7251 - improve startup performance tracking bug

Status: RESOLVED DUPLICATE of bug 447581

(Core Graveyard :: Tracking, defect, P3)

Description

[chris hofmann]

all the areas that need to be fixed to make us start fast...

Comment 1

[Chris Yeh]

removing dependency on 7249, since we do this for the optimized builds already.
the bug is still open to get it working for developer builds.

Comment 2

[leger]

Bulk add of "perf" to new keyword field.  This will replace the [PERF] we were
using in the Status Summary field.

Comment 3

[Chris Waterson]

*** Bug 26224 has been marked as a duplicate of this bug. ***

Comment 4

[chriss]

Adding beta1 to keyword list

Here's the latest test data for start time after a system reboot.

                            4.71       SM
Win98200MHz, Pentium Pro    2:02s      11:23s
64MB RAM

Mac OS 8.5                  2:89s      13:63s
266MHz, PowerPC G3,
64MB RAM 

Linux RedHat 6.0,           5:11s      27:34s
166MHz, Pentium, 80MBRAM

For beta 1 we need to be within 2x 4.71

Comment 5

[chris hofmann]

-> jevering 

Comment 6

[leger]

Putting on PDT+ radar for beta1.

Comment 7

[Jim Roskind]

Removed PDT+  to get a re-eval so that this doesn't show in stats.  I'd rather 
see the actual bugs  rated as PDT+, then see this one in addition to all the 
others (at present, there is one other :-/ ).

Comment 8

[leger]

Adding 28964

Comment 9

[leger]

Adding 28911

Comment 10

[leger]

Adding 26512

Comment 11

[Suresh Duddi (gone)]

Taking over this bug.

Comment 12

[leger]

Updating QA Contact.

Comment 13

[kirun]

*** Bug 50746 has been marked as a duplicate of this bug. ***

Comment 14

[Edward Kandrot]

Using Quantify and the latest build, it appears that we spend 40% of our boot
time in malloc, allocating 180,000+ blocks.

Comment 15

[Dawn Endico]

dp no longer at netscape. assigning to chofmann so this is on someone's radar

Comment 16

[chris hofmann]

maybe selmers that better owner/watcher for this tracking bug 
from an application point of view.

kandrot, do you know if there is a bug that specifically talks
about all this memory allocation going on at startup?
That is the part memory and performance team might be
interested in.  also, do we know if the memory allocation
that is going on rougly mirrors the code size of the .dll's
or is there any sore thumbs that might stand out as allocating
more than they need to or should at startup?

sounds like evaughn's late loading ideas might have a big impact
on the overall startup time if we can shave off lots of code
that really is not used a large pct. of the time.

waterson, did you see this.  holy heck.  40% of startup
time just to cram the monkey into memory...  whoa!

Comment 17

[chris hofmann]

another good piece of data to get would be to figure out of
the 40% of seamonkey start time in malloc ratio also holds true
in gtkembed and winembed.

Comment 18

[selmer (gone)]

->McAfee, CC Vishy

Comment 19

[Jesse Ruderman]

Adding bug 67618 (mallocs, ~40%) and bug 64146 (making charset menu takes ~5%).

Comment 20

[Cathleen]

added more startup bugs

Comment 21

[Chris McAfee]

over to cathleen

Comment 22

[Shiva Thirumazhusai]

Adding myself.

Comment 23

[Brendan Eich [:brendan]]

Adding bug 68045 ("precompile chrome JS and load it incrementally"), which is
really "precompile all XUL/XBL/JS/CSS that costs a lot more to parse than to
fast-load from a pre-compiled form".

/be

Comment 24

[Randell Jesup [:jesup] (needinfo me)]

ARGH!  Someone who got CC'd on the last update to this please re-add all the
dependencies!  Mozilla (or bugzilla?) apparently emptied the dependency field,
and  when I added one it erased all the old ones.  Please add 91231 when you fix
this.  Sorry; I should have been more paranoid when I noticed an empty text
field there.

