[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [MEXT] about the nature of "flows" Re: weird idea about flow bindings



George Tsirtsis escribió:
inline...

On Mon, May 18, 2009 at 12:38 PM, marcelo bagnulo braun
<marcelo at it.uc3m.es> wrote:
George Tsirtsis escribió:
Marcelo,

It might be useful to step back a bit to understand better what the issue
is.

I think one issue in this discussion is that the use of the term
"flow" in flow bindings in the flow bindings and flow description
drafts, is currently a bit wider than that used in flow labels. Flow
bindings, with any of the filter description formats discussed (text
or binary) should be more accurately described as "packet filter rule
binding".

In other words, all our filter descriptions so far, have taken the
view that the MN should have as much flexibility in identifying
"flows" (or more accurately "filter rules") as a regular packet
classifier e.g., firewall. You seem to think this is a bad idea and
that only IP layer (and possibly transport layer?) fields should be
used in the context of flow bindings.

Am I right so far?

correct

GT> Just to make sure we are on the same page, do you agree that flow
label is not sufficient,

flow label is clearly not sufficient since at least the src and dst address are also needed
but i am ok with using any IP layer information
I am not so comfortable with transport information, but as is stated earlier, i guess i can live with that
 and that other IP and transport layer fields
need to be used too? i.e., are you only against adding the Pointers to
binary filter rules? or you want to limit the fields used further?


just the pointers
I think that in the extreme, one could go further and say that packet
routing should really be ONLY based on IP Addresses.
not really, the flow label is explcitly there to provide special treatment
to flows. Same with the traffci class bits. It is perfectly ok to use those
a part of the routing decisions imho

GT> classification is different from routing. Do you know many routers
(beyond lab experiments on QoS routing) that use Flow Labels or even
DSCPs for routing? In any case I was making an "extreme"  argument. Of
course I think it is ok to use flow labels for routing .... but I
think it needs to be combined with other fields too.

 Anything else is
mostly broken since it can never scale and it is not compatible with
regular IP routing. I am actually sympathetic to this view of the
world in which it requires that the MN uses a different source address
for every "flow" it wants to be able to route separately. Then flow
bindings are simply implemented by redirecting the various MN
addresses used to the different CoAs of the MN.

The flow label, however, is IMO mostly an inadequate field to use for
flow bindings. This is because a given MN has no control over when and
how a given CN uses the flow label.Flow Bindings allow the MN to
 affect the 'last hop routing' of packets going from the HA to itself.
It makes no sense to me for the MN to have to rely on the use of what
is basically an optional field by random 3rd party nodes.


I think we are focusing in different ways of splitting the problem.
IMHO, the IP layer does not know what an application flow is. It doens't
knwo how the application set different bits and what it is more it doesn't
have to know.
The flow label is a handle for the IP layer to know what a flow is. Making
the IP layer second guess what a flow is, is not the best idea.
So, while you see the division MN/CN, i am concerned in the divsion IP
layer/app layer


GT> I agree. All this, however, is supposed to become known to the MIP
layer via an API...no one expects the MIP engine itself to figure
these things out.

In contrast, the flexibility provided by the binary flow description
as you know is mostly about IP layer and Transport Layer fields, which
I think are the minimum required to provide effecting flow binding
capabilities. The recent addition of the Pointer idea simply expanded
the capabilities to random upper layer field matching, I believe very
elegantly and powerfully, if you do not mind me saying so ;-)

well, i do mind.


GT> OK, shameless self-promotion did not pan-out :-)

as defined in the draft

 Offset

    The Offset field identifies an integer number of bytes from the
    beginning of the IP header.  It points to the beginning of the
    field of interest in the packet.

So, suppose that the idea is to use application layer bits.
Now what happens if the IP layer adds an extension header or an option
inside a header?

GT> It is a matter of identifying the start of the offset at an
appropriate place e.g., at the end of IP header and all options.

What happens if the TCP inserts one option?

GT> OK, maybe the offset should start at the end of the transport
layer header :-)
Joking aside, this is trivial;
i don't know if i would call it trivial.
I tcan certainly be done, (and it is done in DPI) but that doesn't make it easy or a good idea. Making the IP layer to understand the trasnport options and app data bit patterns still doesn't sound like a good idea to me

 I remind you that these offset-based
pointers have been used in packet classifiers in the field for many
years....not really my idea. The only case it does not work is when
e2e security is used (e.g. IPSEC), in which case of course only IP
layer fields can be used for classification.

What happens with applications that we don't know the format (such as
skype)?

GT> But that's the whole point of the Pointers. You do not have to
know! It is the Skype app that would have to use the API to set-up an
appropriate Pointer.

mmm, i don't follow this.
suppose i own a PDA and i want that the skype traffic is routed through my GPRS link. How do i configure my PDA to do so? Skype won't do it for me. So we have two options. Wait for skype to support the flow binding API or wait for te OS to properly set a random flow label. I really think the second is more reasonable.

