Performance issues with version 5.0 (example videos)

[AEgit]

JabRef 3.8.2
windows 10 10.0 amd64
Java 1.8.0_211

VS

JabRef 5.0-dev--snapshot--2019-06-20--master--6265e2460
Windows 10 10.0 amd64
Java 1.8.0_211

As announced previously (#4430 (comment)), here I add some videos that showcase the performance difference between JabRef 3.8.2 vs. JabRef 5.0 when working with a large database (> 18,000 entries) and several thousand static groups.

These issues have been reported previously, but I thought it would be better to create a new bug report instead of adding the videos to just one of these previous reports (they should apply to all of them, I reckon). Previous reports of issues with the performance include (among others), the following:

#4430
#4756
#4526

Note, that I had to run these tests on a virtual machine, since I do not have a full license for YourKit Java Profiler and had to use a test license (since the test license I had on my actual machine is no longer valid). So the performance on an actual machine should be slightly better (the virtual machine is running on a SSD, and I assigned two cores of a Core i7 Hexacore and 8 GB of 32 GB of RAM to the virtual machine), but the relative performance difference between version 3.8.2 and version 5.0 should be comparable. In fact, my impression is that this setup favours version 5.0 because 3.8.2 appears slower in the virtual machine than in a non-virtualized setting. Version 5.0 on the other hand, is equally slow in a virtual machine and when running on the actual hardware.

Note, that I use the same database when comparing the two versions. The only difference is, that the database in version 5.0 has been saved using version 5.0 (while the other one is still using a database in 3.8.2 format). This allows for better comparison, since version 5.0 stores groups differently.

1. Comparison of first start of JabRef:
   3.8.2:
   https://app.box.com/s/7agrhreojpk85itl9dvsz7mx8ueaydcl
   5.0:
   https://app.box.com/s/79dqowjer214qdi0wiceh03x9g4iuexi

Note, that version 5.0 sometimes seems to have crashed. Furthermore, I did not immediately click onto 3.8.2 when it was actually already ready.

2. Comparison of JabRef search feature:
   3.8.2:
   https://app.box.com/s/fgr84sgpsz6dscbnku91m5o2hhbf6rwt
   5.0:
   https://app.box.com/s/fxwcska8dm5rlilv6a9mqswelkrqkh0t

3. Comparison of working with JabRef groups:
   3.8.2:
   https://app.box.com/s/fc5ku29f39o9r7h86hj309tao4uifo8o
   5.0:
   https://app.box.com/s/cubdi2vvmzjnj5i7rtn57lfxp24kvzb9

Note, how JabRef freezes several times in version 5.0 when using certain groups features (e.g., assigning an item to a group, assigning a group as a new subgroup, ....). Sometimes it appears as if nothing is happening in the video, but it is actually just JabRef freezing from time to time.
Note, the big differences in the time that JabRef spents on some actions (see YourKit Java Profiler). For certain actions, JabRef 5.0 will take much much longer than 3.8.2 (which translates into higher CPU demands).

[buhtz]

#5156

[tobiasdiez]

@AEgit the most recent developement version now includes the latest performance improvements. Could you please check what still needs improvement. Thanks!

[AEgit]

@tobiasdiez Due to the issues with the current Windows installer (see #5580 (comment)) and since the snap package on Linux also does not seem to work yet (see #5417) I cannot test the current developer version, I am afraid.

[LyzardKing]

The snap can be installed from https://builds.jabref.org/master/
It won't auto update and you need to install with --dangerous (since it's not from the store)
But for testing it can be done.

[AEgit]

JabRef 5.0.0-dev--2019-11-17----151a216a2
Windows 10 10.0 amd64
Java 13.0.1

@tobiasdiez Tried the current dev version. This one is much improved in terms of performance. Switching between different groups and articles is now much smoother. Well done!

The following performance issues remain:

1. Assigning an item to a group has a heavy impact on the CPU. Now, if you assign just a single item, this appears to be no major isse for a user with a powerful machine. I have a 6-core i7 and while the assignment of a single item brings the CPU usage to nearly 50% for a few seconds, the item appears (?) to be assigned immediately to this group (this might already be a problem, though, for people with weaker machines).
   What is a problem, however, is when you try to assign multiple items simultaneously to a new group. The CPU usage shoots up and (depending on how many items you try to assign to the group) JabRef freezes for some time. The CPU usage also stays high for a protracted time period: I am not sure, whether the items have already been assigned to the group as indicated by the item counter since the CPU usage stays high even after that for quite some time.
   To recreate this problem, try the following:
   a) Select multiple items in the main table
   b) Drag'n'drop them onto a group they had not be assigned to
   c) Observe the mentioned issues
2. The same problem arises when you try to assign groups as subgroups of other groups (in fact, I think this feature might be bugged, but I have to test this more thoroughly). If you assign just a single group, things appear fine. But if you assign multiple groups simultaneously as subgroups to a new group (again, by using the drag'n'drop feature - but this time in the groups panel), you end up with a high CPU usage over a protracted period of time.
3. When you select multiple groups in the groups panel simultaneously (using left click on the starting group and left click + SHIFT on the ending group), there is a noticeable delay until all groups are selected and the CPU usage shoots up again.
4. There is a noticeable delay (a few seconds) when you select the "All entries" group until the respective items are shown in the main table. I reckon this is directly dependent on the number of entries assigned to a certain group, because this problem also appears to arise, when groups with a high entry count are selected (currently my "All entries" group contains nearly 20,000 items).