Comment 25

[Randell Jesup [:jesup] (needinfo me)]

Hmmmm.  I just noted that a bunch of the other fields got wierded out as well.
(Summary should be "improve startup performance tracking bug").   Sounds like a
Bugzilla bug....

Comment 26

[Jacek Piskozub]

Restoring the fields as they were. Adding 91231 as requested.

Comment 27

[selmer (gone)]

sorry for the spam, the field changes fooled me at a sleepy moment...

Comment 28

[Randell Jesup [:jesup] (needinfo me)]

Now be good this time, bugzilla...

Comment 29

[Myk Melez [:myk] [@mykmelez]]

Mysteriously disappearing values in fields are usually bugs in Mozilla's form
handling.  Randell, were you using Mozilla to update the bug, and if so what
build?

Next time you notice this you should see if it happens with Netscape 4.x or
another browser.  If so, file a bug against Bugzilla; otherwise file one against
Browser->HTML Form Controls or Browser->Form Submission (depending on whether
you notice the problem before or after the form gets submitted) and notify
[e-mail] immediately so we can block that Mozilla build to prevent
corruption of the Bugzilla database.

Comment 30

[Matthew Tuck [:CodeMachine]]

Usually I would blame Mozilla for this sort of thing ... BUT ... I noticed that
when the dependencies were readded the notifications got sent out merged with
the original notifications when the dep got removed.

AIUI when this happens its a Bugzilla bug, and we're not sure why it happens. 
If that's the case, it's likely that the whole thing might be a Bugzilla bug.

Comment 31

[Randell Jesup [:jesup] (needinfo me)]

I was using mozilla; a daily from just before 0.9.3.  Note however that it
replaced the subject line with the subject line from the first bug in my query
list - it would be tough for mozilla to do that, since I never even looked at
that bug (other then it being an entry in the query results).  I still think
it's bugzilla.

Comment 32

[Chris Waterson]

Attachment: jprof output of static build ``cold'' start on 200MHz machine (21sec, with JP_PERIOD=0.0015)

Comment 33

[Chris Waterson]

Some analysis on the above.

I'm going to treat main() as the root for the relevant number of stack
traces (that's 1858, as opposed to 1911). Since this is a sampling
profiler, it should yield good results with respect to time spent in
I/O. Unfortunately, it doesn't appear to have picked up symbols from
NSPR (even though I built with seawood's patch for bug 88045).

- Only 2% of the stack traces (37/1911) fell in NS_InitXPCOM(). Of
  that, 16 (or less than 1%) were spent in
  nsComponentManagerImpl::PlatformPrePopulateRegistry(). So, we seem
  to spend very little time reading in the registry. About 1/2% of the
  time was spent reading the XPTI files. So it appears that there is
  little to gain from optimizing XPCOM initialization code.

- About 8% of the time was spent in
  nsAppShellService::CreateHiddenWindow() (158 stacks). From here, 134
  stacks (or 7% overall) was spent in
  nsAppShellService::GetHiddenWindowAndJSContext(). This winds up in
  what appears to be a CreateInstance() call. (See below.)

- 329 stacks, or 17% of the time, was spent in CreateInstance(). The
  biggest offenders are:

  - nsNSSComponentConstructor(), which accounts for 5.5% (103 stacks).
  - nsChromeRegistryConstructor(), which accounts for 2% (39 stacks).
  - nsAutoConfigConstructor(), which accounts for 1.6% (30 stacks).
  - nsPrefServiceConstructor(), which accounts for 1.6% (29 stacks).
  - NS_NewCharsetMenu(), which accounts for 1% (19 stacks).
  - Construct_nsIScriptSecurityManager(), which accounts for 0.5% (10
    stacks).

  Others to note (that take less than 0.5%, but still show up with
  several stacks) are:

  - nsHTTPHandler::Create()
  - nsProfileConstructor()
  - CreateXULDocument()
  - NS_NewLocalStore()
  - InternetSearchDataSourceConstructor()
  - nsBookmarksServiceConstructor()
  - nsCookieHTTPNotifyConstructor()

  All of which read files from the disk.

- We spend 1.7% of the time in nsStaticComponentLoader::GetFactory(),
  which seems high. Perhaps we're paging in the executable to access
  data at this point?

