Ad b: This is not true. If you assume that the keys grow polynomially with n, let's say n^3, then just let the number of buckets in the radix-sort grow linearly (i.e., use radix n instead of some fixed radix). Then the number of passes will be equal to the degree of the polynomial, so in the cubic case three passes will be sufficient, leading to a linear-time algorithm again.
Radix-sort stops being usable if the integers grow superpolynomially, but then you can use a bunch of other techniques (see papers on sorting integers by Mikkel Thorup), leading to complexities on the order of n*log log n or even better.
As for "human-friendliness": It's much more useful for a human to have negative temperatures denote "it's freezing outside" than "it's colder than Mr. Fahrenheit used to experience":-)
I agree that Fahrenheit temperatures might have a more useful range for everyday life, but it's zero and 100 points don't make any sense to normal people.
I think that everybody would be happy if they released not a driver, but specs to their HW. Many people would be happy to write a driver and that wouldn't need a licence from third parties, maybe except for patents, who knows in the insane US patent system.
Networking protocols with proper congestion control usually require precise timing, which is much harder to achieve in user space. The same case as TCP.
Well, such arguments don't say much – everything you can do over IP, you can do over UDP. You can implement SCTP over UDP, as well as you can do TCP over UDP. However, that doesn't mean it makes sense to do so:)
People are already doing similar things over UDP for years and in the vast majority of cases they get it wrong (poor performance, broken congestion avoidance algorithms and so on). That seems to suggest that the problem is not so trivial as it looks and that it makes sense to finally find a good solution and standardize it.
You would be right if the people cared whether a domain is registered or not. They usually want to know whether they can access the domain (send mail to it etc.), that is to know whether it is present in the DNS. And for that purpose, the DNS protocol is the only right tool.
Slashdot Mirror
If they are interested in these architectures, they should help GCC folks maintain them instead of resurrecting obsolete compilers.
Ad b: This is not true. If you assume that the keys grow polynomially with n, let's say n^3, then just let the number of buckets in the radix-sort grow linearly (i.e., use radix n instead of some fixed radix). Then the number of passes will be equal to the degree of the polynomial, so in the cubic case three passes will be sufficient, leading to a linear-time algorithm again.
Radix-sort stops being usable if the integers grow superpolynomially, but then you can use a bunch of other techniques (see papers on sorting integers by Mikkel Thorup), leading to complexities on the order of n*log log n or even better.
As for "human-friendliness": It's much more useful for a human to have negative temperatures denote "it's freezing outside" than "it's colder than Mr. Fahrenheit used to experience" :-)
I agree that Fahrenheit temperatures might have a more useful range for everyday life, but it's zero and 100 points don't make any sense to normal people.
I think that everybody would be happy if they released not a driver, but specs to their HW. Many people would be happy to write a driver and that wouldn't need a licence from third parties, maybe except for patents, who knows in the insane US patent system.
With the kernel changes Van Jacobson suggests, it's likely, but I very much doubt that you can do well with the current socket interface.
Networking protocols with proper congestion control usually require precise timing, which is much harder to achieve in user space. The same case as TCP.
People are already doing similar things over UDP for years and in the vast majority of cases they get it wrong (poor performance, broken congestion avoidance algorithms and so on). That seems to suggest that the problem is not so trivial as it looks and that it makes sense to finally find a good solution and standardize it.
Actually, that's not true in many cases -- GCC 3.x generates very good code, but the 4.x versions still haven't caught up with the 3.x line.
You would be right if the people cared whether a domain is registered or not. They usually want to know whether they can access the domain (send mail to it etc.), that is to know whether it is present in the DNS. And for that purpose, the DNS protocol is the only right tool.