[Top][All Lists]

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

Re: [Discuss-gnuradio] comments on stream tags and metadata storage

From: Marcus Müller
Subject: Re: [Discuss-gnuradio] comments on stream tags and metadata storage
Date: Fri, 25 Jul 2014 15:46:57 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:24.0) Gecko/20100101 Thunderbird/24.5.0

Hi Peter,

I agree that this is a very relevant topic, and especially the
performance of tag handling might prove to be problematic soon...
However, it's a bit hard to start a discussion like that; a lot of
things in GNU Radio are like they are because someone wrote them like
that, and they proved to just work, or if they didn't, they got remodeled.
That being said, your mail was very long, and it took me multiple
sessions to read it. I've now decided to share my reply as partial as it is.

sooo let me just whip up a few comments:
(1) That's a documentation issue, isn't it? Anyway, I'm not quite sure
you're right; the insertion of tags is mutexed IIRC, and the
get_tags_in_range() functionality, too, so once the user got his vector
of tags, that won't change anymore. There's the possibility that he
*misses* some of the tags for the range that get inserted after he
got_tags_in_range(), but that's only fair -- it's quite intuitive not to
insert tags after you've handed off the samples to which they would
belong to downstream tags.

(2) that's an interesting point.
> In the current implementation it's further necessary that tags be
added to an output in monotonic non-decreasing offset order.
Uh, that's news to me, can you point me to the reason? If a block
assumes things to be ordered, but they aren't... again, this is not
well-documented, so you're right for raising this issue!

I'm a bit worried that you always suffer at one end: If tags are always
stored ordered, than inserting tags gains computational complexity, even
if the getter doesn't need them sorted.

In my opinion we shouldn't go for the "generate tags only in work()"
because that would increase the complexity of the insertion (inserting
would have to device a check that it's being called from work, or this
will only be a contract...) and is kind of unnecessary. A block always
(even outside of work()) has access to nitems_written() so it's always
able to avoid generating tags for samples that might already have been
read downstream.

(3) I don't see the relation to the discussion, as you said ;) but that
sounds like a bug, so if you opened up a new thread or filed a bug at
gnuradio.org, that would be awesome :)

(4) I'm fairly certain the buffers use a deque to store tags, not a map
of any kind. So maybe I'm misunderstanding you, or you misread code
I think what you're describing might be a bug in metadata_filesink, so
that might need some attention! see (3).

(5)>(5) All stream tags are placed in the extras block,
sorry, can't follow you there. Extras block?


On 25.07.2014 13:00, Peter A. Bigot wrote:
> I'd hoped my comments below would start a more extensive dialog on GNU
> Radio's metadata infrastructure. Several years experience that I have
> with this capability in a non-commercial C++ DSP framework suggests
> many enhancements in flow, representation, and utilities.
> I have a slight itch to contribute to a solution, but without
> community involvement can't hope to provide anything mergable.  Is
> this simply not something anybody feels needs to be addressed, or did
> I ask in the wrong forum?
> Peter
> On 07/17/2014 05:11 PM, Peter A. Bigot wrote:
>> Some comments after playing with stream tags and metadata this
>> afternoon.
>> (1) Although the discussion of stream tag insertion hints that this
>> should be done within the scheduler's call to work() it could be more
>> clear that doing it in any other context can result in race conditions.
>> (I did think I saw it stated more clearly somewhere, but can't find
>> that now, so maybe this point has been addressed.)
>> (2) In the current implementation it's further necessary that tags be
>> added to an output in monotonic non-decreasing offset order.
>> file_meta_sink does not sort the return value from get_tags_in_range(),
>> and emits all data up to the timestamp of the next tag, so a subsequent
>> tag with an earlier offset is dropped from the archive.
>> (I note that tagged_file_sink() does sort the tags it receives in one
>> case, but not in others.)
>> I don't see this requirement on ordered generation documented.  In some
>> cases, it may be inconvenient to do this, e.g. when a block's analysis
>> discovers after-the-fact that something interesting can be associated
>> with a past sample.  Similarly, a user might want a block to associate
>> a tag with sample that not yet arrived, to notify a downstream block
>> that will need to process the event.
>> A simple solution for the infrastructure is to require that tags only be
>> generated from within work(), with offsets corresponding to samples
>> generated in that call to work(), and in non-decreasing offset order
>> (though this last requirement could be handled by add_item_tag()).  The
>> developer must then handle the too-late/too-early tag associations
>> through some other mechanism, such as carrying the effective offset as
>> part of the tag value.
>> (3) Qt GUI Range with widget Counter + Slider invokes callbacks twice,
>> even if the value itself was set exactly once through the counter text
>> entry.  If the callback records the change by queuing a stream tag for
>> addition to the output, multiple tags with the same offset/key/value
>> will be generated.
>> There are ugly solutions to this but it's probably sufficient to note
>> somewhere that it can happen.  It's really not specific to tags, but is
>> clearly visible in that case.
>> (4) The in-memory stream of tags can produce multiple settings of the
>> same key at the same offset.  However, when stored to a file only the
>> last setting of the key is recorded.
>> I believe this last behavior is incorrect and that it's a mistake to use
>> a map instead of a multimap or simple list for the metadata record of
>> stream tags associated with a sample.
>> One argument is that it's critical that a stream archive of a processing
>> session faithfully record the contents of the stream so that re-running
>> the application using playback reproduces that stream and thus the
>> original behavior (absent non-determinism due to asynchrony). This
>> faithful reproduction is what would allow a maintainer to diagnose an
>> operational failure caused by a block with a runtime failure when the
>> same tag is processed twice at the same offset.  This is true even if
>> the same key is set to the same value at the same sample offset multiple
>> times, which some might otherwise want to argue is redundant.
>> A corollary argument is that the sample number at which an event like a
>> tuner configuration change occurs usually can't be exactly associated
>> with a sample; the best estimate is likely to be the index of the first
>> sample generated by the next call to work.  But depending on processing
>> speed an application might change an attribute of a data source multiple
>> times before work was invoked.  The effect of those intermediate changes
>> may be visible in the signal, and to lose the fact they occurred by
>> discarding all but the last change affects both reproducibility and
>> interpretation of the signal itself.
>> (5) All stream tags are placed in the extras block, and when a segment
>> is completed file_meta_sink will generate a new header.  The new header
>> contains copies of the unique tags, but updates their offsets to be the
>> start of the new segment.
>> This is incorrect as the original stream did not have those tags
>> associated with those samples, so re-playing will introduce a behavioral
>> difference.  For example, a tag that is meant to be associated with the
>> start of a packet will be duplicated at an offset that is probably not
>> the start of a packet.
>> Solutions include (a) leave the original offset setting for tags in the
>> extras section when they're reproduced in a new segment, even though
>> that offset is not present in the segment; (b) treat stream tags as
>> ephemeral and do not persist them in the extras section when generating
>> a new segment; (c) extend the add_item_tag API to record whether the
>> tag is ephemeral or persistent.  Offhand I can see no argument
>> supporting persisting a tag and updating its offset, and only rare cases
>> where it's appropriate to replicate outdated information in a new
>> segment, so (b) seems to be the right move.
>> All the above is based on my understanding and expectations of how
>> stream  tags are/should be used.  If my understanding is mistaken,
>> please let me know.
>> Peter
>> _______________________________________________
>> Discuss-gnuradio mailing list
>> address@hidden
>> https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
> _______________________________________________
> Discuss-gnuradio mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/discuss-gnuradio

reply via email to

[Prev in Thread] Current Thread [Next in Thread]