- We spend 18% of the time (330 stacks) in poll(), from
  g_main_poll(). What are we doing?

- We spend 5% of the time (241 stacks) parsing style sheets, that is,
  in CSSParserImpl::Parser().

- We spend 8% of the time (149 stacks) handling ContentInserted()
  notifications from asynchronous XBL loads. It appears that most of
  the time is spent constructing frames, so it's not clear that
  synchronously loading the XBL (i.e., so it would have been ``there
  already'') would buy much.

- It appears that we spend

  - 10.8% of the time (201 stacks) in
    nsCSSFrameConstructor::ConstructFrame(),

  - 10.7% of the time (199 stacks) in
    nsCSSFrameConstructor::ConstructFrameInternal(),

  - 4.9% of the time (90 stacks) in
    nsCSSFrameConstructor::ConstructXULFrame(), and

  - 5.9% of the time (109 stacks) in
    nsCSSFrameConstructor::ProcessChildren().

  That is, when the sample was taken, the stack pointer was somewhere
  in these routines (not somewhere in a descendant of the
  routine). Take this evaluation with a grain of salt: since these
  routines recursively call each other, I'm not completely positive
  that I've properly understood jprof's accouting.

More later...

Comment 34

[Randell Jesup [:jesup] (needinfo me)]

Chris Waterson writes:
>- We spend 1.7% of the time in nsStaticComponentLoader::GetFactory(),
>  which seems high. Perhaps we're paging in the executable to access
>  data at this point?

        Probably.  dlopen() (depending on the OS) doesn't really load
the entire library; quite a bit will get paged in later.

>- We spend 18% of the time (330 stacks) in poll(), from
>  g_main_poll(). What are we doing?

        Those will be mostly waiting on IO I'd guess; you'll have to look
at the exact stacks.  The question what IO are we waiting on...???  I don't
think we're doing async file IO so that wouldn't show up, and I don't think
we should be doing network IO or DNS (but perhaps we are).  Perhaps some
things/Events are handled or generated off of timers as well, and we're
waiting for those?  This would be good to spend some time understanding.
Perhaps it's a gtk/etc thing?  What was used to end the profile, or shut
mozilla down?  If you did it by hand, that might account for it.  Also,
it may have been waiting for X - I don't know how gtk/gdk handles X;
it may dispatch things async to X and then go back to the main event
loop to poll.(!)

        Also, we have 41 hits (2.2%) in select() (i.e. sleeping waiting
for some form of IO or timer).

>- 329 stacks, or 17% of the time, was spent in CreateInstance(). The
>  biggest offenders are:
>
>  - nsNSSComponentConstructor(), which accounts for 5.5% (103 stacks).
>  - nsChromeRegistryConstructor(), which accounts for 2% (39 stacks).
>  - nsAutoConfigConstructor(), which accounts for 1.6% (30 stacks).
>  - nsPrefServiceConstructor(), which accounts for 1.6% (29 stacks).
>  - NS_NewCharsetMenu(), which accounts for 1% (19 stacks).

        NSSComponentConstructor looks suspicious, especially since we
haven't hit any https links yet.  It spends 1/2 it's time in
NewTempCertificate, a lot of in ASN1 decoding.  Probably parsing key/etc
files?  The others may each deserve a short look for easy gains.  The
CharSetMenu is a good target for lazy creation.

>- We spend 5% of the time (241 stacks) parsing style sheets, that is,
>  in CSSParserImpl::Parser().

        This may be worth looking at a little.

>- We spend 8% of the time (149 stacks) handling ContentInserted()
>  notifications from asynchronous XBL loads. It appears that most of
>  the time is spent constructing frames, so it's not clear that
>  synchronously loading the XBL (i.e., so it would have been ``there
>  already'') would buy much.

        Agreed.

