You can usually cut the cold start performance in half by enabling Tiered Compilation. It's just one parameter in your project file, or via an environment variable.
Disclaimer: I work on performance in the ASP.NET team.
I should say (since i've got your attention) that i'm almost 100% sure that all of your customers would be more than willing to sacrifice a startup time for a well-coordinated callback when everything is warmed up, as a toggleable option of course.
To explain this a bit better, i'd definitely be happy even waiting 10+ minutes for all of my controllers/hub methods to be instant response. I could then subscribe to an event when it is ready and let haproxy/nginx or whatever else know that the server is "good to go". This way, the server goes into the pool, and everyone is happy. I remember seeing that the Bing guys did something like this, but again it was majorly hacky.
Thanks for all the great work you're doing on C#/.NET Core in general by the way.
Tiered Compilation is tweaked in a way that it provides a very good balance between startup time and throughput, which is what a majority of users would need. If you want to privilege throughput and don't care about startup time, then you should try COMPlus_ReadyToRun=0 as an environment variable. This will force to use non-crossgened code so the JIT can better optimize your code.
To explain this a bit better, i'd definitely be happy even waiting 10+ minutes for all of my controllers/hub methods to be instant response. I could then subscribe to an event when it is ready and let haproxy/nginx or whatever else know that the server is "good to go". This way, the server goes into the pool, and everyone is happy. I remember seeing that the Bing guys did something like this, but again it was majorly hacky.
Thanks for all the great work you're doing on C#/.NET Core in general by the way.