Running GUI application in separate application domain

[Guest]

I'm trying to isolate "applications" into their own application domain within
a single process. I've quoted applications because it's a logical
representation of an application. Basically it consists of a bunch of
components supplied by some application group. I got this to work, somewhat.
The problem is that the application performs roughly (and this has not been
measured, but a guess based on the rendering of the application GUI) 10x
slower than a similar (almost identical) application in which all of the
components are created within the main (/default) application domain. Is
there some inherent performance problem with running GUI applications in
separate application domains?

By the way, I'm also posting this to the dotnet general newsgroup and dotnet
framework clr newsgroup.

 

[Guest]

I've got some additional information. Via configuration I was able to move
all these components into a single application domain, separate from the
default domain. The performance of this matches the performance of the
existing application where all the components are in the default domain. In
both of these scenarios I noticed the % Time in GC is around 7%. When these
components are spread across two application domains the % Time in GC is
around 20% and as I mentioned the performance appears to be about 10x worse.
I'm running the application from my laptop which is a single processor
machine. Why would separating the work of a single application to multiple
application domains have such a drastic change on the GC?

 

[Lloyd Dupont]

components are created within the main (/default) application domain. Is

there some inherent performance problem with running GUI applications in
separate application domains?

your question is tricky because neither yes or no could be replied.
on one hand GUI has nothing to do with decreasing performance
on the other hand Remoting has a lot to do with them.

When 'application' or 'component' are in different application domain they
can't call each other, you have to go through remoting to marshall call and
return value.
Although a specialized version of remoting with a faster in memory channel
is used for inter domain communication, there is still a price to pay which
visually perceptible, as you saw...

 

[Guest]

What remoting are we talking about. I create two components in one
application domain and two in another. They don't talk to each other. There
should be no remoting/marshaling going on, that I know of. When I compare
the performance of this configuration against a "similar" application which
creates these same components all in the default application domain the
performance is drastic. The multi-appdomain model is at least 10x slower.

 

[Guest]

By the way, forget my second post. When I tried running all the components
in one appdomain, which was not the default appdomain, again the performance
was very poor. I guess the first time was a fluke. By the way, I talk about
performance but I guess some may be confused about what I mean. The
components are grid components (classes derived from the SyncFusion grid)
which are displaying real-time data updates over a TIB bus. When these same
components are all run in the default application domain the application can
keep up with the real-time updates. When I create two of the grids in one
appdomain and two in another the performance decreases drastically.

 

[Lloyd Dupont]

What remoting are we talking about. I create two components in one

application domain and two in another. They don't talk to each other.
There
should be no remoting/marshaling going on, that I know of. When I compare
the performance of this configuration against a "similar" application
which
creates these same components all in the default application domain the
performance is drastic. The multi-appdomain model is at least 10x slower.
--

Now I wonder, if you want to display this component you have:
1. to put them on a visible form
2. respond to all the mouse move, mouse down, key and paint events
3. feed them some data.

all of this takes a lot of method => a lot of remoting call

or could you explain more in detail what you do?
how you could avoid remoting while still having multiple app domain per
component (assuming component = something inheriting from SWF.Control) ?

 

[Guest]

I create two application domains. In each of those domains I create two GUI
controls. They do not communicate amongst each other. There should be no
remoting going on, at least as far as I can tell. Think of it this way:

I have this current application, which you can think of as your typical
application. It has a mainform and some controls (though all of my controls
are modeless dialogs) and they all run in the default application domain.
Now you decide you want to run this whole thing in a different application
domain, other than the default. So in Main() of the default appdomain you
create a new appdomain and execute an assembly on a different thread. The
main of this new assembly does exactly what the original program did. Don't
know if this explains it any better. I hope so.

 

[Lloyd Dupont]

I create two application domains. In each of those domains I create two

GUI
controls. They do not communicate amongst each other. There should be no
remoting going on, at least as far as I can tell. Think of it this way:

I have this current application, which you can think of as your typical
application. It has a mainform and some controls (though all of my
controls
are modeless dialogs) and they all run in the default application domain.
Now you decide you want to run this whole thing in a different application
domain, other than the default. So in Main() of the default appdomain you
create a new appdomain and execute an assembly on a different thread. The
main of this new assembly does exactly what the original program did.
Don't
know if this explains it any better. I hope so.

mmhh...
yep that's clear.
does you Main() method is tagged with [STAThread] ?
I *suppose* that [STAThread] and multiple app domain would create some
internal remoting overhead....
but probably not that much.... mhh... just an idea...