>- It appears that we spend
>
>  - 10.8% of the time (201 stacks) in
>    nsCSSFrameConstructor::ConstructFrame(),
>
>  - 10.7% of the time (199 stacks) in
>    nsCSSFrameConstructor::ConstructFrameInternal(),
>
>  - 4.9% of the time (90 stacks) in
>    nsCSSFrameConstructor::ConstructXULFrame(), and
>
>  - 5.9% of the time (109 stacks) in
>    nsCSSFrameConstructor::ProcessChildren().
>
>  That is, when the sample was taken, the stack pointer was somewhere
>  in these routines (not somewhere in a descendant of the
>  routine). Take this evaluation with a grain of salt: since these
>  routines recursively call each other, I'm not completely positive
>  that I've properly understood jprof's accouting.

        Yes, those all wind in an out of each other a lot, then they
devolve into a load of smaller hits all over.  Untangling these (maybe
looking at the raw stack frames) may help.  Also, profiling some pageloads
this way may show better where the FrameConstructers are spending time.

Other notes:

    - 6.2% is spent in read().

    - 4.8% is spent in nsContainerFrame::ReflowChild

    - 3.6% is spent in free() - and fair bit of that in locking/unlocking
           the pthread mutexes.
    - 3.6% is spent in malloc() - about 2/3's in malloc, 1/3 in pthreads.

    - 3.2% is spent in ~nsCOMPtr_base; about 1/4 in that itself(!?) and the
           rest in a bazillion ::Release methods.  (I was a little
           surprised by the number of hits in this directly.  The hits in
           ::Release are destructors and free().

    The nsObjectHashtable destroy/reset code is hit a lot, but it all
    devolves to ::Release calls (i.e. destructors & free).

    nsXULPrototypeDocument::~nsXULPrototypeDocument(void) seems to get more
    hits than I'd expect.

    - 2.4% is spent in js_GetToken, 1% of it direct hits.

    - 2.3% is spent in js_GC, and separately about 1% in gc_root_marker

    - 2.3% is spent in doContent (from XML_Parse).

    - 1.3% is direct hits in kill (?)

    - 1.1% is spent in nsCSSScanner::Next

    - 1.1% is spent in nsHashtable::Get (from all over)

    - 1% is spent in libc_write

    - 1% is spent inflating zip files

    - 1% is spent parsing StdURL's; probably much of that allocing CStrings

    we're spending a couple of % in the dlopen and dl_runtime code.

    - 0.5% doing stat's (from a bunch of places)

    - 0.5% direct hits in nsCOMPtr_base::begin_assignment (seems odd)

    - 0.5% direct hits in nsStr::StrTruncate (seems odd)

    - 0.5% is spent in JS_DHashTableOperate (mostly direct hits in
           SearchTable)

    - 0.2% direct hits in nsVoidArray::InsertElementAt (I'll look at this)
           plus 0.3% in GrowArrayBy (really in Realloc)

    - 0.2% direct hits in nsVoidArray::ElementAt (I'll look at this but I
           think it's optimal)

I noted there's ShutdownXPCOM called - I assume then this is a
startup/shutdown pair?  Perhaps startup, then signal (or JS?) to stop
the logging would be better, though you'll pick a bunch of poll() hits.

Some hotspots/solutions aren't obvious from a profile, especially when a
lot of it is interpreting XUL/JS/etc.  For example, lazy creation of things
not immediately needed (especially if the lazy creation can be dropped onto
a low-pri background thread; perhaps not possible in this case - or find
some way to evaluate it on the main thread when it's otherwise idle, to
avoid adding locking all over.).

Comment 35

[Chase Tingley]

> Those will be mostly waiting on IO I'd guess; you'll have to look
> at the exact stacks.  The question what IO are we waiting on...???  I don't
> think we're doing async file IO so that wouldn't show up, and I don't think
> we should be doing network IO or DNS (but perhaps we are).

FWIW, if a PAC file is specified, it will be loaded during startup (prefs
initialization) rather than waiting for a network event. (bug 83984)

Comment 36

[Randell Jesup [:jesup] (needinfo me)]

I ran a jprof (JP_REALTIME JP_START JP_PERIOD=0.0003) on a dual 450 Linux
box with plenty of RAM.  This was a warm start, done seconds after having
started and stopped it.  I stoped the profile by hand after it had come up
(with www.mozilla.org as the home, and tinderbox as the sidebar).

