Re: Are dumb segments dumb?

[Francesco Mari]

It might be a variation in the process I tried. This shouldn't affect
much the statistics anyway, given that the population sample is big
enough in both cases.

2016-07-25 17:46 GMT+02:00 Michael Dürig :
>
> Interesting numbers. Most of them look as I would have expected. I.e. the
> distributions in the dumb case are more regular (smaller std. dev, mean and
> median closer to each other), bigger segment sizes, etc.
>
> What I don't understand is the total number of records. These numbers differ
> greatly between current and dumb. Is this a test artefact (i.e. test not
> reproducible) or are we missing out on something.
>
> Michael
>
>
> On 25.7.16 4:01 , Francesco Mari wrote:
>>
>> I put together some statistics [1] for the process I described above.
>> The "dumb" variant requires more segments to store the same amount of
>> data, because of the increased size of serialised record IDs.  As you
>> can see the amount of records per segment is definitely lower in the
>> dumb variant.
>>
>> On the other hand, ignoring the growth of segment ID reference table
>> seems to be a good choice. As shown from the segment size average,
>> dumb segments are usually fuller that their counterpart. Moreover, a
>> lower standard deviation shows that it's more common to have full dumb
>> segments.
>>
>> In addition, my analysis seems to have found a bug too. There are a
>> lot of segments with no segment ID references and only one record,
>> which is very likely to be the segment info. The flush thread writes
>> every 5 seconds the current segment buffer, provided that the buffer
>> is not empty. It turns out that a segment buffer is never empty, since
>> it always contains at least one record. As such, we are currently
>> leaking almost empty segments every 5 seconds, that waste additional
>> space on disk because of the padding required by the TAR format.
>>
>> [1]:
>> https://docs.google.com/spreadsheets/d/1gXhmPsm4rDyHnle4TUh-mtB2HRtRyADXALARRFDh7z4/edit?usp=sharing
>>
>> 2016-07-25 10:05 GMT+02:00 Michael Dürig :
>>>
>>>
>>> Hi Jukka,
>>>
>>> Thanks for sharing your perspective and the historical background.
>>>
>>> I agree that repository size shouldn't be a primary concern. However, we
>>> have seen many repositories (especially with an external data store)
>>> where
>>> the content is extremely fine granular. Much more than in an initial
>>> content
>>> installation of CQ (which I believe was one of the initial setup for
>>> collecting statistics). So we should at least understand the impact of
>>> the
>>> patch in various scenarios.
>>>
>>> My main concern is the cache footprint of node records. Those are made up
>>> of
>>> a list of record ids and would thus grow by a factor of 6 with the
>>> current
>>> patch.
>>>
>>> Locality is not so much of concern here. I would expect it to actually
>>> improve as the patch gets rid of the 255 references limit of segments. A
>>> limit which in practical deployments leads to degeneration of segment
>>> sizes
>>> (I regularly see median sizes below 5k). See OAK-2896 for some background
>>> on
>>> this.
>>> Furthermore we already did a big step forward in improving locality in
>>> concurrent write scenarios when we introduced the
>>> SegmentBufferWriterPool.
>>> In essence: thread affinity for segments.
>>>
>>> We should probably be more carefully looking at the micro benchmarks. I
>>> guess we neglected this part a bit in the past. Unfortunately CI
>>> infrastructure isn't making this easy for us... OTOH those benchmarks
>>> only
>>> tell you so much. Many of the problems we recently faced only surfaced in
>>> the large: huge repos, high concurrent load, many days of traffic.
>>>
>>> Michael
>>>
>>>
>>>
>>>
>>>
>>> On 23.7.16 12:34 , Jukka Zitting wrote:


 Hi,

 Cool! I'm pretty sure there are various ways in which the format could
 be
 improved, as the original design was based mostly on intuition, guided
 somewhat by collected stats
 
 and
 the micro-benchmarks 
 used
 to optimize common operations.

 Note though that the total size of the repository was not and probably
 shouldn't be a primary metric, since the size of a typical repository is
 governed mostly by binaries and string properties (though it's a good
 idea
 to make sure you avoid things like duplicates of large binaries).
 Instead
 the rationale for squeezing things like record ids to as few bytes as
 possible is captured in the principles listed in the original design doc
 :

- Compactness. The formatting of records is optimized for size to
 reduce
IO costs and to fit as much content in caches as possible. A node
 stored in
SegmentNodeStore typically consumes only a fraction of the size it

Re: Are dumb segments dumb?

