Re: [Bug-wget] [RFC] Extend concurrency support

[Daniel Stenberg]

On Wed, 21 May 2014, Tim Ruehsen wrote:

> > libcurl offers a substantial amount of more functionality in the network 
> > layer than what wget has. And yes, libcurl has a DNS cache.
>
> As I understood Guiseppe, he wants to concentrate on FTP(S) and HTTP(S). 
> Additional functionality like POP3, IMAP, ... is not going to be used.

Let me rephrase: libcurl supports much more features in HTTP(S) and FTP(S) 
than wget does. Like more auths, more proxy auths, socks, FTPS, http2, more 
TLS library support, happy eyeballs and more. All that even event-based if 
wanted.

> You are using chained lists to store cache entries !?

I'm not sure what cache you're talking about. libcurl has several. But also we 
don't expose those things in the API so if we'd run into a bottle neck it 
isn't that big of a deal to improve it. I figure you talk about the DNS cache 
here and that's only storing names for a short period (since it doesn't have 
the real underlying TTL) so it'll typically never grow very large.

> Does this scale with Wget corner cases like 'download the internet' ? I know 
> it sounds it bit nit- picky... but I like to mention it rather sooner than 
> too late.

libcurl is used to download thousands of parallel transfers in existing 
applications today, in way more "heavy" use than wget ever does and it works 
perfectly fine and with speed. But of course, there are always room for 
improvements and we accept patches.

> I am not sure, how we find enough people-power for this task. On the other 
> hand side, that's what I've done in the Mget project. I guess, a merge of 
> Mget and Wget would be less work. Mget already implements most of Wget's 
> options plus a bunch more.

If mget > wget, why bother to merge anything to wget? Why not just either keep 
working on mget as it is or just rename mget to wget2?

> > > - Cookie logic (incl. public suffix handling)
> >
> > libcurl also provides cookie support.
>
> Yes, to set cookie headers in requests and get cookies from responses.
> This would not replace relevant code in Wget's cookie.c

libcurl offers more cookie API than that, but more importantly I did 
explicitly say that libcurl's functionality in FTP/HTTP/cookies etc would NOT 
completely remove the need to add necessary adaptions. Perhaps cookie code 
would be example of an area that would need more adaptions. Or perhaps not.

> the checksum has to calculated and verified. With a non-threaded approach, 
> you would serialize this task to a single CPU core. While checksumming, the 
> parallel download is paused. Not so in a threaded model.

If you ask me, the overhead and complexity of threading does not motivate the 
small performance win this gives you. But that's just me and I'm not forcing 
this opinon on anyone.

> The same goes for DNS resolving as long as the resolving does not work 
> asynchronous.

libcurl offers asynchronous integrated name resolving already.

> Not sure how and if this works with libcurl, but I guess that it will make 
> the clients code more complex.

Not at all. The client doesn't have to know nor care.

> One question that came to my mind while I was looking at libcurl API. What 
> about type safety (thinking of e.g. curl_easy_setopt()) ?

It only has limited type checks (especially if gcc isn't used) but it is 
typcially not a big problem to users and the API has been tested and remained 
stable for quite some time by a fair amount of users. I'm sure it wouldn't be 
a problem for you either.

--

 / daniel.haxx.se