Build: from about a week ago, with my ContentIterator fixes.  Compiled with
-O2 optimization.

Summary:

Note: %'s are always % of the total number of non-poll hits (157)!
'direct' is hits directly in the routine.

157 non-poll/etc hits.
   59 (38%) in ResumeParse()
      art of this (<10%) is handling initial reflows.
      Most appears to be in doing *::OpenContainer calls.

   32 (20%) is doing CSSLoaderImpl::ParseSheet()
      Most of that is inserting content, eventually calling
      nsCSSFrameConstructor::ContentInserted.
     7 (4.5%) are in CSSParserImpl::Parse()

   36 (23%) are in CreateInstance() (pre dp's patch)
      22 (14%) are in nsDll::Load, probably all going to dlopen.  Almost
         all of those are in dl_lookup_symbol (21, 13.5%)
      15 (9.6%) in nsGenericFactory::CreateInstance calling Constructors
         and ::Create() methods  

   30 (19%) are in nsCSSFrameConstructor::ConstructFrame (note: see ParseSheet
      above, this overlaps with that.)

   25 (16%) in js_Interpret
      A fair bit of this is XBL.
      6 (3.8%) in js_GetToken, 3 of them direct hits(!)
      5 (3.2%) in js_ForceGC
      5 (3.2%) in js_AtomizeString (1 direct)

   17 (11%) in nsXBLService::LoadBindings
      10 (6.4%) in ::InstallProperties, a lot of it executing JS

   13 (8.3%) in nsDocShell::CreateContentViewer

   13 (8.3%) in nsContainerFrame::Reflow (some overlap with above/below entries)
      8 (5.1%) in nsBlockFrame::Reflow

   13 (8.3%) in XPCWrappedNative::CallMethod, including 3 direct hits(!?)

   13 (8.3%) in malloc()/new

   12 (7.6%) in free(), including 2 direct hits.

   11 (7.0%) in nsBox::Layout, called from Reflow() (see containerframe above)

   11 (7.0%) in XML_Parse

   10 (6.4%) in InitializeProfileService.  Most of that is in:
      8 (5.1%) nsChromeRegistry::FlushCaches calling GetServiceByContractID
          (this may be mostly in opening shared libs)

   10 (6.4%) in nsPipe::nsPipeInputStream, almost all from:
      9 (5.7%) in HaveDecodedRow, drawing images

    7 (4.5%) destroying webshells/docshells and their frames

    3 direct in memset (from 3 places)
    2 direct in nsString::nsString(void)
    2 direct in nsSupportsArray::QueryInterface
    2 direct in memcpy
    2 direct in nsString::Length
    2 direct in JS_GetPrivate()

I can upload the jprof or email it if someone likes

Comment 37

[Randell Jesup [:jesup] (needinfo me)]

js_GetProperty seems to be a hotspot in the JS code.  (6.4%) and it doesn't fan
out too far.

Comment 38

[Randell Jesup [:jesup] (needinfo me)]

Attachment: Warm boot jprof, dual-450 PIII Linux, -O2, realtime, to home page (mozilla.org), tinderbox panel open

Comment 39

[Randell Jesup [:jesup] (needinfo me)]

Sorry, I meant _js_LookupProperty for the hotspot (JS_GetProperty is one of it's
callers).  The jprof will show you.

Comment 40

[Cathleen]