[Michael Dürig]

Interesting numbers. Most of them look as I would have expected. I.e. 
the distributions in the dumb case are more regular (smaller std. dev, 
mean and median closer to each other), bigger segment sizes, etc.


What I don't understand is the total number of records. These numbers 
differ greatly between current and dumb. Is this a test artefact (i.e. 
test not reproducible) or are we missing out on something.


Michael

On 25.7.16 4:01 , Francesco Mari wrote:

I put together some statistics [1] for the process I described above.
The "dumb" variant requires more segments to store the same amount of
data, because of the increased size of serialised record IDs.  As you
can see the amount of records per segment is definitely lower in the
dumb variant.

On the other hand, ignoring the growth of segment ID reference table
seems to be a good choice. As shown from the segment size average,
dumb segments are usually fuller that their counterpart. Moreover, a
lower standard deviation shows that it's more common to have full dumb
segments.

In addition, my analysis seems to have found a bug too. There are a
lot of segments with no segment ID references and only one record,
which is very likely to be the segment info. The flush thread writes
every 5 seconds the current segment buffer, provided that the buffer
is not empty. It turns out that a segment buffer is never empty, since
it always contains at least one record. As such, we are currently
leaking almost empty segments every 5 seconds, that waste additional
space on disk because of the padding required by the TAR format.

[1]: 
https://docs.google.com/spreadsheets/d/1gXhmPsm4rDyHnle4TUh-mtB2HRtRyADXALARRFDh7z4/edit?usp=sharing

2016-07-25 10:05 GMT+02:00 Michael Dürig :


Hi Jukka,

Thanks for sharing your perspective and the historical background.

I agree that repository size shouldn't be a primary concern. However, we
have seen many repositories (especially with an external data store) where
the content is extremely fine granular. Much more than in an initial content
installation of CQ (which I believe was one of the initial setup for
collecting statistics). So we should at least understand the impact of the
patch in various scenarios.

My main concern is the cache footprint of node records. Those are made up of
a list of record ids and would thus grow by a factor of 6 with the current
patch.

Locality is not so much of concern here. I would expect it to actually
improve as the patch gets rid of the 255 references limit of segments. A
limit which in practical deployments leads to degeneration of segment sizes
(I regularly see median sizes below 5k). See OAK-2896 for some background on
this.
Furthermore we already did a big step forward in improving locality in
concurrent write scenarios when we introduced the SegmentBufferWriterPool.
In essence: thread affinity for segments.

We should probably be more carefully looking at the micro benchmarks. I
guess we neglected this part a bit in the past. Unfortunately CI
infrastructure isn't making this easy for us... OTOH those benchmarks only
tell you so much. Many of the problems we recently faced only surfaced in
the large: huge repos, high concurrent load, many days of traffic.

Michael





On 23.7.16 12:34 , Jukka Zitting wrote:


Hi,

Cool! I'm pretty sure there are various ways in which the format could be
improved, as the original design was based mostly on intuition, guided
somewhat by collected stats 
and
the micro-benchmarks  used
to optimize common operations.

Note though that the total size of the repository was not and probably
shouldn't be a primary metric, since the size of a typical repository is
governed mostly by binaries and string properties (though it's a good idea
to make sure you avoid things like duplicates of large binaries). Instead
the rationale for squeezing things like record ids to as few bytes as
possible is captured in the principles listed in the original design doc
:

   - Compactness. The formatting of records is optimized for size to
reduce
   IO costs and to fit as much content in caches as possible. A node
stored in
   SegmentNodeStore typically consumes only a fraction of the size it
would as
   a bundle in Jackrabbit Classic.
   - Locality. Segments are written so that related records, like a node
   and its immediate children, usually end up stored in the same segment.
This
   makes tree traversals very fast and avoids most cache misses for
typical
   clients that access more than one related node per session.

Thus I would recommend keeping an eye also on benchmark results in
addition
to raw repository size when evaluating possible improvements. Also, the
number and size of data segments are good size metrics to look at in
addition to total disk usage.

BR,

Jukka Zitting

On Fri, Jul 22, 2016 at 

Re: Are dumb segments dumb?

