If you are going to write a big article on a 'problem', then it would be a good idea to spend some time explaining the problem, perhaps with some scenarios (real world or otherwise) to solve. Instead, this article just leaps ahead with a blind-faith 'we must do this!' attitude.
That's great if you are just toying with this sort of thing for fun, but perhaps worthless if you are advocating a style of server design for others.
Also, the decade-ago 10k problem could draw some interesting parallels. First of all, are machines today 1000 times faster? If they are, then even if you hit the 10M magic number, you will still only be able to do the same amount of work per-connection that you could have done 10 years ago. I am guessing that many internet services are much more complicated than a decade ago...
And if you can achieve 10M connections per server, you really should be asking yourself whether you actually want to. Why not split it down to 1M each over 10 servers? No need for insane high-end machines, and the failover when a single machine dies is much less painful. You'll likely get a much improved latency per-connection as well.
Comments
If you are going to write a big article on a 'problem', then it would be a good idea to spend some time explaining the problem, perhaps with some scenarios (real world or otherwise) to solve. Instead, this article just leaps ahead with a blind-faith 'we must do this!' attitude.
That's great if you are just toying with this sort of thing for fun, but perhaps worthless if you are advocating a style of server design for others.
Also, the decade-ago 10k problem could draw some interesting parallels. First of all, are machines today 1000 times faster? If they are, then even if you hit the 10M magic number, you will still only be able to do the same amount of work per-connection that you could have done 10 years ago. I am guessing that many internet services are much more complicated than a decade ago...
And if you can achieve 10M connections per server, you really should be asking yourself whether you actually want to. Why not split it down to 1M each over 10 servers? No need for insane high-end machines, and the failover when a single machine dies is much less painful. You'll likely get a much improved latency per-connection as well.