13009 - service manager has too many locks

Status: RESOLVED WORKSFORME

(Core :: XPCOM, defect, P2)

Description

[Chris Waterson]

The service manager has APIs that are protected by a monitor, as well as using
a threadsafe hashtable for internal use. Travis said "service manager is
dying", so I'm handing him this bundle o' joy to remind him that the new
incarnation shouldn't do that.

Comment 1

[travis]

This probably belongs to you dp as you are the one working on that side of the
service manager.

Comment 2

[Suresh Duddi (gone)]

It might take a while before the servicemanager goes away. In the mean time, any
suggestions for improvement here.

Comment 3

[Suresh Duddi (gone)]

This is a performance bottleneck. We should see if we can eliminate the
excessive locking.

Comment 4

[Warren Harris]

If the service manager entry points are protected by a lock, the hashtable
doesn't need to use one. I see line 180 of nsServiceManager.cpp says "Get a
threadSafe hashtable" which may not be necessary any more. I'd investigate
setting that parameter to false.

Comment 5

[Suresh Duddi (gone)]

Just curious, why are you eliminating the other case: remove locks from apis and
protect the data (hashtable entries).

Comment 6

[daver]

performance team should nominate this to make it happen.

Comment 7

[daver]

Correction, performance team should nominate this IF they want to make it
happen.

Comment 8

[Suresh Duddi (gone)]

Performance guys, can you evaluate this and let us know if we should fix it.

Comment 9

[Suresh Duddi (gone)]

whatever. We are going to leave it like this for a while.

Comment 10

[leger]

Moving "perf" to keyword field.  This is the are we use now :-)

Comment 11

[Peter Trudelle]

moving from leftover dogfood to beta1 radar per beta criteria priority #2

Comment 12

[leger]

Need to quanify what impact this has on the spec and where you think it will 
help and how much please.

Comment 13

[Mike Shaver (:shaver emeritus)]

Where is the spec?

Comment 14

[Suresh Duddi (gone)]

This bug came out of the timing analysis done by the performance effort. I am 
sure this was a blaring number and hence the bug. We need to quantify to really 
get the number by which this is a problem and further by how much we can reduce 
it since there is no way we can eliminate all the locks.

I know this doesn't help much. Just thinking aloud at this point. My general 
sense was that this wont improve performance a great deal. So let us go after 
bigger fish...

Comment 15

[leger]

Putting on PDT- radar for beta1.  If you have date to make a better call, let us 
know - pdt

Comment 16

[Warren Harris]

A recent Quantify run shows that GetService accounts for 15.18% of total 
runtime. The enter/exit monitors there account for about 2% of that.

The service manager also uses a threadsafe hashtable, which accounts for 
another 1%. That hashtable is threadsafe (unnecessarily?), and 44% of all time 
spent in nsHashtable::Get is entering and exiting the lock. 

Attempting to calculate this all out, I come up with (.02 + .44 * .01) * .15 
==> .366% of execution time manipulating these locks. However, the simple fix 
is probably to change the flag passed to the hashtable constructor to make it 
not be threadsafe since the higher level locks are already protecting it.

Comment 17

[Jim Roskind]

Based on Warren's analysis, there is a tiny .366% gain (even if we got it down 
to zero locks) and notable risk.  PDT gave this a PDT- for beta1.

Comment 18

[leger]

Moving all current open XPCOM and XPCOM Registry bugs to rayw since dp is on 
sabbatical.  rayw is now default assignee for these components.

Comment 19

[Warren Harris]

Assigning rest of xpcom stuff to myself (from Ray).

Comment 20

[chris hofmann]

-> kandrot 

now is a good time for smaller and larger perf wins...

Comment 21

[Doug Turner (:dougt)]

reassign all kandrot xpcom bug.

Comment 22

[Doug Turner (:dougt)]

neeti is working on this as part of her cleanup/performance work in xpcom.

Comment 23

[Henry Jia]

Add myself to the CC list. I'll try to see if I can do some work on it.

Comment 24

[neeti]

This was fixed by bug 96461. nsComponentManager now uses PLDHashtable, instead 
of the threadsafe version of nsHashtable.