From:	Kevin Grigorenko/Austin/IBM@IBMUS
To:	Memory Analyzer Dev list <mat-dev@xxxxxxxxxxx>,
Date:	04/26/2013 08:24 AM
Subject:	Re: [mat-dev] MAT Argument Processing
Sent by:	mat-dev-bounces@xxxxxxxxxxx

From:	"Buchen, Andreas" <andreas.buchen@xxxxxxx>
To:	Memory Analyzer Dev list <mat-dev@xxxxxxxxxxx>,
Date:	04/26/2013 08:02 AM
Subject:	Re: [mat-dev] MAT Argument Processing
Sent by:	mat-dev-bounces@xxxxxxxxxxx

Hey y'all, as part of bug #404679, I added in an HPROF preference page and support for -D parameters to change the behavior of the HPROF parser. The added UI plugin dependencies are optional but if headless mode is run without SWT/jface plugins, then a ClassNotFoundException will be thrown at runtime. The broader question is how best to process and coalesce preference choices and command line arguments. I propose a new extension point called something like argumentsProvider which will allow a plugin to register a callback which has a method that is called before processing the dump (e.g. ParseHeapDumpJob and ParseSnapshotApp) with a Map<String, String>. Then this map can be added to the SnapshotInfo so that lower code can query for arguments without any direct dependencies. With this design, then we can move the HPROF preferences into a new plugin which has an arguments provider. Any thoughts? Is this overengineering to use the plugin architecture for this?

--
Kevin Grigorenko
IBM WAS SWAT Team
kevin.grigorenko@xxxxxxxxxx
Blog: https://www.ibm.com/developerworks/mydeveloperworks/blogs/kevgrig/?lang=en_______________________________________________
mat-dev mailing list
mat-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/mat-dev_______________________________________________
mat-dev mailing list
mat-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/mat-dev

_______________________________________________
cross-project-issues-dev mailing list
cross-project-issues-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev

From:	"Buchen, Andreas" <andreas.buchen@xxxxxxx>
To:	Memory Analyzer Dev list <mat-dev@xxxxxxxxxxx>,
Date:	04/26/2013 08:02 AM
Subject:	Re: [mat-dev] MAT Argument Processing
Sent by:	mat-dev-bounces@xxxxxxxxxxx

Hey y'all, as part of bug #404679, I added in an HPROF preference page and support for -D parameters to change the behavior of the HPROF parser. The added UI plugin dependencies are optional but if headless mode is run without SWT/jface plugins, then a ClassNotFoundException will be thrown at runtime. The broader question is how best to process and coalesce preference choices and command line arguments. I propose a new extension point called something like argumentsProvider which will allow a plugin to register a callback which has a method that is called before processing the dump (e.g. ParseHeapDumpJob and ParseSnapshotApp) with a Map<String, String>. Then this map can be added to the SnapshotInfo so that lower code can query for arguments without any direct dependencies. With this design, then we can move the HPROF preferences into a new plugin which has an arguments provider. Any thoughts? Is this overengineering to use the plugin architecture for this?

--
Kevin Grigorenko
IBM WAS SWAT Team
kevin.grigorenko@xxxxxxxxxx
Blog: https://www.ibm.com/developerworks/mydeveloperworks/blogs/kevgrig/?lang=en_______________________________________________
mat-dev mailing list
mat-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/mat-dev

Attachment: problemlog.png
Description: Binary data

From:	"Tsvetkov, Krum" <krum.tsvetkov@xxxxxxx>
To:	Memory Analyzer Dev list <mat-dev@xxxxxxxxxxx>,
Date:	03/15/2013 06:15 AM
Subject:	Re: [mat-dev] Proposal: MAT Warning System
Sent by:	mat-dev-bounces@xxxxxxxxxxx

-          Assuming that we introduce the functionality post 1.3, we could go for a major version jump afterwards (i.e. mat 2.0) and signal that there are incompatible changes. It is then clearer, but still this doesnât help any adopters.
-          Add another interface, and make out object implementation implement both â wound need some casting on all the places we would like to use the warnings
-          Use some kind of singleton to pass the object and get the warnings - we will need one anyway (at least internally) to read all extensions.

Hey Andrew, Krum,

Sorry for the delay, I was on vacation. Great feedback. Here are some more comments:

> In the last days I also thought that we should first collect some more concrete examples of what problems we could tackle.

Some other ideas:
- For the WebSphere Application Server product, I'd like to add a warning when a thread has been marked as potentially hung by the hung thread detection - just a boolean check.
- WAS Classloader marks that an application is stopped but it hasn't been garbage collected (somebody is holding on to something that they shouldn't) - just checking a field on the ClassLoader.
- Check if various WAS caches are full - just comparing two integer fields.

