[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [Syslog-sec] RE: Maximum message size
Hi Rainer,
Thanks for making the changes.
I'd like to make one point a little more strenuously:
"For interoperability reasons, all receiver implementations SHOULD be
able to accept messages up to and including 65,535 octets in length."
Also, I think the last two sentences could be combined into one - "If
a receiver receives a message with a length larger than 65,535 octets,
or larger than it supports, the receiver MAY discard the message or
truncate the payload."
dbh
-----Original Message-----
From: syslog-sec-bounces@www.employees.org
[mailto:syslog-sec-bounces@www.employees.org] On Behalf Of Rainer
Gerhards
Sent: Thursday, October 14, 2004 7:34 AM
To: syslog-sec@employees.org
Subject: [Syslog-sec] RE: Maximum message size
David and all,
many thanks for all the good comments.
I agree that 16 MB for a single message is far more than actually
should be used. Let's look a bit at the history of this:
Initially, we noticed that the 1024 octet limit for current syslog is
too short. I think there is still general agreement on this. We had a
good discussion about the size issue in September 2003. A link to an
archive is here:
http://www.syslog.cc/ietf/autoarc/msg00855.html
The general issue is that 1K is too short for many applications. For
example, Marshall Glen mentions the healthcare needs (RFC3881) for a
larger message. As far as I know, the IHE initiative currently
standardizes on a maximum XML stream size of 32K (which is then
supposed to be transmitted via syslog). I think a message size limit
of 64K would probably be suffcient for a long time to come (and leaves
some headroom for message content increases).
Besides that, I myself identified a need - in very rare occasions - to
send information larger than 64K. In extreme cases, this information
could be as large as one megabyte. Again, these were cases that might
happen once in a year. All of this, of course, only if an operator
*really* wants to submit that large information. So the initial
approach and suggestion was to introduce a way to build "message
groups". That would have been a way to split "oversize messages" (as
they were called
then) into multiple syslog messages. They key to that concept was that
each single message could be kept within a reasonable limit (it would
have worked even with 1K max message size). It would only have been a
way to transmit very large messages in those very seldom cases that
they would actually be used.
Some time later, we came to the conclusion that it should be possible
for a single syslog message to be of the largest imaginable size. It
was argued that fragmentation is a transport issue and as such
application-level "fragmentation" (the message groups) would not do
good. We agreed to that.
Even some time later, we needed to define a maximum message size, just
to limit it somewhere (if I recall correctly, that was a transport
requirement). We picked 16MB just as a number that we thought would
never have be reached and such be sufficient for all imaginable time
to come. As far as I was concerend, that decision was driven by the
fact that in the past there were often limits that nobody expected to
reach - and yet they wer quickly reached.
Som 16MB was picked more or less as "larger than anybody will need
anytime in the forseeable future". We did not even have a big
discussion on that limit.
I think this are the key points on how we reached this limit. Now,
thanks to your call and looking the whole situation, I really think
we've gone overboard. That 16MB limit is something nobody intends to
use. Even a lower - but high limit - is extremely unlikely. I'd expect
to see a 1MB "message" in maybe 0.0001% of the cases - if at all (but
I expect to see e.g. 70KB "message" in maybe 0.001% of the cases).
Also, I have read your proposal to set a maxium size for the message,
but allow an implementation to go beyond that. Maybe something like
this:
####
5.1 Message Length
A receiver MUST be able to accept messages up to and including 480
octets in length.
A receiver SHOULD be able to accept messages up to and including
65,535 octets in length.
If a receiver receives a message with a length larger than it
supports, the receiver MAY
discard the message or truncate the payload.
A receiver MAY accept messages larger than 65,535 octets.
####
I've dropped the wording that transports MAY support only a smaller
max size - I think UDP was the most critical in this regard and it
obviously handles 64K. So I assume this clause is not really needed if
we go for 64K as the max size.