 

[Guest]

Yes, I did add the [STAThread] attribute to the Main() of this loaded
assembly. However, for the thread I started that calls
AppDomain.ExecuteAssembly() I did not specify any apartment setting. I
assume that's OK.
--
Thanks,
Nick

I create two application domains. In each of those domains I create two
GUI
controls. They do not communicate amongst each other. There should be no
remoting going on, at least as far as I can tell. Think of it this way:

I have this current application, which you can think of as your typical
application. It has a mainform and some controls (though all of my
controls
are modeless dialogs) and they all run in the default application domain.
Now you decide you want to run this whole thing in a different application
domain, other than the default. So in Main() of the default appdomain you
create a new appdomain and execute an assembly on a different thread. The
main of this new assembly does exactly what the original program did.
Don't
know if this explains it any better. I hope so.

mmhh...
yep that's clear.
does you Main() method is tagged with [STAThread] ?
I *suppose* that [STAThread] and multiple app domain would create some
internal remoting overhead....
but probably not that much.... mhh... just an idea...

 

[Guest]

I just ran some more tests and captured perfmon logs for the two different
scenarios. When I run the GUI components in the default app domain I get the
following results (I only lists the ones that are significantly different and
could potentially indicate the problem).

..NET CLR Security\Total Runtime Checks: Min=122,443 Max=358,890 Delta=~236,000
..NET CLR Memory\# Gen 0 Collections: Min=149 Max=9927 Delta=~9800
..NET CLR Memory\% Time in GC: Average=6
..NET CLR Memory\Allocated Bytes/sec: Average=12MB
..NET CLR Memory\Finalization Survivors: Average=344

The same counters for the scenario where the components are run in an
appdomain different from the default domain is:

..NET CLR Security\Total Runtime Checks: Min=124,677 Max=1,475,320 Delta=~1.3M
..NET CLR Memory\# Gen 0 Collections: Min=762 Max=69,766 Delta=~69,000
..NET CLR Memory\% Time in GC: Average=17
..NET CLR Memory\Allocated Bytes/sec: Average=32MB
..NET CLR Memory\Finalization Survivors: Average=38

For for some reason running these components in a different application
domain drastically increased the # of runtime security checks, # of gen 0
collections, % Time in GC, and the allocated bytes/sec. Note that the server
which generates the real time data feed is the same in both cases.

Any ideas?

 

[Guest]

I'm pretty certain is has to do with Security Runtime Checks. When I turn
Code Access Security off (caspol -s off) the performance is fine and of
course the runtime security checks perfmon counter is zero.

This is where some of the confusion comes in. When CAS is turned on
sometimes the performance is good and sometimes it isn't. However, ever
since I've started looking at the security runtime checks performance counter
I've noticed that every time the performance is slow I'm doing huge amounts
of runtime security checks. On the order of 3200/sec. When CAS is on and
the peformance is good the runtime security checks counter is much lower,
maybe 300/sec - 500/sec. I can't figure out why sometimes this counter will
be high and other times it will be low for the same assemblies. I've
literally ran the program once and noticed poor performance and then ran the
same program again (no changes) and noticed good performance. Again, the
poor performing run was generating loads of runtime security checks and the
good performing run was not.

 

[Lloyd Dupont]

if you've got VS2005 you could run a performance session.
It will tells how much time you spend in your method (in absolute and %
value).
by running 2 session: 1 with multiple app domain, one without.
you will be able to know which method takes more time, hence finding out
which security check are done and why.

 

[Willy Denoyette [MVP]]

Yes, I did add the [STAThread] attribute to the Main() of this loaded
assembly. However, for the thread I started that calls
AppDomain.ExecuteAssembly() I did not specify any apartment setting. I
assume that's OK.

No, it's not, see my response to the same thread in dotnet.clr NG.

Willy.