585482 - Firefox fsyncs excessively during page load of mapquest.com

Status: RESOLVED WORKSFORME

(Toolkit :: Places, defect)

Description

[Jeff Muizelaar [:jrmuizel]]

When loading mapquest.com we seem to fsync about 97-122 times. Chrome fsyncs about 13-15 times and these happen well after the page has finished loading.

Comment 1

[Shawn Wilsher :sdwilsh]

I presume that because you are filing this in places you have stacks of these fsyncs showing places at fault?  These would be helpful to see.

Comment 2

[Matthew Gregan [:kinetik]]

Attachment: linux fsync/fdatasync stacks

I see 42 fdatasyncs on Linux loading that site.  Stacks attached.

Comment 3

[Shawn Wilsher :sdwilsh]

The good news is that every single one of these is happening off of the main thread.

This is likely a regression from bug 556400.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

(In reply to comment #1)
> I presume that because you are filing this in places you have stacks of these
> fsyncs showing places at fault?  These would be helpful to see.

From what I recall, most if not all were from mozilla::storage::AsyncExecuteStatements::Run which makes it difficult to assign blame.

Perhaps it's worth adding a thread safe singleton object that is passed as a parameter to the IO submission that AsyncExecuteStatements::Run can call through so that blame for IO can be assigned back to code that caused it.

Comment 5

[Shawn Wilsher :sdwilsh]

(In reply to comment #4)
> Perhaps it's worth adding a thread safe singleton object that is passed as a
> parameter to the IO submission that AsyncExecuteStatements::Run can call
> through so that blame for IO can be assigned back to code that caused it.
Well, it's usually pretty easy to track what file is getting the fsync if you also look at open and track the file descriptor that's handed back.  With dtrace, this is pretty easy (and I've got a script for that somewhere).

Like I said though, I suspect this is places, and I suspect it's bug 556400 that caused it.

Comment 6

[Jeff Muizelaar [:jrmuizel]]

(In reply to comment #5)

It would be handy for profilers like shark though

Comment 7

[Marco Bonardo [:mak] (Away Apr 25 - May 5)]

this doesn't happen anymore from when we moved to WAL journal