Re: [bug-myserver] task #9307: Propose to improve GNU MyServer Throughput and Reliability

[Alexandru Iancu]

On Tue, Apr 14, 2009 at 1:44 PM, Giuseppe Scrivano <address@hidden> wrote:

Alexandru Iancu <address@hidden> writes:

> On the other hand, as I said in that paper, up until now, optimizations were made 
> in application software like this sendfile OS feature while using this new transport
> we are trying to improve the performance other places: at transport layer.

Yes but SCTP can take advantage from the sendfile optimization too in
the same way TCP does.

 

I hope it may use this because it's really nice feature, unfortunately I tried it yesterday

evening with default parameters and it didn't worked. There are about 20 SCTP options

so maybe it'll work with little bit of tuning ... neither did I looked into kernel sources

to have a hint about what option to use to make it work. Actually I never did this ...

 

Without sendfile you have (assuming the file is not cached):

file system -> kernel space -> user space -> network.

with sendfile it is:

file system -> kernel space -> network.


> On the other hand this feature is marginal to our server because it&#39;s not available on

> windows and used just by HTTP(S).

I think FTP could use it too, are there any reason to don't use it?

 

Unlike HTTP, FTP transfers also big files so the spec. requires the server to cancel ongoing

download/upload. I don't know how to stop a sendfile API? May I register a callback that OS

checks from time to time to stop a sending file?

 

> Did you measure into an real world environment how myserver&#39;s performs using

> sendfile vs. not using this feature? Or do you know studies that show how much
> gain ones can have using sendfile instead read/send? Because I think in Internet

> the real server&#39;s botleneck is not the processor&#39;s power but the number of connections

> the server(as a whole: user and kernel spaces) can handle concurrently. 
> SCTP helps me to handle more connections concurrently.

I don't have numbers too but when I started to use sendfile for HTTP I
saw big differences.

> So, I&#39;m optimistic about using SCTP even without sendfile.

Thought it can be better to have both :)

Have you already wrote any test case for SCTP that I can use as starting point?

 

I agree, I also would like to have them both. Again, maybe it's an socket option problem or a stupid bug in my code :)

I tryed to focus on SCTP with 2 applications: a "server" listening on a one-to-many socket and a multiple threads client connecting on a one-to-one socket. Before you need to install a kernel loadable module(lksctp). I'll send you the sources later today.

 

I didn't yet have enough time to propose how to change myserver's code. I'll focus on myserver starting with this next weekend.

As a time frame, about the end of next week(April, 25th) I hope I'll have an answer to this question.

 

Andu.

 

Thanks,
Giuseppe