Hi Andrei,
Le 26/02/2017 à 23:48, Andrei Chis a écrit :
Hi Thierry,
Thanks for the measurements. Things are more clear now :)
On Fri, Feb 24, 2017 at 11:52 PM, Thierry Goubier
<[email protected] <mailto:[email protected]>> wrote:
Hi Andrei,
Le 24/02/2017 à 17:02, Andrei Chis a écrit :
Hi Thierry,
Strangely enough I'm getting different times on my machine
Just to start from the same baseline in a fresh Pharo 60411
image with
no changes to any of the inspectors over a series of runs I get:
array := (1 to: 1000000) asArray.
[array inspect] timeToRun. "0:00:00:00.145"
[GTInspector inspect: array] timeToRun. "0:00:00:00.031"
Yes, this is expected. I'm not comparing to the standard
EyeInspector, but with my experiments with a FastTable-derived
widget and the EyeInspector model. GTInspector is a lot faster than
the normal EyeInspector.
As default (the only one displaying 100k elements is the AltInspector)
| array |
array := (1 to: 1000000) asArray.
[AltInspector inspect: array ] timeToRun. "0:00:00:00.096"
[EyeInspector inspect: array] timeToRun. "0:00:00:00.364"
[GTInspector inspect: array] timeToRun. "0:00:00:00.065"
Indeed this is what I expected.
If I change indexableDisplayLimit to 50000 and remove the Items view
then I get:
[GTInspector inspect: array] timeToRun. "0:00:00:00.124"
[GTInspector inspect: array] timeToRun. "0:00:00:00.256"
So this time is to be compared to the "0:00:00:00.096".
Can you just also try to remove the Items view and set
indexableDisplayLimit to 50.
In the previous times, the Item view was removed. If
indexableDisplayLimit is set at 50, then the time become:
[GTInspector inspect: array] timeToRun. "0:00:00:00.156"
Which is interesting in itself: multiplying the limit by 1000 (from 50
to 50k) increase the time by a factor of ~1.6.
I think there is some small performance degradation in the current tree
data source used by Glamour but I'm not 100% sure.
It looks pretty efficient to me. There isn't much increase in times.
I looked at the scaling in my experiment: the lower bound in opening the
inspector is reached for a 100k elements array. Anything smaller
has the same time (~80ms). The 1000k array increase that by ~15/20ms.
I'm really interested in seeing what makes GTInspector slower on
your
machine. Did you do other optimizations to SpecInspector?
The optimisations are:
- remove Spec (revert to a pure morphic application),
- use a FastTable-derived widget for a tree view (with a 100k
element limit),
- have the GT views as subitems in the widget.
Nice. Would be also interesting to see if in the future we can merge
more the models used by GTInspector and EyeInspector.
Could be. I just used them out of convenience, and I needed an inspector
with the ability to view large trees.
For now, for the Raw view I set the limit to 5000 elements. This seems
fast/large enough and it allows to add the automatic refresh back to the
Raw view.
I wonder if it wouldn't be possible to tune the automatic refresh.
Increase the time interval if it takes to long (for example when
refreshing a set with 1M elements). But 5000 elements is already a good
limit.
Regards,
Thierry
Cheers,
Andrei
I'm using a MacBook Pro Retina - 2.8 GHz Intel Core i7
Back on a Acer Chromebook C720p 1.4GHz Intel Celeron, with vmLatest.
5.0-201702231204 Thu Feb 23 12:36:20 UTC 2017 gcc 4.6.3 [Production
Spur ITHB VM]
CoInterpreter * VMMaker.oscog-EstebanLorenzano.2136 uuid:
40534c32-ca6b-4e97-91ec-31d509e49b0c Feb 23 2017
StackToRegisterMappingCogit * VMMaker.oscog-EstebanLorenzano.2136
uuid: 40534c32-ca6b-4e97-91ec-31d509e49b0c Feb 23 2017
VM: 201702231204 https://github.com/pharo-project/pharo-vm.git
<https://github.com/pharo-project/pharo-vm.git> $ Date: Thu Feb 23
13:04:59 2017 +0100 $
Plugins: 201702231204 https://github.com/pharo-project/pharo-vm.git
<https://github.com/pharo-project/pharo-vm.git> $
Linux testing-docker-b6b0368d-4817-4638-86be-f022b8a58580
4.8.12-040812-generic #201612020431 SMP Fri Dec 2 09:33:31 UTC 2016
i686 i686 i386 GNU/Linux
Pharo VM version:
CoInterpreter VMMaker.oscog-HolgerHansPeterFreyther.1880 uuid:
16138eb3-2390-40f5-a6c8-15f0494936f8 Oct 10 2016
StackToRegisterMappingCogit
VMMaker.oscog-HolgerHansPeterFreyther.1880
uuid: 16138eb3-2390-40f5-a6c8-15f0494936f8 Oct 10 2016
[email protected]:pharo-project/pharo-vm.git Commit:
06744effac0f0aa3b4b32e17636448f9d51d6707 Date: 2016-09-30
08:40:43 +0200
By: GitHub <[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>>
Cheers,
Andrei
On Fri, Feb 24, 2017 at 3:24 PM, Thierry Goubier
<[email protected] <mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>> wrote:
2017-02-24 14:29 GMT+01:00 Andrei Chis
<[email protected] <mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>>:
On Fri, Feb 24, 2017 at 12:16 PM, Thierry Goubier
<[email protected]
<mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>>
wrote:
Hi Andrei,
2017-02-24 11:31 GMT+01:00 Andrei Chis
<[email protected]
<mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>>:
Hi Thierry,
Indeed that's the simplest option now that we
are using
fast table.