These are the performance issues I have observed so far. Nevertheless, again a big thanks to @tobiasdiez for the major performance improvements that the current version offers now - this is definitely a major step into the right direction.

[tobiasdiez]

Thanks for the feedback. I'll try to have a look at the remaining performance issues in the next days.

[AEgit]

JabRef 5.0.0-dev--2019-11-24----8780a5632
Windows 10 10.0 amd64
Java 13.0.1

Note also, that it appears, as if the search has become slower again. Not as bad as it was in the past (#2852), but it seems, that the latest versions have seen a performance decrease there as well. Entering a search term using a large database results in lagged input (similar to a low frame rate in a video game).

[Siedlerchr]

@tobiasdiez related to the add tab threading?

[tobiasdiez]

@AEgit We implemented some performance improvements recently. Can you please see if they might have fixed some of the issues you described in #5071 (comment). Moreover, it would be good which of the remaining issues are critical, i.e. need to be fixed before we release 5.0. Thanks

[AEgit]

JabRef 5.0-beta.354--2020-01-18--2612e22
Windows 10 10.0 amd64
Java 13.0.2

The new versions brings some improvements, but the problems are still too big for a release version. I will explain this in detail below (based on the issues reported in #5071 (comment)):

1. Now, if you assign multiple items to a single group, JabRef still has a high CPU usage (60%) over a protracted period of time (several minutes!), but it does not freeze anymore. However, now another issue can be observed. Namely, the cycling through entry previews does not work while the CPU usage is high. When you try to go from an entry preview of one item to the next one, the next item will be selected in the main table (indicated by the blue marking), but the corresponding entry preview is not shown. Instead, the entry preview of the previously selected item is shown. This means, that for several minutes (until the CPU usage has dropped from the item assignment to a new group) it is NOT possible to see the entry previews of all entries (except for the one, that had been selected initially).
   I think this is a major problem, that needs to be fixed before release.
2. The problem with high CPU usage over time, if you assign a group as a subgroup of another still persists and appears to be even worse than before. I have just tried to assign one group as subgroup of another and that assigned another group as the subgroup of the previous one, i. e. we end up with a 3-leveled order: group - subgroup - sub-subgroup
   In total, just two groups have been moved. Despite that, the CPU usage stayed high for several minutes (!). It might be, though, that this problem only appears, if issue 1 has been triggered first, since I have not been able to trigger this behaviour consistently.
3. Issue 3 appears to be fixed.
4. The delay mentioned in issue 4 has been much reduced (it is now less than a second, as far as I can tell).
5. When you now make changes to group structure (e.g., assign two groups as the subgroup of another group) and then selected another group again, NO items at all are shown in the main table. This means, that a few changed to the group structure lead to a non-functional main table (it just remains empty).
6. The slowed down search persists (see Performance issues with version 5.0 (example videos) #5071 (comment)).

I wonder whether some of these performance issues could be solved by a simple workaround: Make it possible to switch off the calculation of the number of items in the groups. In version 3.8.2 this was possible and was also suggested to do, when having performance problems. Indeed, I switched it off for my database, since otherwise it would not be possible to work with the database due to performance problems.
However, you shouldn't get rid of the background box surrounding the group item account. This box changes colour according to whether an item selected in the main table is part of a group or not, so this information should be available. In the old JabRef version this was achieved by having the name of a group underlined, which contained a selected item.

[ilippert]

I have 10500ish entries; simply sorting the entry table by a column, e.g. title, takes typically several seconds.

JabRef 5.0--2020-03-08--93de138
Linux 5.5.7-200.fc31.x86_64 amd64
Java 13.0.2

[m-mauersberger]

Performance issues persist for me when using a large database (~5000 entries) as well: database operations take ~5-10s. I want to use it as shared DB but the problem is the same for local and shared databases (see also #4461). Even if no changes are present, reloading of (shared) database takes the same time!

Maybe it has something to do with the EventBus? Just as a quick idea: Can it be the reason for slow down that each entry has its own EventBus and the listening methods are "overcharged"?

[ilippert]

Recent versions of JabRef have worked well for me.
but
JabRef 5.3--2021-01-07--129abaa
Linux 5.9.16-200.fc33.x86_64 amd64
Java 15.0.1

seems very slowly reacting (takes several seconds to switch between entries in entry table). Although the use of the system's resources is fine.