trace-malloc results for mozilla startup, with blank page
Type                    Count    Bytes %Total
TOTAL                  115396  5863421 100.00
void*                   23517   943767  16.10
registry-Buffer             1   512000   8.73
xpti-unclassified         250   274080   4.67
unclassified-string      5623   203441   3.47
JSScopeProperty          4460   160560   2.74
gtk-unclassified          928   159473   2.72
nsZipArchive             3438   147662   2.52
JS-Array                 4654   144448   2.46
js_MatchToken            5046   135536   2.31
FrameArena                 31   127565   2.18
JS-script                 866   125423   2.14
nsXULElement             2086   118296   2.02
JS-atom                  4085   114218   1.95
CSSDeclarationImpl       2732   109940   1.88
nsCSSRule                1342   106488   1.82
JSScope                  1140    95760   1.63
nsComponentManagerImpl   2312    90464   1.54
nsTextFragment            360    90106   1.54
JS-unclassified          2110    87392   1.49
nsStaticCaseInsensitiveNameTable   2574    85184   1.45
orkin-unclassified         50    82694   1.41
X-unclassified           3187    82483   1.41
nsCStringKey             2525    77323   1.32
JS-GC-arena                 8    73912   1.26
JS-function              2412    71840   1.23
nsXULAttribute              5    71680   1.22
AtomImpl                 1749    65758   1.12
nsXULAttributes           865    62280   1.06
xptiInterfaceInfo        1409    62232   1.06
nsXMLElement             1137    59124   1.01
RuleHash                 3782    58652   1.00
InMemoryDataSource         92    56074   0.96
nsPersistentProperties   1923    55962   0.95
sscanf                   1504    55325   0.94
xptiManifest             1874    46336   0.79
JS-slots                  818    45536   0.78
nsImageGTK                110    44008   0.75
nsXULPrototypeElement     891    42768   0.73
nsSupportsArray           582    32592   0.56
OTHER                   22918   885039  15.09

Comment 41

[Dimitrios]

I made my own start up time "profiling", using FileMon from
www.sysinternals.com. There were 9 sec stopwatch measured cold start time, 5.5
sec hot start (filemon activated). See attachments following (a .txt compilation
of the largest times measured, an excel worksheet containing the cold start
profiling data for the mozilla.org tree and an excel worksheet for the cold
start profiling data for the Application Data tree). 

\PROGRAM FILES\MOZILLA.ORG AND SUBDIRECTORIES
7.39 sec for 3245 uncached (cold start) file operations. 0.104 sec for 2649
cached (hot start) file operations on mozilla.org dir and subdirs. Cold start
values:
TOTAL 2.07 sec for MOZILLA.ORG\MOZILLA
TOTAL 2.3 sec for MOZILLA.ORG\MOZILLA\CHROME 
TOTAL 2.41 sec for MOZILLA.ORG\MOZILLA\COMPONENTS, including 0.90 sec for gecko
(gk*.dll)
TOTAL 0.327 sec for MOZILLA.ORG\MOZILLA\PREF
TOTAL 0.361 sec for MOZILLA.ORG\MOZILLA\RES

$HOME\APPLICATION DATA\MOZILLA AND SUBDIRECTORIES
0.566 sec for 436 uncached (cold start) file operations. 0.04 sec for 415 cached
(hot start) file operations.

MY HUMBLE CONCLUSIONS: 
- Huge part of cold start time (7.39 out of 9 sec) is due to file i/o (...you
didn't already know that , heh? :-)
- Oddly enough, hot start time is moderately reduced (9 -> 5.5 sec), although
file io is reduced to a minimal 0.566 sec in that case. Is filemon reporting
exaggerated numbers on cold start?
- chrome is a beast (2.3 sec) with some jar files taking up to 0.482 sec (comm.jar)
- xpcom is a small beast (0.621 sec) 
- Where is this precompiled chrome stuff ? xul.mfl (0.123 sec) maybe ?
- At least 0.45 sec (5.65 % of total file io) could be gained from loading of
messenger chrome and msgmapi.dll with Mail&News, instead of doing it at browser
initialisation. Should I file a bug for it?
- At least 0.106 sec (1.33 % of total file io) could be gained from delayed
loading of chatzilla chrome
- Are ucvlatin.dll (0.109 sec) and other similar libraries needed for creating
the initial browser window?
- Venkman doesn't add significant startup delay
- 721 operations on Windows registry. I speculate many of them are unecessary at
startup, but I have no idea of the time gained. Probably every file io gain
might be significant on a heavily fragmented drive. 
- Oddly, 151 operations on Windows registry resulted in NOT FOUND, most of them
are like this: HKCU\CLSID\{00021401-0000-0000-C000-000000000046}\InprocServer32.
Is this normal ?
- Extremely high level of component.reg usage reported by Daniel Bratell in bug
27510 seem not existing anymore (filemon reports 1064 KB total).