[Francesco Mari]

I opened OAK-4596 to track the segment leak.

2016-07-25 16:01 GMT+02:00 Francesco Mari :
> I put together some statistics [1] for the process I described above.
> The "dumb" variant requires more segments to store the same amount of
> data, because of the increased size of serialised record IDs.  As you
> can see the amount of records per segment is definitely lower in the
> dumb variant.
>
> On the other hand, ignoring the growth of segment ID reference table
> seems to be a good choice. As shown from the segment size average,
> dumb segments are usually fuller that their counterpart. Moreover, a
> lower standard deviation shows that it's more common to have full dumb
> segments.
>
> In addition, my analysis seems to have found a bug too. There are a
> lot of segments with no segment ID references and only one record,
> which is very likely to be the segment info. The flush thread writes
> every 5 seconds the current segment buffer, provided that the buffer
> is not empty. It turns out that a segment buffer is never empty, since
> it always contains at least one record. As such, we are currently
> leaking almost empty segments every 5 seconds, that waste additional
> space on disk because of the padding required by the TAR format.
>
> [1]: 
> https://docs.google.com/spreadsheets/d/1gXhmPsm4rDyHnle4TUh-mtB2HRtRyADXALARRFDh7z4/edit?usp=sharing
>
> 2016-07-25 10:05 GMT+02:00 Michael Dürig :
>>
>> Hi Jukka,
>>
>> Thanks for sharing your perspective and the historical background.
>>
>> I agree that repository size shouldn't be a primary concern. However, we
>> have seen many repositories (especially with an external data store) where
>> the content is extremely fine granular. Much more than in an initial content
>> installation of CQ (which I believe was one of the initial setup for
>> collecting statistics). So we should at least understand the impact of the
>> patch in various scenarios.
>>
>> My main concern is the cache footprint of node records. Those are made up of
>> a list of record ids and would thus grow by a factor of 6 with the current
>> patch.
>>
>> Locality is not so much of concern here. I would expect it to actually
>> improve as the patch gets rid of the 255 references limit of segments. A
>> limit which in practical deployments leads to degeneration of segment sizes
>> (I regularly see median sizes below 5k). See OAK-2896 for some background on
>> this.
>> Furthermore we already did a big step forward in improving locality in
>> concurrent write scenarios when we introduced the SegmentBufferWriterPool.
>> In essence: thread affinity for segments.
>>
>> We should probably be more carefully looking at the micro benchmarks. I
>> guess we neglected this part a bit in the past. Unfortunately CI
>> infrastructure isn't making this easy for us... OTOH those benchmarks only
>> tell you so much. Many of the problems we recently faced only surfaced in
>> the large: huge repos, high concurrent load, many days of traffic.
>>
>> Michael
>>
>>
>>
>>
>>
>> On 23.7.16 12:34 , Jukka Zitting wrote:
>>>
>>> Hi,
>>>
>>> Cool! I'm pretty sure there are various ways in which the format could be
>>> improved, as the original design was based mostly on intuition, guided
>>> somewhat by collected stats 
>>> and
>>> the micro-benchmarks  used
>>> to optimize common operations.
>>>
>>> Note though that the total size of the repository was not and probably
>>> shouldn't be a primary metric, since the size of a typical repository is
>>> governed mostly by binaries and string properties (though it's a good idea
>>> to make sure you avoid things like duplicates of large binaries). Instead
>>> the rationale for squeezing things like record ids to as few bytes as
>>> possible is captured in the principles listed in the original design doc
>>> :
>>>
>>>- Compactness. The formatting of records is optimized for size to
>>> reduce
>>>IO costs and to fit as much content in caches as possible. A node
>>> stored in
>>>SegmentNodeStore typically consumes only a fraction of the size it
>>> would as
>>>a bundle in Jackrabbit Classic.
>>>- Locality. Segments are written so that related records, like a node
>>>and its immediate children, usually end up stored in the same segment.
>>> This
>>>makes tree traversals very fast and avoids most cache misses for
>>> typical
>>>clients that access more than one related node per session.
>>>
>>> Thus I would recommend keeping an eye also on benchmark results in
>>> addition
>>> to raw repository size when evaluating possible improvements. Also, the
>>> number and size of data segments are good size metrics to look at in
>>> addition to total disk usage.
>>>
>>> BR,
>>>
>>> Jukka Zitting
>>>
>>> On Fri, Jul 22, 2016 at 5:55 AM Francesco Mari 

Re: Are dumb segments dumb?

