On Wednesday, 6 July 2016 at 14:57:08 UTC, chmike wrote:
On Wednesday, 6 July 2016 at 11:33:53 UTC, Eugene Wissner wrote:
The only reason libev was choosen is that it is the simplest
implementation I know about. A few C files. I had an
educational purpose: I wanted to see how an event loop works
on low level. Asyncio was for me no-go, since I've not written
a lot in C++, and can only read the code a bit. So I'm not
hanging on libev. The only another implementation would be the
python event library. It is also pure C code and it was much
more cleaner written than libev on the first sight.
Your project and work is valuable on many aspects. I didn't
mean to depreciate it.
Now there are two problems with my work:
1) The first is something we all are tired to talk about:
manual memory management. I make a proof of concept and am
writing the code that is 100% marked as @nogc. It has side
effects. For example I allocate exceptions and thay should be
freed after catching - it is something, phobos doesn't do. As
said it is an experiment; I would like to see how it works.
2) Performance. The main reason I started the writing at all
is that the existing implementations seem to have only
performance as criterium. Performance is super important, but
not on the cost of design, usability and extensibility. For
example in vibe.d (and libasync) everything possible is
defined as structs, everything that would be interesting to
extend is final; and after it you go to phobos and see a
"workaround". For example Mallocator is a struct but you have
to be sure, it is an allocator. How do you force the right
interface?
That is a valid point. I know it is hard to get the best
performance and optimal API design at the same time. The reason
the methods are final is to avoid the overhead of virtual
method indirection.
static if (hasMember!(Allocator, "deallocate"))
{
return impl.deallocate(b);
}
else
{
return false;
}
Somethinkg like this would be ok for C. But for a language
with interfaces, it is ugly design independent of how it
performs. Everything I say here is IMHO.
This would mean we need a new design pattern that supports both.
Except these two points I'm interested in some kind of
collective work aswell. It is very difficult as one man job. I
didn't know other people are also working on similar
implementations. Nice to know.
Are you aware of any benchmark tools in other languages that
could be used?
The benchmark tools available are mainly testing web servers.
And the exact operation pattern is not very clear. One of such
benchmark tool is wrk [https://github.com/wg/wrk] that measure
HTTP request speed. The problem is that it measure the
performance of the I/O and HTTP handling.
Here are some benchmark results:
* https://www.techempower.com/benchmarks/
* https://github.com/nanoant/WebFrameworkBenchmark
How can D be worse than Java ?
My strategy would be to split the problem. First get the async
I/O optimal. Then get HTTP handling optimal, and finally get
database interaction optimal (the optimal async I/O should
help). An investigation on the methods used by Java/undertow to
get these performances could help.
I would suggest to implement a benchmark client doing some
predefined patterns of I/O operations similar to web
interactions or the most common types of interactions. And a
server implemented in C using native system I/O operations.
This server implementation would then be our reference.
What would be measured is how much slower our different D
implementations are relative to the C reference implementation.
This will allow us to have a benchmarking test that doesn't
depend that much of the hardware.
Yes. it would be a way to go.
As I see Kore performs pretty well. One could write bindings
first and then begin to rewrite step for step. On every
development step you have a working http server and you can use
these http-benchmark tools for the benchmarking.
I have to look in it in the next weeks