Just now in the case of the Raw view an
OrderedCollection is used to store all displayed
elements.
If you display large collections every time you
open the
Raw view it will instantiate a collection of
size 100k
and instantiate 100k objects of
type GTInspectorIndexedNode. With FastTable we can
lazily load elements so we should be able to
remove this
behaviour and the limit. Just we need to make
sure it
will play nicely with automatic refresh. There
is also
the issue that when expanding an element in the
tree if
it's a collection you don't want to expand all
elements.
I'm not sure you need to worry too much about that
one; in
practical experiments, creating that 100k collection for
viewing (and the associated nodes instances) isn't too
costly (unless creating the GTInspectorIndexedNodes has
hidden costs: I've only experimented with the
EyeInspector
framework).
There should be no hidden costs in GTInspectorIndexedNodes.
I made some experiments in the latest Pharo version and
opening
the Raw view on an array with one million numbers takes
around
120ms when 100k elements are computed.
I'll be curious how much it takes on your machine. To test
update indexableDisplayLimit to 50000 in
Object>>#gtInspectorVariableNodesIn: and remove the
annotation
from Collection>>#gtInspectorItemsIn: (so that the Items
presentation is not loaded)
arrayLarge := (1 to: 1000000) asArray.
arrayLarge inspect.
This is the values I get on my work laptop (core i3-2350M
2.30 Ghz);
both inspectors displays 100k elements.
(1 to: 1000000) asArray in: [ :s | [s inspect] timeToRun]
0:00:00:00.064
(1 to: 1000000) asArray in: [:s | [GTInspector inspect: s]
timeToRun] 0:00:00:00.381
Pharo 6 60411
Opening the tree with all elements works fine in my
experiments. Tuning scrolling as done in Bloc is
necessary,
however.
I'm looking now on a lazy data source for
FastTable that
plays nicely with GTInspector. Let's see how it
will go.
Help is always welcomed :)
As I said: do not overoptimize that part... just
remove that
limitation on the raw view and measure.
If I measure the Items view on the previous array it takes
around 35ms.
What I'd like to have is the same opening time for the Items
view on large Sets and Dictionaries.
On my machine, the experiment is that displaying the set is
fast,
but the system becomes totally unresponsive... which may be
an issue
with the self refresh of the inspector. Yes, it was the culprit.
(1 to: 1000000) asSet in: [ :s | [s inspect] timeToRun]
0:00:00:00.034
(1 to: 1000000) asSet in: [:s | [GTInspector inspect: s]
timeToRun]
0:00:00:00.199
Regards,
Thierry
I think I used the word paginator in the wrong
way. If
you have a very large collection (millions of
elements)
I do not want to scroll through elements but
most likely
jump to a certain element or view just a subset
of all
elements. Not really add a paginator like in web
pages.
Ok, millions of elements is a bit out of my scope...
I'll
look for filtering then.
Yes, definitely filtering is the way to go there :)
Cheers,
Andrei
Regards,
Thierry
Cheers,
Andrei
On Fri, Feb 24, 2017 at 10:37 AM, Thierry Goubier
<[email protected]
<mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>> wrote:
Hi Andrei,
if you're using fasttable for the Raw view, you
should be able to reach one 100k elements
without
issues. I did some experiments and it
handles the
load very well.
Avoid the paginator at any cost. This thing is
really user-unfriendly.
Regards,
Thierry
2017-02-23 20:19 GMT+01:00 Andrei Chis
<[email protected]
<mailto:[email protected]>
<mailto:[email protected]
<mailto:[email protected]>>>:
Hi Stef,
Currently that's the default behaviour
of the
Raw view: it displays for collections
only the
first and the last 21 elements. The
Items view
however always should display all the
elements
of a collection.
The main problem with the Raw view in
Pharo 5 is
the speed. In Pharo 6 now the speed of
the Raw
view is greately improved so we could
increase
those limits. Still for now there should
still
be some limit for the Raw view. Ideally we
should add a small widget, something like a
paginator, for navigating through large
and very
large collections.
Cheers,
Andrei
On Feb 23, 2017 19:35, "stepharong"
<[email protected]
<mailto:[email protected]> <mailto:[email protected]
<mailto:[email protected]>>>
wrote:
Hi
I'm trying to debug citezen
generation and I
have to compare strings.
Now I think that the raw views (in
Pharo 50)
is not good because we cannot see
all the
items in raw format.
See the attachements. It jumps from
21 to
174 ...
and what I want to see is of course
in the
middle.
Is it me or there is something wrong
there.
Stef
