[boomlinde]

1) You can have the kernel notify you upon a file system change via a watch descriptor. It's not rocket science, and you don't have to do any polling.

2) It might not buy you much (I disagree), but is hardly a disadvantage over any other approach where you'd want to send compressed http responses.

3) Let me get this straight; you're talking about the CPU data cache? It's not going to be the bottleneck in a scenario where you are basically pumping data into a socket. If it was, how would allocating a lot small buffers have an advantage over a single large allocation, if the data was the same in the end?

As for elephants...

1) It depends on the expected result. A page built on static files doesn't need to be static in any other sense.

2) I'm not sure why you think that dynamically generated "server pages" are different from static ones in this case. The GWT pagespeed optimizations you are talking about are not relevant to the discussion, since they are not inherent to either static/dynamic backends.

3) As far as I'm concerned, this is the only valid point you have brought to the discussion.

[waps]

1) and the atime updates (I am saying this because apache does make sure it updates them, which runs them -at least- 1ms per file they serve, and yes they accumulate these accesses, but they still cause limits and slowness) ? Caching the file will, incidentally, make watch descriptors not work anymore for other apps.

2) so it's identical. Meaning it's not an advantage of static files.

3) Actually the cpu data cache is going to be the bottleneck when you have 2x10G going into a single server (because you're using 92%+ of your data bandwidth just pumping data into the network cards). Not writing (or reading) to/from memory when you don't absolutely have to is what's going to make the difference in speed. If you get to the point where you can expand templates and compress them without hitting main memory (not that hard in chips with 16M third level cache), it will make a huge difference.

And of course this is assuming it actually fits in physical memory. If templates fit, but expanded they don't, templates win hands down (so expanding per-user stats pages into static files, for example, is guaranteed to cost you more than it buys you).

Elephants.

1) true, but it's going to cost you in roundtrips

2) yes they are. The optimization states that you optimize by having an "initialize" call, to get all data you need to render the full user's page. Then the dynamic backend calls that function, stores the result at the bottom of the GWT serving page, and boom, you go from n roundtrips to 1. Since a roundtrip is going to cost you at least 80-90 ms in any case except a local LAN, this is one of those optimizations that if you don't have them, you can forget about having your page displayed in less than 400-500ms.

Dynamic serving gets a bad rap because people are doing 10 mysql queries to a database server (meaning a different physical machine) with explicitly disabled caching and then complaining that actually takes time ... it's ridiculous. Looking at some PHP code and the speed complaints on fora that accompany it, it feels like putting your grandmother's dead grandmother behind the wheel of a ferrari and complaining it doesn't move fast.

Having small programs generate data on cue is always going to be faster than reading the actual data. I don't get why this isn't obvious to everyone. Hell, it's how (and why) compression works.

[boomlinde]

1) You don't need to poll atime descriptors. Neither do you need to use watch descriptors (although I don't understand your argument about them). You just need a way to tell the server that you updated a page. It might have been a good point if we were discussing specific implementations.

2) Are you saying that it was a completely irrelevant thing to mention in your list of problems with generating static files to begin with? If so, I agree.

3) Whether you're right or wrong (I can't tell at this point, and I won't simply decide on either until there are numbers to prove it), you are arguing on the basis of very limited and somewhat unlikely scenarios. I could argue on the basis of having very complex templates and page logic to generate extremely small pages, but I don't since I realize that it's not very likely to happen.

Elephants again...

1) No. If your use case is such that your static page contains the same information what your dynamically generated equivalent would (and no one is arguing for using static pages if this isn't possible), you don't need additional round trips.

2) If all data you need to render to a page is static, you could easily go from one roundtrip to none.

Let's bring another animal into the room:

With static files, you can enjoy the luxuries of very expressive languages with slow runtimes to build your pages, without any performance impact on serving them. Generating dynamic pages in a way that will even come close to serving static files in terms of performance is going to mean spending more effort than is viable for most projects.