Still, I think it is valid to question whether such higher layer
matching capability is a good thing architecturally. As of yet,
however, I am not sure I see what the issue is with providing such
capability, beyond a somewhat academic issue of regular IP routing
compatibility.


well, here i guess we disagree. including application data bit pattern in
mip signalling doesn't seem the best architecture to me. Layer violation
seems at least one issue here


GT> As I said, I think you are making a valid argument for discussion.
I would in fact be willing to consider dropping the Pointers from the
binary flow description draft as long as "flexibility" is not used as
an argument when comparing it to text based flow description; remember
this was the main reason the Pointers where added to the draft in the
first place.

mmm, i have even more issues with any other flow language descriptors that enter even more into the app land.

Regards, marcelo


George

Regards, marcelo


As a side note, I have no problem whatsoever requiring that HAs
perform deep packet inspection as the binary flow description draft
requires. Commercial HAs are able to perform packet filter inspection,
and are actually expected to do so, for other needs already
(firewalling, lawful interception, value-add bells and whistles etc)
so I do not see any issue there. Note that the use of Pointers for
such deep header inspection is trivial compared to other methods since
it does not require that the HA understands the upper layer header
formats. It simply moves to byte X and matches Y number of bytes. On
the MN side things are different and as I explained the complexity on
the MN side is implementation specific.

George


On Sun, May 17, 2009 at 11:34 AM, marcelo bagnulo braun
<marcelo at it.uc3m.es> wrote:

George Tsirtsis escribió:

Marcelo,

Flow Movement is supposed to work in combination with DSMIPv6. Are you
suggesting the flow bindings should only work for IPv6 traffic then?


no, i ma suggesting that for IPv6 traffic, we should use the flow label
for ipv4, we certainly need to use something else
What i am not buying is that we cannot build a method that relies in the
flow label cause IPv4 doesn't support it. In other words, i think it is
perfeclty ok to build our reccomended method relying on the usage of the
flow label, and define some other hacks to support ipv4.

why would we place such a restriction?

BTW, as far as I know flow label is the least well defined aspect of
IPv6.

but that is not the porblem of IPv6, but it is an inherent problem of
defining a flow. The IP layer has a problem cause it is agonstic to
flows.
So, only the application knows what really a flow is. And IPv6 does
provide
the means for the application to signal the packets that belong to flow
to
the IP layer i.e. the flow label.
In other words, this is a fundamental problem: IP does not knows about
flows
and we are trying that MIP which is in the IP layer to discriminate
flows.
We can either try the IP layer to guess what are the common bits of a
flow
or we can leverage on the explicit definition of a flow, which is what
the
flow label is for.

 How does the MN know whether a given CN marks each relevant flow
with a different flow label?

becuase it needs to be compliant with RFC3697?

Also note that flow labels are source
address specific so the flow label by itself is not sufficient.



sure, the flow description is src add, dst add and flow label


Also the binary format
(http://tools.ietf.org/html/draft-tsirtsis-mext-binary-filters-00) in
fact allows for a very simple MN implementation if the only thing you
want to classify is flow labels. An implementation could only allow
the flow label to be set resulting in the following flow descriptor:

    0                   1                   2                   3
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |0|0|0|0|0|0|0|0|0|0|0|0|M| 0000 |       (M)   Flow Label
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
       FB (cont)     |
   +-+-+-+-+-+-+-+

Or actually something like the above but including the source address
too (flag C), do me more correct..

If an MN implementation does not want to deal with any of the other
fields, it does not need to implement them.
In that sense multiple compatible implementations are allowed as long
as the HA implements the flow descriptors in their full complexity.



this is my problem. I don't want to impose the HAs to be able to parse
arbitrary bits in the application data. I think this is not the right way
to
do it. I don't think the IP layer should be snooping into application
data
bits


Finally note that what constitutes a "flow" could be something like
"ALL TCP Traffic", no matter which source address and what flow label
is used. How would the MN be able to say push all TCP to one CoA and
all UDP to another based on your scheme?


well, you can do this based on the protocol field of the IP header
As i stated in my earlier reply to Hesham, i don't have a problem with
using
any IP header fields as flow descriptors and i may be ok with sing the
transport header bits (not what i would preffer, but hey, you can't
always
get what you want)
My problem is with supporting application data bits as part of the flow
description, which i really think is broken

Makes sense?

Regards, marcelo



George


On Sun, May 17, 2009 at 10:03 AM, marcelo bagnulo braun
<marcelo at it.uc3m.es> wrote:


George Tsirtsis escribió:


Hi Marcelo, I now understand what you propose better. and it makes
more
sense than I initially thought.

Two issues, though:

1) this requires the CN to cooperate i.e., make proper use of flow
labels.


according to RFC3697:

 To enable Flow Label based classification, source nodes SHOULD assign
 each unrelated transport connection and application data stream to a
 new flow.
...

 To enable applications and transport protocols to define what packets
 constitute a flow, the source node MUST provide means for the
 applications and transport protocols to specify the Flow Label values
 to be used with their flows.  The use of the means to specify Flow
 Label values is subject to appropriate privileges (see section 5.1).
 The source node SHOULD be able to select unused Flow Label values for
 flows not requesting a specific value to be used.

