Hi Andy,
From Michel and you I understand that the problems with rehashing come
from:
- Different manufacturers may provide servers with the same module set,
but may have chosen different names to rehash in case of clash.
- Servers may be upgraded with modules, and may need rehashing.
Not choosing module identifiers in the name hash, has as consequence
that in case of clash in a server, the server needs to announce the name
that is rehashed.
This announcement can be done in a standardized format and pre-loaded
servers with clashes MUST be accompanied by such a file.
Given that small clients need to be prepared when servers are added or
modules are added to existing servers, Appropriate code can be loaded in
these clients when needed.
The existence of rehashes does not unnecessarily complicate this work.
The choice is then between:
- no module IDs: rehashes need to be administrated and taken into
account when servers are added or functionality is added.
- module ID: Registry needs to be defined, maintained, and the module ID
must be transported in every exchange.
The module ID is compulsory when the client does an automatic discovery
of servers and needs to use the same hash values independent of the
arrival of new servers or the extension of servers.
Is this a correct summary of the arguments involved?
Peter
Andy Bierman schreef op 2015-04-13 18:14:
On Mon, Apr 13, 2015 at 1:34 AM, peter van der Stok
<[email protected]> wrote:
Dear all,
Trying to work out all your recommendations, I fail to understand the
need
for the solution.
Large clients and small servers is not tru any more.
The assumption is that clients and servers are small, they are not
dynamically augmented with new modules. They only know about hash
values and
have no knowledge what so ever of the corresponding names.
The term "dynamically loaded" is misleading, because a new
revision of the device, or a device that supports firmware upgrade,
can have new YANG modules added.
Let's concentrate on the one module clash first.
Apparently, it is envisaged to load modules that already contain hash
clashes inside.
For me that is an unsolvable problem: There is one hash value that
points to
different names (memory stores) in the server.
Servers only know about hash values, they have no name tables. Suppose
one
of the names has been rehashed (where? certainly not within the
server).
We might add a statement in the server that the old hash value goes to
the
new rehashed value.
It is not clear to me how the server can decide which memory location
goes
with the old hash value and which with the new value.
If a value gets rehashed, it is because there is a collision.
I am not sure if module numbers can be added to YANG,
but it should be possible to either avoid 1-module collisions
or add a YANG extension to the module to resolve the collision
so that every implementation uses the same rehash (known in advance).
Apparently, something needs to be done before the module is loaded
into the
server. For me that means that only modules can be loaded without hash
clashes.
For modules intended for CoMI, we can certainly
make sure no names used in the module produce any collisions.
Suppose in a given server two names from two modules clash. The
solution is
then that the module ID distinguishes between the two clashing values.
It is then proposed that module names are hashed. That solution is the
same
as saying that you need more than 32 bits for a hash value to reduce
clash
probability.
The only valid remaining approach is a registry that maps module name
to a
32 bit? ID. Transporting the module ID means additional transport
costs.
It should be possible to come up with a short module-id format,
and only use the long form (full module name) if no short-form exists.
I can imagine though that in a given installation the servers and
their
module combinations are known.
Any clashes can be solved before load time. The appropriate code can
be
loaded in the small clients and servers.
This does not allow for the old client/new server scenario to keep
working,
if the new server adds a module with a hash collision.
The problem arrives when a new server with a set of modules (including
a new
module) finds a clash.
This clash does not affect the old clients and servers (they are not
aware
of the new module)
But what if the module they know about is the one that gets rehashed?
There is no saved or canonical order for processing YANG modules.
Consequently, the clash can be solved by renaming the clashing name in
the
newly added module.
Only if the collision is detected and corrected before
the new module is published.
All clients, and servers which have to use the clashing name of the
new
module will use the hash value of the renamed name, and do not need to
be
aware of the rehashing.
At the operational level, there are only unique hashes within an
installation.
At the organization level, the rehash value of the new module needs to
be
registered and used when code for new clients or servers are
generated.
Consequently, I do not see a need for module IDs.
YANG is modular, and different naming authorities (SDOs and vendors)
work at different speeds and publish independently of each other.
Applications can use different modules on the same server without
impacting each other. Vendors can add their own modules before
or after standard modules are written and added.
If module A and module B both have objects that hash to the same value
then they cannot be used together without re-hashing 1 of the objects.
Any mistakes in the above?
I think the combinations of modules that will be available
on a given implementation cannot be controlled in advance.
We cannot assume all client devices and server devices can
be upgraded together (flag day upgrade).
