Hi Andy,
as I said, I did not touch it, even because at the moment there is no
support/interface for the vectored reads (except for the patch sent by
Leandro, which I will sort out now).
Hence for now there are no compatibility problems I suppose...
Anyway I did not propose to change the length of the trailing 32 bits.
In fact, my proposal had the purpose of keeping the compatibility, not
breaking it...
I proposed to keep the dlen field exactly as it is (32 bits which
specify the length of the subsequent chunk, like any other request) and
move everything else (flags+path+prqlist) to the data part (which is not
part of the header) of the request. Please note that breaking that field
will require to define an exception to the general rule behind the protocol.
IMHO this is more coherent, but, again, I have no problems.
Right now the choice for the prqlist is not important to me. What I am
waiting for is the kxr_bind server side implementation. The low level
multistream code seems fine now at the client side, but I cannot go
further until the server implements kxr_bind. Please refer to my
proposal for that, and tell me if something has to be changed.
ah, I was forgetting:
every parallel stream (not the main one) initializes in this way:
- connection
- handshake
- kxr_bind to the session of the main login stream
note that XrdClient fires the parallel streams after the successful
kxr_open (hence, the multistream init is parallelized inside the
parallel open mechanism!!)
Ciao ciao
Fabrizio
Andrew Hanushevsky wrote:
> Hi Fabrizio,
>
> Actually, we have a couple of problems with redefining the "prqlen"
> field. The main one is forwarde/backward compatability. 1) Old servers
> must work with new clients and old clients must work with new servers.
> Therefore, the trailing field *must* stay as 32 bits. That means
> subdividing it. The best way is 8-8-16 as I indicated. Additionally, 2)
> Leandro has already worked up that particular change with his
> implementation of vectored reads. So, I strongly insist that the field
> be subdivided as 8-8-16.
>
> Andy
>
> ----- Original Message ----- From: "Fabrizio Furano"
> <[log in to unmask]>
> To: "Andrew Hanushevsky" <[log in to unmask]>
> Cc: "Xrootd Mailing List" <[log in to unmask]>; "Peter Elmer"
> <[log in to unmask]>
> Sent: Wednesday, June 28, 2006 1:54 AM
> Subject: Re: kxr_bind again
>
>
>> Hi Andy!
>>
>> I am very advanced in the first impl of the multistream client. I
>> want only to summarize here the version of the bind/read req I am
>> expecting:
>>
>> kxr_bind:
>>
>> Request Normal Response
>>
>> kXR_char streamid[2] kXR_char streamid[2]
>> kXR_unt16 kXR_bind kXR_unt16 0
>> kXR_char sessid[16] kXR_int32 2
>> kXR_int32 0 kXR_unt16 substreamid
>>
>>
>> kxr_read:
>>
>> honestly I don't realize if the structs you specified can generate
>> misalignments in certain architectures. Also, I don't see the point in
>> sparing 1 byte for such a relatively long request. I'd be more
>> generous for the flags/path/prglen fields, just to keep consistent
>> with all the other requests, which have a 32bit dlen field as last
>> field in the req header. I'd keep that also to maintain a little more
>> protocol backcompatibility. Anyway, here it is my version:
>>
>> Request Normal Response
>>
>> kXR_char streamid[2] kXR_char streamid[2]
>> kXR_unt16 kXR_read kXR_unt16 status
>> kXR_char fhandle[4] kXR_int32 dlen
>>
>> kXR_int64 roffset kXR_char data[dlen]
>> kXR_int32 rlen
>> kXR_int32 dlen
>>
>>
>> kXR_unt32 flags
>> kXR_unt32 path
>> struct prqlist[prqlen]
>>
>> struct prqlist {
>> kXR_char prhandle[4];
>> kXR_int32 prlen;
>> kXR_int64 proffset;
>> };
>>
>>
>> Where prqlen is the number of pre-read reqs. Hence the relationship:
>>
>> dlen = prqlen*sizeof(prqlist) + 2*sizeof(kXR_unt32)
>>
>> NOTE:
>> - path = 0 means that the respone has to be sent through the main
>> login stream
>> - the value of path also denotes the substream where to send the data
>> chunk requested in the req header
>> - the client does not care about the substream which will carry the
>> response to a read request. So the server is allowed not to honor the
>> path field for any reason. The idea behind is that a bunch of
>> substreams is considered internally as one mega-stream with
>> individually selectable substreams.