So, yes but node should do this if they expect things to work properly



It is not clear to me that even if all nodes populated the flow label
(which I think is not even true)


does anybody has information about the support of RFC3697 in current
OSes?



, its use would be universally consistent in a way that would ensure
that the setting of the flow label by the CN would be inline with what
the
MN would like to identify as a flow.


that would determine the granularity of the method
I am not sure the spectrum of choices is so big that this can be a
problem
could you think of some very different criteria to define what a flow
is
that this would be an issue?


2) This does not work for IPv4 traffic.



i don't buy this argument. If we apply this argument, then we could
never
use the flow label for anything, since it doesn't work for IPv4.

Reagrds, marcelo



Thoughts?
George
On Sun, May 17, 2009 at 8:58 AM, marcelo bagnulo braun
<marcelo at it.uc3m.es <mailto:marcelo at it.uc3m.es>> wrote:

 ok, let me try again because it seems i am failing miserably to
 communicate what i want to say (or it simply so broken that people
 cannot believe what they are reading :-)

 Let's try with an example.

 We have the MN who has two CoAs, CoA1 and CoA2.

 The MN has a policy that says that it wants flows from app1 to be
 exchanged through CoA1 (well, the policy would say that they
 should be received by the interfece in which CoA1 is configured)
 and that flows from app2 to be exchanged through CoA2 (same
 comment about the interface applies here)

 In addition, MN has defined CoA1 to be the primary CoA i.e. the
 CoA with highest priority.

 So, we have two different situations. Either the MN starts the
 communication or the CN starts the communication. Actually, this
 is pretty irrelevant, cause the two directions of a connections
 are handled independently. the MN will send packets from app1
 through CoA1 and packets from app2 through CoA2 and it doesn't
 need any additional actions from the HA for outgoing packets.

 Now, the difficulty is with incoming packets. the MN needs to
 inform the HA about which CoA it wants to use to receive any flow.

 Suppose that the HA has no flow bindings for the MN so far.

 So, suppose that the app1 in CN starts a flow to the MN. The HA
 will receive the packet and will send it though the CoA1 cause it
 is the one with higher priority. the MN is happy and there is no
 need to do anything else.

 Suppose now that the app2 in the CN starts a flow to the MN.
 According to RFC3697, the CN SHOULD assign a different flow label
 to the packets belonging to this flow e.g. to this TCP connection.
 Let's assume it is actually a TCP connection to make the example
 more concrete.
 So, app2 opens a TCP connection to MN. CN then sends the TCP SYN
 to the MN.
 The HA receives the TCP SYN and since there is no flow binding
 state for this flow it forwards it through the CoA1 (the one with
 highest priority)
 The MN receives the TCP SYN through CoA1 and it realizes that this
 is not compliant with its flow policy. So, it extracts the flow
 label value, the src address and the dest address and it sends a
 BU to the HA stating that the flow identified by the flow label,
 src add and dest add needs to be bound to the CoA2.
 Simultaneously, it also sends the TCP SYN+ACK back to the CN.
 The CN receives the TCP SYN+ACK and sends the Ack back. the HA
 receives the ACk and will forward it based on the flow binding,
 i.e. through the CoA.

 the result is a method that never looks into higher layer
 information for setting the flows.

 makes sense?

 Regards, marcelo


 Hesham Soliman escribió:


     On 17/05/09 3:04 PM, "marcelo bagnulo braun"
     <marcelo at it.uc3m.es <mailto:marcelo at it.uc3m.es>> wrote:


         Hesham Soliman escribió:

             The flow label should be part of any filter descriptor
             anyway. So at anytime
             A MN can send a BU specifying that packets with flow
             label X should be
             directed to CoA1. I'm not sure how your proposal
             changes/adds to this.

         what i am saying is that you don't need anything else
         other than this


     => see below


             As a general issue, the flow label is unlikely to be
             used for this purpose
             until we resolve it's mutability issues

         what mutability issues are you referring to?
         AFAIU, the flow label is set randomly b the source and
         remains unchanged
         e2e and for the duration of the flow.


     => I thought RFC 3697 talked about it being changed en route
     then changed
     back before it hits the destination. But I just scanned it and
     I can't find
     this text. Maybe it was never included.


              and until it is actually _used_ in
             the API. At least until recently no one was using it.

         why do we need to have it in the API? It is only imporntat
         the the
         source sets it to a random value and keep it unchenged e2e
         and for the
         duration of the flow. I understand that these conditions
         are met today,
         right?


     => But are you saying the src MIP6 module overwrites the
     potential flow
     label set in the API? Because it is possible that an app sets
     the flow
     label. RFC 3697 assumes the presence of an e2e mechanism to
     negotiate the
     flow label. If not then how do you make sure it's always
     unique? And how do
     you make sure that the CN echoes back the same value in the
     flow label?
     And what if the CN initiated the connection?

     Hesham






 _______________________________________________
 MEXT mailing list
 MEXT at ietf.org <mailto:MEXT at ietf.org>
 https://www.ietf.org/mailman/listinfo/mext