With that clause, we would limit the size for all interop cases, but
leave a window open so that the message length could be extended if
there is a real need AND the operator and software vendor wants this.
I assume nobody will implement this burden if there is no market need
to do so. But if it is, it clould at least be implemented in a
standards-compliant way.
If I look back at the original requirement - provide a way to transmit
very rare oversize messages, I think we can safely implement this the
bounds of the current proposals, even if the size limit does not go
beyond 64K. If somebody needs it, this can be easily done via an
experimental (or vendor-specific) SD-ID. Again, I do not think
somebody will accept this burden without market need.
In the light of this, I, too, would recommend that we limit the
message size to 64K. I would prefer, however, that we keep the maximum
open as in the suggested wording above.
Rainer
> -----Original Message-----
> From: David B Harrington [mailto:ietfdbh@comcast.net]
> Sent: Wednesday, October 13, 2004 5:10 PM
> To: Rainer Gerhards; 'Anton Okmianski'; syslog-sec@employees.org
> Subject: Maximum message size
>
> Hi,
>
> I am concerned that we are in danger of promoting uses of syslog
that
> will defeat its current usefulness as a human-readable log of system
> activity.
>
> Do network operators really believe the 16,777,216 octet size is a
> needed improvement to syslog, or are we designing this size in
> explicitly for the vendors of applications who want to use syslog as
a
> programmatic interface? I believe Syslog should be designed to meet
> the needs of operators. Most of the discussion ssems to be from
> vendors of syslog receivers or applications. Other protocols such as
> FTP might be better used for such a specialty use case.
>
> Does allowing messages of 16,777,216 octets to be accumulated within
> the system log make it harder to use some widely-available minimal
> text editors, like Notepad, to view the system log? Will having huge
> application-specific messages in the system log make it harder for
an
> operator to locate useful troubleshooting messages within the system
> log? Can the information in such a huge message fit within an 80x25
> screen, when an operator is troubleshooting network problems and
needs
> to use a bare-bones terminal attached to the serial port of a device
> to view the system log?
>
> Syslog was designing to be an operator's tool, and Syslog should be
> designed to meet the needs of operators. Will allowing messages of
> this size negatively impact its usefulness in the primary use case?
>
> What do network **operators** think about the need for messages of
> this size? Have we deliberately asked them their opinion on this by,
> for example, taking a survey or going to NANOG to ask them?
>
> I don't think we should support such large messages in the system
> logging protocol.
>
>
> David Harrington
> dbharrington@comcast.net
>
>
>
> -----Original Message-----
> From: syslog-sec-bounces@www.employees.org
> [mailto:syslog-sec-bounces@www.employees.org] On Behalf Of Rainer
> Gerhards
> Sent: Wednesday, October 13, 2004 5:17 AM
> To: Anton Okmianski; syslog-sec@employees.org
> Subject: RE: [Syslog-sec] UDP message size issue proposal
>
> Anton,
>
> I agree to your conclusion.
>
> Although I am the one who always voted for larger sizes, the
> discussion has shown that this yields to unnecessary complexity. I
can
> satisfy my needs with a vendor-extension structured data element,
> which I think shows the flexibility of the new drafts.
>
> So I support your move. I would tend to even remove the transport
> version header. If there is need to, we could always include that in
a
> later release (e.g. "v1", which would make it clearly distingshable
> from the current frame format). I see little need to do so, though.
>
> Regarding -protocol, I think we should still keep an upper limit in
> the spec. Even I don't see any reason why a syslog message should go
> above 16MB. So for -protocol, I propose the following new text:
>
> ####
> 5.1 Message Length
>
> The maximum length of any syslog message is 16,777,216 octets.
Any
> receiver receiving a larger message MUST discard the message. A
> diagnostic message SHOULD be logged in this case.
>
> A receiver MUST be able to accept messages that are 480 octets
(or
> less) in length. A receiver SHOULD be able to accept messages
that
> are 65,535 octets (or less) in length. It is RECOMMENDED that
> receivers be able to accept messages up to the maximum message
> length
> of 16,777,216 octets.