[ilippert]

Recent versions of JabRef have worked well for me.
but
JabRef 5.3--2021-01-07--129abaa
Linux 5.9.16-200.fc33.x86_64 amd64
Java 15.0.1

seems very slowly reacting (takes several seconds to switch between entries in entry table). Although the use of the system's resources is fine.

Ok, today under
JabRef 5.3--2021-01-07--aef49f4
Linux 5.9.16-200.fc33.x86_64 amd64
Java 15.0.1
JavaFX 15.0.1+1
it works well again ;)

[ilippert]

JabRef 5.3--2021-01-24--ae43548
Linux 5.10.9-201.fc33.x86_64 amd64
Java 15.0.2
JavaFX 15.0.1+1

When starting up, JabRef needs 60-70% of CPU. It then goes down to 0% If I don't do anything. When selecting an entry it goes up to 2-3%. Switching a library goes up to 7%. Well, this might be normal operations. Adding a few lines in some entries. And then CPU use goes up to 60, 70, >80%! Closing JabRef then does not stop the process.

[tobiasdiez]

Thanks for the report. Do you remember the last version were you didn't experienced these problem?

[ilippert]

Thanks for the report. Do you remember the last version were you didn't experienced these problem?

Unfortunately, not exactly. If it helps, I would say last week. (If there was a repository with all the last 2 week's master builds, I would be happy to test these; but I don't think this exists.)

[ilippert]

JabRef 5.3--2021-03-15--b068ce7
Linux 5.10.22-200.fc33.x86_64 amd64
Java 15.0.2
JavaFX 16+8

have not done anything with JabRef for a few minutes, and it uses 50% of CPU.
happens regularly that JabRef uses up to 100% and then crashes.

[ilippert]

JabRef 5.3--2021-03-19--049acb9
Linux 5.11.7-200.fc33.x86_64 amd64
Java 15.0.2
JavaFX 16+8

I am sorry to say that JabRef over the last versions reliably crashes and often crashes the system, requiring a cold restart of the computer.

[Ali96kz]

@ilippert could you provide your .bib file? I will try to investigate it

[Ali96kz]

Could we test it with that fix #7486

[ilippert]

JabRef 5.3--2021-03-29--2948e6d
Linux 5.11.10-200.fc33.x86_64 amd64
Java 15.0.2
JavaFX 16+8

Over the last two days I experienced no issues.

[Ali96kz]

@AEgit Could you test it with that fix #7486, please don't forget to make backups

[AEgit]

JabRef 5.3--2021-03-29--2948e6d
Windows 10 10.0 amd64
Java 14.0.2
JavaFX 16+8

The issue mentioned in #5071 (comment) persists.

[ilippert]

JabRef 5.7--2022-06-12--e3b2ab2
Linux 5.17.13-300.fc36.x86_64 amd64
Java 18.0.1
JavaFX unknown

I am quite happy with the performance, however,

• when typing into fields in the entry editor, the process is often far too slow in actually accepting the input and showing it in the field.

[ThiloteE]

meta-issue: #8906

[ThiloteE]

JabRef 5.7--2022-06-15--cb5fe60
Linux 5.4.0-120-generic amd64
Java 18.0.1
JavaFX unknown

Tried to reproduce #5071 (comment)

Took me ca. 1-4 minutes to even select 7360 entries from my 100 000 entries database (15,5 mb) with Shift+PageUp. This was a lot faster than Shift+Up by the way. I don't want to know how long it will take to select 7000 entries with CTRL+Shift+Up....

Then i tried to delete 7360 entries: Used my stop watch to count: 1 hour, 37 minutes ... + X. Did a force close, because I don't want this PC to run at max capacity for hours. Just to delete a simple text in a text-file this seems a little bit over the top.

CPU was pretty high. RAM was normal and stable.

This was done with an old laptop from 2013: Aspire E13 (ES1-311-P87D), the HDD was replaced with a SSD at one point.
The CPU is a Intel Pentium N3540, which is designed for long battery life and consequently has weak benchmarking results: https://www.cpubenchmark.net/cpu.php?cpu=Intel+Pentium+N3540+%40+2.16GHz&id=2408

[ThiloteE]

Instead of reporting your findings with regard to performance here directly, in future I would ask you to create separate issues for issues that have previously been mentioned in this thread, but still have not been solved and then link to this issue here.

With this workflow, we can more easily ignore issues that have already been solved and it saves us all the time to re-read walls of text over and over again :)) Thaaanks in advance :)

[koppor]

IMHO, the main issue that our main table is tightly coupled with the data of the BibDatabase: We load all data into memory and work in the in memory representation with Java and JavaFX data structures.

"Fix" in the medium way:

• Update to Lucene for the search (that way, the generation of search data is decoubled of the search itself) (happening at Lucene search backend #8963)
• Investigate our JavaFX observable collections and event bus - and see if "too much" things are going on when using JabRef
• Really personal thinking: We should update the internal data structures to Eclipse Collections, because they are more efficient.

"Good architecture" fix:

Use an embedded database to provide pre-calculated views, indexing, search, ...