> it may help to run a scan once, and save the results somehow, so that opening a view later does not result in new computations.

I worry about running the scan during the initial open since dump open time is already a big complaint about MAT. I also worry about making it some button that you have to click since most users won't click it. I like the approach of IClassSpecificNameResolver, although maybe we should add a SoftReference cache of the resolved warning results?

> For example warning about closed files or disposed graphics objects might make sense if they were being retained when no longer needed, but in another situation excessive numbers of unclosed files could indicate a resource leak.

Yes, I could also think of some situations where a warning only occurs in some situations (for example, one large HTTP session is okay, but a thousand is probably not), but I would rather the default set be speedy. What about a resolver marking itself using an annotation which predicts the impact (e.g. @Impact(ImpactLevel.High))? Then there could be a preference that controls the threshold of resolvers to run.

> A standard Eclipse way is to use an overlay (a decorator) on the icon. For example a warning triangle or red error cross on projects which are in error. Would that be more intuitive?

That is probably better. I can try that first but I'd like to see if it's too subtle or not.

> Should we then have a problems view? You could then show warnings/errors on a selection or across the dump, and group the errors by type, then jump to an instance of the error.

Yes, that makes sense. Can we just reuse the existing Problems View or would we have to create or own?

> We could also allow the user to annotate an object with some text (that would be visible in other views), or possibly tasks (e.g. TODO items for objects the user wants to hand inspect).

Do you mean that a user could right click on an object and add a note or TODO marker?

> We could also have a preferences page - should there be a per snapshot preferences for warnings? How might that be done - should that be like per project Java warnings.

What's the benefit of a per-snapshot preference?

> Should snapshots map to Eclipse projects in the projects explorer?

What's the benefit?

> Should we make it easy to have multiple workspaces for MAT? E.g. all snapshots for one end-user application could be in one workspace, and then by switching workspaces you could then see all the snapshots from another application.

Interesting, I rarely use MAT in the process of development, but instead on post-mortem dumps and I usually don't have the matching source code, so I hadn't thought about this use case.

> Would showing details of the warning in a tab in the object inspector be a good location?

That certainly makes sense to add, but I wouldn't want that to be the only place (marker or color I think is the primary notification).

> Would an extension point to specify the icon for an object be useful?

Could this be one of the fields returned in the WarningResolution struct? Does a custom icon conflict with the IMarker/Problems View approach?

> The IClassSpecificWarningResolver seems a reasonable start, though markers seem a good model and could allow the linking of help.

Do you think IClassSpecificWarningResolver should extend IMarker?

> The name resolver implementation has a slight problem that only one resolver per class is allowed.

I think this is because IObject.getClassSpecificName() returns a single String instead of a List. This is one reason I suggested List<WarningResolution> to support multiple warning resolvers for a single class.

> Adding methods to IObject is not a compatible change

What's the alternative? If I have an IObject and I want to get the list of warnings for it, would I have to use some sort of singleton service?

> Why tree view and not table view as well?

Yes, table view should also show it. Your suggestion of using IQueryPart.IS_IMPORTANT and coloring the row seems to be the way to go.

> How do we save/display the warning text in HTML for saved reports?

Good question. How about footnotes that link to the bottom of the HTML?

> @Subject("java.lang.Object") would probably do the same as "*".

I didn't know @Subject supported "instanceof" resolution. So if I have class B which extends class A, and I write a name resolver for class A but I have an instance of class B, then the name resolver of class A will run?

--
Kevin Grigorenko
IBM WAS SWAT Team
kevin.grigorenko@xxxxxxxxxx
Blog: https://www.ibm.com/developerworks/mydeveloperworks/blogs/kevgrig/?lang=en


"Tsvetkov, Krum" ---03/01/2013 04:34:06 AM---I also think it would be a nice improvement to MAT to add some more automatically shown warnings. >>

From:	"Tsvetkov, Krum" <krum.tsvetkov@xxxxxxx>
To:	Memory Analyzer Dev list <mat-dev@xxxxxxxxxxx>,
Date:	03/01/2013 04:34 AM
Subject:	Re: [mat-dev] Proposal: MAT Warning System
Sent by:	mat-dev-bounces@xxxxxxxxxxx

-          Big objects (> % of the total heap)
-          Closed files
-          Un-/disposed widgets https://bugs.eclipse.org/bugs/show_bug.cgi?id=274369
-          Small Strings objects, keeping a huge char[] (as result of substring() call)
-          A big, empty collection ?? - not sure if this is useful on a single object. The reports showing millions of empty collections are good, but on a single object I donât know.