TEST CONDITIONS: Build 2001101909, new clean profile, modern theme, Win2K SP2,
K6-III/400, 192MB ram, 5400 rpm hdd

Comment 42

[Dimitrios]

Attachment: Cold start profiling data (files spending more than 20msec)

Comment 43

[Dimitrios]

Attachment: W2K cold start profiling data (excel file) from FileMon. Details inside

Comment 44

[Dimitrios]

Attachment: Second part of W2K cold start profiling data (excel file) from FileMon. Details inside

Comment 45

[Randell Jesup [:jesup] (needinfo me)]

Coldstart -> warmstart doesn't drop as much as you expected because as a
multi-threaded app, other things (CPU use) are happening during the (cold) file
loads.  On warmstart, you still have all the CPU use but can't overlap it with
file IO.  (Note: this is an assumption.)

Loading mapi/messenger chrome with mail&news would be good; I suspect this has
been thought of before and a bug exists.  Do a query (or check bugs that this
depends on) and add it if you can't find it.

I suspect that a lot of the time for warmstart (and cold) is taken doing
relocations, memory allocations and object/factory creations/initializations.
See some of the jprofs.

Comment 46

[Daniel Veditz [:dveditz]]

The mailnews and chatzilla chrome are opened because both components overlay the
browser to add items to the Tasks menu (and possibly other overlays). If we
can't cache this somehow we could create a small "mailoverlay.jar" if it helps.

Comment 47

[Dimitrios]

Randell, thank you for the clarification.  With all my ignorance about
javascript and the Mozilla structure, I do agree there's a lot of room on
reducing memory allocations and other cpu/memory intensive operations during
startup. But there's also a lot of work on reducing file i/o too. Warm start
times are quite acceptable on low end machines with at least 64MB of memory.
Cold start is what draws the attention of an average user (and he can't use
turbo mode on low memory machines, because of footprint problems). Even worse,
cold start is terribly affected by fragmentation.
Couldn't we keep library and chrome loading to a minimum needed for showing a
simple browser window? Even if start page is not blank, some delays from
late-loaded files wiil be hidden by the network latencies, once a request for
web data retrieval is issued.

Bug 12896 isn't for the mail files I was talking about. I can't find a similar
bug, so I will file a new one. As dveditz suggests, it might be feasible to find
a solution for late loading those files.

Comment 48

[Cathleen]

moving tracking/meta bugs out of milestone pools. 

Comment 49

[Henri Sivonen (:hsivonen) (not reading bugmail while on leave)]

*** Bug 174091 has been marked as a duplicate of this bug. ***

Comment 50

[Simon Paquet [:sipaq]]

*** Bug 185265 has been marked as a duplicate of this bug. ***

Comment 51

[Pham]

*** Bug 216801 has been marked as a duplicate of this bug. ***

Comment 52

[Soe L. Tun]

*** Bug 235263 has been marked as a duplicate of this bug. ***

Comment 53

[Worcester12345]

Could somebody post the link to the graphs or charts which showed the startup
times? Thanks.

Comment 54

[Worcester12345]

(In reply to comment #53)
> Could somebody post the link to the graphs or charts which showed the startup
> times? Thanks.

Anyone? Thanks again.

Comment 55

[Adam Guthrie]

*** Bug 308628 has been marked as a duplicate of this bug. ***

Comment 56

[Julien Moorrees]

I don't know if this is the right place to post this bug, but since about 3 weeks now, minefield has extremely slow startup. Just upgraded to Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv:1.9a7pre) Gecko/2007072505 Minefield/3.0a7pre and it took my 5 minutes! to boot it up. If I should post a new bug, just let me know. 

p.s. Is there any way I can make a trace of why this is taking so long?

Comment 58

[Steve Fink [:sfink] [:s:]]

How did this get reopened?

Comment 59

[Tobias B. Besemer [:BesTo] (QA)]

What's about the open 'Depends on' & 'Blocks' Bugs ???
Move to Bug 447581 ???