crash or explode. Just restart and have the
clients re-register ... nothing of value (actual call state or even
CDRs) would be lost.

Cluster-wide replication via the likes of redis or hazelcast or some
other replicated, in-memory database might completely remove this
bottleneck of registrations and have NGCP scale with the actual calls
which are handled (and worth having their state stored persistently
somewhere).



Regards

Christian