Re: [9fans] 9ttp

[John Floren <john@jfloren.net>]

On Thu, Sep 8, 2011 at 12:44 PM, Bakul Shah <bakul@bitblocks.com> wrote:
> On Thu, 08 Sep 2011 09:56:10 PDT John Floren <john@jfloren.net>  wrote:
>> On Thu, Sep 8, 2011 at 9:31 AM, erik quanstrom <quanstro@quanstro.net> wrot=
>> e:
>> >> > On Thursday 08 of September 2011 14:54:40 erik quanstrom wrote:
>> >> > > On Thu Sep =A08 04:52:08 EDT 2011, 23hiro@googlemail.com wrote:
>> >> > > > HTTP is technically different and not easily comparable to 9p. HTT=
>> P is
>> >> > > > not a good example of how to do things, but over high-latency link=
>> s 9p
>> >> > >
>> >> > > =A0 =A0 =A0 with a single outstanding request
>> >> > >
>> >> > > > is much slower for getting files.
>> >> > >
>> >> > > there, fixed that for ya.
>> >> >
>> >> > is 9p windowable at all? is that implemented?
>> >> >
>> >> >
>> >> >
>> >> 9p has tagged requests.
>> >
>> > cf. /sys/src/cmd/fcp.c
>> >
>> > - erik
>> >
>> >
>>
>> I do not think it is acceptable to have to fork repeatedly merely to
>> efficiently read a file. Also, as far as I can tell, exactly one
>> program (fcp) does that.
>>
>> Can a single process have multiple outstanding requests? My
>> investigations indicated not, but then again I may have mis-read
>> things.
>
> Is there a way to distinguish between files backed by real
> storage & synthetic files? Seems to me that the server
> wouldn't know if you pipelined multiple read/write requests on
> a given connection (in-order delivery). May be the client can
> do read-ahead of N blocks. But one issue with read-ahead /
> write-behind is the problem of head of line blocking --
> further non-r/w requests queue up behind them. That is why FTP
> uses a control connection for all the commands & responses but
> data is delivered on a fresh tcp connection.
>
>

See my thesis for an FTP-like extension to 9P
(https://bitbucket.org/floren/tstream/src/67c7419ad84a/documents/Thesis.pdf)
in which 9P messages are used to negotiate a separate TCP data stream,
avoiding the blocking problem. It achieved transfer performance
equivalent to that of HTTP over a high-latency link.

Deja vu here--I know we just discussed this about a month ago :)


John