>> - I don't have any idea about the flags which we need. From my point
>> of view we can also consider to drop that field.
>>
>>
>>
>> NOTE2:
>>
>> Today/tomorrow I will commit my code, which seems to start working.
>> For now it does not deal with the kxr_read modifications. It includes
>> the mods/additions to XProtocol.hh, so, if you have any change for
>> that, please wait a bit until I commit, and then change that version,
>> just to avoid confusion... Here is some more info:
>>
>> - if the multistream support is disabled, then the old sock code is
>> used. The multistream support is implemented at low level in a
>> subclass of XrdClientSock.
>> - I did a test with multistream on but with only one stream and it
>> seems to work OK. More news to come.
>> - well, I cannot bind any test substream because there is no kxr_bind
>> yet... so anything can happen when you implement it and try to run the
>> advanced client... :-)
>>
>>
>> What do you think?
>> Bye!
>>
>> Fabrizio
>>
>>
>>
>>
>>
>> Andrew Hanushevsky wrote:
>>> Hi Fabrizio,
>>>
>>> A quick response as I am on vacation right now. I was planning for the
>>> server to return the substream ID and it will be an integer from 1 to n.
>>> As for directing responses to a particular stream. Recall that all
>>> requests go on the original login stream. The client directs which
>>> stream
>>> is to be used for the response. Recall that we were going to modify the
>>> read and write requests so that the stream ID could be inserted. This is
>>> also compatible with vectored reads that Renee wants. The change is:
>>>
>>>> kXR_int32 prqlen
>>>>
>>>> to
>>>>
>>>> kXR_char flags
>>>> kXR_char path
>>>> kXR_unt16 prqlen
>>>
>>> in the read request and similarly for the right request (there is a 4
>>> byte
>>> reserved field there). Note that "path" is the substream ID and it is a
>>> character field as the stream id will be 1 to 255 (actually much
>>> smaller).
>>> You may want to reflect that in the bind response.
>>>
>>> Notice that only read/write is multi-stream capable. All other requests
>>> are always done on the login stream. It simplifies things very much.
>>>
>>> Andy
>>>
>>> On Wed, 21 Jun 2006, Fabrizio Furano wrote:
>>>
>>>> Hi Andy,
>>>>
>>>> I am working on the multistream support, and I choosed the second way
>>>> I commented in the previous message. The server has to generate an
>>>> integer which will identify univoquely that substream in the pool which
>>>> belongs to a session. I recall to you that part of the prev message.
>>>>
>>>> The client does not expect a praticular way to generate that number.
>>>> Even random is good, so you can do what you prefer. The client also
>>>> will
>>>> not care about the (sub)stream it will get a response from. The only
>>>> important thing is that all the msgs belonging to a particular request
>>>> (e.g. kxr_oksofar+kxr_ok) all go through the same substream. No matter
>>>> which. Otherwise it's impossible to keep the ordering.
>>>>
>>>> So, if you want, a first server side version could also randomize the
>>>> choice of the response socket. this would be fine, since at the moment
>>>> the protocol has no way of requesting the server to send a response
>>>> through a particular substream.
>>>>
>>>> Right now I only wrote the low level support for the multistream
>>>> xfers, but I believe that also the client will randomize the choice of
>>>> the substream where to send requests to. We'll see.
>>>>
>>>> Fabrizio
>>>>
>>>>
>>>>
>>>> -------- Original Message --------
>>>> Subject: kxr_bind
>>>> Date: Thu, 15 Jun 2006 11:25:52 +0200
>>>> From: Fabrizio Furano <[log in to unmask]>
>>>> To: Andrew Hanushevsky <[log in to unmask]>
>>>>
>>>> Hi Andy,
>>>>
>>>> following our recent discussions on the multistream xfer, I propose
>>>> this change in the kxr_bind spec:
>>>>
>>>> Request Normal Response
>>>>
>>>> kXR_char streamid[2] kXR_char streamid[2]
>>>> kXR_unt16 kXR_bind kXR_unt16 substreamid
>>>> kXR_char sessid[16] kXR_int32 0
>>>> kXR_int32 0
>>>>
>>>>
>>