[Francesco Mari]

I put together some statistics [1] for the process I described above.
The "dumb" variant requires more segments to store the same amount of
data, because of the increased size of serialised record IDs.  As you
can see the amount of records per segment is definitely lower in the
dumb variant.

On the other hand, ignoring the growth of segment ID reference table
seems to be a good choice. As shown from the segment size average,
dumb segments are usually fuller that their counterpart. Moreover, a
lower standard deviation shows that it's more common to have full dumb
segments.

In addition, my analysis seems to have found a bug too. There are a
lot of segments with no segment ID references and only one record,
which is very likely to be the segment info. The flush thread writes
every 5 seconds the current segment buffer, provided that the buffer
is not empty. It turns out that a segment buffer is never empty, since
it always contains at least one record. As such, we are currently
leaking almost empty segments every 5 seconds, that waste additional
space on disk because of the padding required by the TAR format.

[1]: 
https://docs.google.com/spreadsheets/d/1gXhmPsm4rDyHnle4TUh-mtB2HRtRyADXALARRFDh7z4/edit?usp=sharing

2016-07-25 10:05 GMT+02:00 Michael Dürig :
>
> Hi Jukka,
>
> Thanks for sharing your perspective and the historical background.
>
> I agree that repository size shouldn't be a primary concern. However, we
> have seen many repositories (especially with an external data store) where
> the content is extremely fine granular. Much more than in an initial content
> installation of CQ (which I believe was one of the initial setup for
> collecting statistics). So we should at least understand the impact of the
> patch in various scenarios.
>
> My main concern is the cache footprint of node records. Those are made up of
> a list of record ids and would thus grow by a factor of 6 with the current
> patch.
>
> Locality is not so much of concern here. I would expect it to actually
> improve as the patch gets rid of the 255 references limit of segments. A
> limit which in practical deployments leads to degeneration of segment sizes
> (I regularly see median sizes below 5k). See OAK-2896 for some background on
> this.
> Furthermore we already did a big step forward in improving locality in
> concurrent write scenarios when we introduced the SegmentBufferWriterPool.
> In essence: thread affinity for segments.
>
> We should probably be more carefully looking at the micro benchmarks. I
> guess we neglected this part a bit in the past. Unfortunately CI
> infrastructure isn't making this easy for us... OTOH those benchmarks only
> tell you so much. Many of the problems we recently faced only surfaced in
> the large: huge repos, high concurrent load, many days of traffic.
>
> Michael
>
>
>
>
>
> On 23.7.16 12:34 , Jukka Zitting wrote:
>>
>> Hi,
>>
>> Cool! I'm pretty sure there are various ways in which the format could be
>> improved, as the original design was based mostly on intuition, guided
>> somewhat by collected stats 
>> and
>> the micro-benchmarks  used
>> to optimize common operations.
>>
>> Note though that the total size of the repository was not and probably
>> shouldn't be a primary metric, since the size of a typical repository is
>> governed mostly by binaries and string properties (though it's a good idea
>> to make sure you avoid things like duplicates of large binaries). Instead
>> the rationale for squeezing things like record ids to as few bytes as
>> possible is captured in the principles listed in the original design doc
>> :
>>
>>- Compactness. The formatting of records is optimized for size to
>> reduce
>>IO costs and to fit as much content in caches as possible. A node
>> stored in
>>SegmentNodeStore typically consumes only a fraction of the size it
>> would as
>>a bundle in Jackrabbit Classic.
>>- Locality. Segments are written so that related records, like a node
>>and its immediate children, usually end up stored in the same segment.
>> This
>>makes tree traversals very fast and avoids most cache misses for
>> typical
>>clients that access more than one related node per session.
>>
>> Thus I would recommend keeping an eye also on benchmark results in
>> addition
>> to raw repository size when evaluating possible improvements. Also, the
>> number and size of data segments are good size metrics to look at in
>> addition to total disk usage.
>>
>> BR,
>>
>> Jukka Zitting
>>
>> On Fri, Jul 22, 2016 at 5:55 AM Francesco Mari 
>> wrote:
>>
>>> The impact on repository size needs to be assessed with more specific
>>> tests. In particular, I found RecordUsageAnalyserTest and
>>> SegmentSizeTest unsuitable to this task. It's not a coincidence that
>>> these tests 

Re: Are dumb segments dumb?