Peter
Andy
Michel Veillette schreef op 2015-04-02 20:37:
If I summarize:
Reducing the scope of uniqueness of YAND hash values to each module
allows detection of hash collision(s) at design time instead of at
run
time.
To reduce this scope, we need to:
- Create a unique module identifier (e.g. 20 bits organization ID
registered at IANA, 10 bits module ID)
- Add this module identifier to the URI (e.g. one to five base64
characters)
- Add a map at the root of the CBOR objects carry within the CoAP
payload to associate module IDs with modules data nodes (e.g. one to
five byes per module)
The problem that still need to be resolved is how we document data
nodes rehash offline.
- One option is to use the description statement of the YANG module
or
a new YANG statement specifically added for this purpose.
- Another option is create a IANA registry for these rehash values.
Hash collisions within YANG modules are very unlikely to happen and
the solution don't need to be highly scalable. In that respect, using
a IANA registry might be a good solution since it can be used for
already published yang modules and can be use after the fact with any
future YANG modules.
Michel Veillette
System Architecture Director
Trilliant Inc.
Tel: 450-375-0556 ext. 237
[email protected]
www.trilliantinc.com
-----Original Message-----
From: Pascal Thubert (pthubert) [mailto:[email protected]]
Sent: 2 avril 2015 13:09
To: Carsten Bormann; Andy Bierman
Cc: Michel Veillette; [email protected]; [email protected];
[email protected]
Subject: RE: [6tisch] [core] COMI hash values globally unique vs.
unique within a module
Yes, we asked about that as well.
It help to have a registry that guarantees the uniqueness of the
first
element in the path, so as to make it easier to check for collision
within one path only.
Cheers,
Pascal
-----Original Message-----
From: 6tisch [mailto:[email protected]] On Behalf Of Carsten
Bormann
Sent: jeudi 2 avril 2015 18:41
To: Andy Bierman
Cc: Michel Veillette; [email protected]; [email protected];
[email protected]
Subject: Re: [6tisch] [core] COMI hash values globally unique vs.
unique within a module
I think I'm losing track of what "this" is.
Assigning numbers when doing a (version of a) module should be
doable,
even if it requires some additional tooling.
Assigning numbers to modules would require a registry.
No rehashing (or hashing at all) required.
(Now, that registry could be filled using a hash...)
BTW, I'm not sure that the "rehashing" in the current spec is the
easiest way to handle collisions -- I also don't understand how
multiple paths that arrive at the same hash value are handled.
Grüße, Carsten
Andy Bierman wrote:
> On Thu, Apr 2, 2015 at 8:25 AM, Andy Bierman <[email protected]>
wrote:
>> Hi,
>>
>> I should make it clear that I was trying to use a module-id so the
>> objects in the module could be given a simple integer OID.
>> The goal was to support identifiers that could fit in 1 or 2 bytes.
>> But this was dropped when we changed to using hashes.
>>
>> It might be a huge effort to create and use numeric module IDs, but
>> YANG module names tend to be long strings. It might be worth it, if
>> a per-module hash is used.
>>
>> If the collisions were coupled to specific modules, then they might
>> be avoided before publication or the client can be hard-wired to
>> use the rehashed values from the start.
>>
>> A client would not need to know the XPath strings for any objects.
>> It could be hard-coded to work with module X, Y, Z, and it would
>> always work, no matter what additional modules were added to the
>> server.
>> This does seem very useful for constrained clients.
>>
>
> A standard, deterministic renaming algorithm would be needed to make
> this viable. If each server is free to choose its own rehash, then
> the client still needs to store all the strings, and resolve the
> collisions at run-time instead of compile-time.
> (So therefore a module-name or module-id would not solve anything).
>
>> Andy
>
> Andy
>
>>
>> On Thu, Apr 2, 2015 at 5:55 AM, Carsten Bormann <[email protected]> wrote:
>>> Michel Veillette wrote:
>>>> strings represents */5323 bytes/*
>>> If we want to free the implementations completely of having to use
>>> these strings, we'll need to assign numbers. This can be done at
>>> the module level, either manually or algorithmically. Using a
>>> per-module hash to do this is suboptimal; it is much better to
>>> fill a small
linear space.
>>>
>>> We then need a way to compose these numbers with numbers assigned
>>> to module IDs. SMIv2 had this way to name modules and their
>>> components...
>>>
>>> Grüße, Carsten
>
> _______________________________________________
> 6tisch mailing list
> [email protected]
> https://www.ietf.org/mailman/listinfo/6tisch
>
>
_______________________________________________
6tisch mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/6tisch
_______________________________________________
6tisch mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/6tisch