>
> If a receiver receives a message within the maximum length, but
> with
> a length larger than it handles, the receiver MAY discard or
> truncate
> it.
>
> A transport mapping MAY define a maximum length below the one
> specified here. Each transport mapping MUST support a maximum
size
> of 480 octets or more.
> ####
>
> If there is agreement on this, I can post a new version of
-protocol.
> That version will also include some changes made thanks to Andrew
Ross
> comments (sent via private mail). I have finished this version. It
is
> available for immediate posting, but I would like to wait for
> agreement on the size issue (at least if that can be reached
quickly).
>
> Rainer
>
> > -----Original Message-----
> > From: syslog-sec-bounces@www.employees.org
> > [mailto:syslog-sec-bounces@www.employees.org] On Behalf Of Anton
> > Okmianski
> > Sent: Tuesday, October 12, 2004 11:25 PM
> > To: syslog-sec@employees.org
> > Subject: [Syslog-sec] UDP message size issue proposal
> >
> > Hi!
> >
> > Sorry for a long delay on this issue - I was on a 2 week vacation.
> I
> > have spoken with a number of TCP/UDP/IP experts regarding the
sizing
>
> > issue. I am ready to propose the following changes:
> >
> > 1. Syslog-protocol will state that the max message will be defined
> by
> > the transport layer.
> >
> > 2. Syslog-transport-udp will support messages up to maximum UDP
> > datagram size of 64K. UDP is a very bad choice for large message
> > transmissions, so it does not make sense for us to stretch it by
> > adding our own fragmentation without other transmission control
> > features such as found in TCP.
> >
> > 3. Syslog-transport-udp will rely on IP fragmentation and we will
> get
> > rid of "proprietary" fragmentation which was designed to handle
> > messages over 64K and deal with various
> > non-compliant network hosts.
> >
> > 4. Syslog-transport-udp will recommend sending messages within the
> > boundaries of prevalent MTU size on a given network to be safe. It
> > will recommend Ethernet's 1500 bytes less headers and will draw
> reader
> > attention to the minimum MTU size hosts on the network are
required
> to
> > support for
> > IPv6 (576 bytes) and IPv6 (1280 bytes).
> >
> > 5. Path MTU discovery may not work robustly and some TCP/IP stacks
> may
> > not support UDP packets of full 64K size and truncate them. We
will
>
> > mention this and bite this bullet.
> > We should not restrict the protocol due to incompliant
> implementations
> > because it is a moving target and penalizes compliant
> implementations
> > with extra overhead. The above size recommendation would partially
> > deal with this, but leave the
> > final choice up to the administrator.
> >
> > 6. We will get rid of most syslog transport headers for UDP as
they
> > will no longer be needed. The only thing that will be left is the
> > transport protocol version prefixed to every syslog message.
Should
>
> > we even bother with that?
> >
> > This is a major change to the syslog-transport-udp. I'd like to
get
>
> > positive feedback before I proceed with this update.
> > The best part is that if we all agree on the above, the next draft
> > will be 1/3 of the size -- easier read for you. :)
> >
> > Thanks,
> > Anton.
> >
> >
> > _______________________________________________
> > Syslog-sec mailing list
> > Syslog-sec@www.employees.org
> > http://www.employees.org/mailman/listinfo/syslog-sec
> >
> _______________________________________________
> Syslog-sec mailing list
> Syslog-sec@www.employees.org
> http://www.employees.org/mailman/listinfo/syslog-sec
>
>
>
_______________________________________________
Syslog-sec mailing list
Syslog-sec@www.employees.org
http://www.employees.org/mailman/listinfo/syslog-sec
_______________________________________________
Syslog-sec mailing list
Syslog-sec@www.employees.org
http://www.employees.org/mailman/listinfo/syslog-sec