[Michael Dürig]

Hi Jukka,

Thanks for sharing your perspective and the historical background.

I agree that repository size shouldn't be a primary concern. However, we 
have seen many repositories (especially with an external data store) 
where the content is extremely fine granular. Much more than in an 
initial content installation of CQ (which I believe was one of the 
initial setup for collecting statistics). So we should at least 
understand the impact of the patch in various scenarios.


My main concern is the cache footprint of node records. Those are made 
up of a list of record ids and would thus grow by a factor of 6 with the 
current patch.


Locality is not so much of concern here. I would expect it to actually 
improve as the patch gets rid of the 255 references limit of segments. A 
limit which in practical deployments leads to degeneration of segment 
sizes (I regularly see median sizes below 5k). See OAK-2896 for some 
background on this.
Furthermore we already did a big step forward in improving locality in 
concurrent write scenarios when we introduced the 
SegmentBufferWriterPool. In essence: thread affinity for segments.


We should probably be more carefully looking at the micro benchmarks. I 
guess we neglected this part a bit in the past. Unfortunately CI 
infrastructure isn't making this easy for us... OTOH those benchmarks 
only tell you so much. Many of the problems we recently faced only 
surfaced in the large: huge repos, high concurrent load, many days of 
traffic.


Michael




On 23.7.16 12:34 , Jukka Zitting wrote:

Hi,

Cool! I'm pretty sure there are various ways in which the format could be
improved, as the original design was based mostly on intuition, guided
somewhat by collected stats  and
the micro-benchmarks  used
to optimize common operations.

Note though that the total size of the repository was not and probably
shouldn't be a primary metric, since the size of a typical repository is
governed mostly by binaries and string properties (though it's a good idea
to make sure you avoid things like duplicates of large binaries). Instead
the rationale for squeezing things like record ids to as few bytes as
possible is captured in the principles listed in the original design doc
:

   - Compactness. The formatting of records is optimized for size to reduce
   IO costs and to fit as much content in caches as possible. A node stored in
   SegmentNodeStore typically consumes only a fraction of the size it would as
   a bundle in Jackrabbit Classic.
   - Locality. Segments are written so that related records, like a node
   and its immediate children, usually end up stored in the same segment. This
   makes tree traversals very fast and avoids most cache misses for typical
   clients that access more than one related node per session.

Thus I would recommend keeping an eye also on benchmark results in addition
to raw repository size when evaluating possible improvements. Also, the
number and size of data segments are good size metrics to look at in
addition to total disk usage.

BR,

Jukka Zitting

On Fri, Jul 22, 2016 at 5:55 AM Francesco Mari 
wrote:


The impact on repository size needs to be assessed with more specific
tests. In particular, I found RecordUsageAnalyserTest and
SegmentSizeTest unsuitable to this task. It's not a coincidence that
these tests are usually the first to be disabled or blindly updated
every time a small fix changes the size of the records.

Regarding GC, the segment graph could be computed during the mark
phase. Of course, it's handy to have this information pre-computed for
you, but since the record graph is traversed anyway we could think
about dynamically reconstructing the segment graph when needed.

There are still so many questions to answer, but I think that this
simplification exercise can be worth the effort.

2016-07-22 11:34 GMT+02:00 Michael Dürig :


Hi,

Neat! I would have expected a greater impact on the size of the segment
store. But as you say it probably all depends on the binary/content

ratio. I

think we should look at the #references / repository size ratio for
repositories of different structures and see how such a number differs

with

and without the patch.

I like the patch as it fixes OAK-2896 while at the same time reducing
complexity a lot.

OTOH we need to figure out how to regain the lost functionality (e.g. gc)
and asses its impact on repository size.

Michael



On 22.7.16 11:32 , Francesco Mari wrote:


Hi,

Yesterday I took some time for a little experiment: how many
optimisations can be removed from the current segment format while
maintaining the same functionality?

I made some work in a branch on GitHub [1]. The code on that branch is
similar to the current trunk except for the following changes:

1. Record IDs are always