Performance Results for Ingo's O(1)-scheduler

Performance Results for Ingo's O(1)-scheduler

[Partha Narayanan]

Here are some results from running VolanoMark on different
versions of O(1)-scheduler based on 2.4.17.

VolanoMark 2.1.2 Loopback test,
8-way 700MHZ Pentium III,
1GB Kernel,
IBM JVM 1.3. (build cx 130 -20010626)
Throughput in msg/sec


KERNEL              UP          4-way       8-way
=========           ======      ======      ======

2.4.17              11005       15894       11595

2.4.17 + D2 patch   10606       23300       29726

2.4.17 + G1 patch   10415       23038       31098

2.4.17 + H6 patch   10914       22270       32300

2.4.17 + H7 patch   11018       23427       31674

2.4.17 + J2 patch   13015       23071       33259




Partha Narayanan
partha@us.ibm.com

Re: Performance Results for Ingo's O(1)-scheduler

[Ingo Molnar]

On Mon, 21 Jan 2002, Partha Narayanan wrote:

> KERNEL              UP          4-way       8-way
> =========           ======      ======      ======
> 2.4.17              11005       15894       11595
> 2.4.17 + D2 patch   10606       23300       29726
> 2.4.17 + G1 patch   10415       23038       31098
> 2.4.17 + H6 patch   10914       22270       32300
> 2.4.17 + H7 patch   11018       23427       31674
> 2.4.17 + J2 patch   13015       23071       33259

thanks for the testing - i'm happy that it's the kernel with the best
interactive properties (-J2) that has the best VolanoMark results as well.

	Ingo

Re: Performance Results for Ingo's O(1)-scheduler

[Adam Keys]

On January 21, 2002 06:03, Partha Narayanan wrote:
> Here are some results from running VolanoMark on different
> versions of O(1)-scheduler based on 2.4.17.
>
> VolanoMark 2.1.2 Loopback test,
> 8-way 700MHZ Pentium III,
> 1GB Kernel,
> IBM JVM 1.3. (build cx 130 -20010626)
> Throughput in msg/sec
>
>
> KERNEL              UP          4-way       8-way
> =========           ======      ======      ======
>
> 2.4.17              11005       15894       11595
>
> 2.4.17 + D2 patch   10606       23300       29726
>
> 2.4.17 + G1 patch   10415       23038       31098
>
> 2.4.17 + H6 patch   10914       22270       32300
>
> 2.4.17 + H7 patch   11018       23427       31674
>
> 2.4.17 + J2 patch   13015       23071       33259

I'm curious about the performance of the 4-way and 8-way systems.  I know 
nothing about this benchmark.  IIRC correctly it simulates chat clients 
connecting to a server and talking to each other.  Is it a CPU, memory, or 
disk bound benchmark?  What is causing the 4-way machines to be only 2x the 
performance of the 1-way machine and the 8-way machines to be < 3x the 
performance?  Is the system bus the limiting factor on those machines?

Curiosity aside, it looks like Ingo's scheduler is coming along nicely.

-- 
akk~

Re: Performance Results for Ingo's O(1)-scheduler

[Ryan Cumming]

On January 21, 2002 19:55, Adam Keys wrote:
> I'm curious about the performance of the 4-way and 8-way systems.  I know
> nothing about this benchmark.  IIRC correctly it simulates chat clients
> connecting to a server and talking to each other.  Is it a CPU, memory, or
> disk bound benchmark?  What is causing the 4-way machines to be only 2x the
> performance of the 1-way machine and the 8-way machines to be < 3x the
> performance?  Is the system bus the limiting factor on those machines?

Memory bus, lock contention, syncronization issues. SMP really isn't as 
magical as people think after the overhead is taken in to account.

-Ryan

Re: Performance Results for Ingo's O(1)-scheduler

[Rusty Russell]

On Mon, 21 Jan 2002 20:16:28 -0800
Ryan Cumming <bodnar42@phalynx.dhs.org> wrote:

> On January 21, 2002 19:55, Adam Keys wrote:
> > I'm curious about the performance of the 4-way and 8-way systems.  I know
> > nothing about this benchmark.  IIRC correctly it simulates chat clients
> > connecting to a server and talking to each other.  Is it a CPU, memory, or
> > disk bound benchmark?  What is causing the 4-way machines to be only 2x the
> > performance of the 1-way machine and the 8-way machines to be < 3x the
> > performance?  Is the system bus the limiting factor on those machines?
> 
> Memory bus, lock contention, syncronization issues. SMP really isn't as 
> magical as people think after the overhead is taken in to account.

Volcanomark is a Java(TM) chatroom benchmark: multiple rooms, where for each
room, every input from a client generates a write to every other client
(think broadcast storm).

chat (which is a C version of Volcanomark) is useful for testing, as is
hackbench2 (which is cut down to just exhibit the runqueue problem, and
doesn't even use threads).

Hope that helps,
Rusty.
-- 
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.

Re: Performance Results for Ingo's O(1)-scheduler

[Mark Hahn]

> connecting to a server and talking to each other.  Is it a CPU, memory, or 
> disk bound benchmark?  What is causing the 4-way machines to be only 2x the 

none of the above: it's scheduler-bound: billions of runnable threads
that do almost nothing but wake up other threads. a fine example of 
where to use coroutines, not kernel threads...

Re: Performance Results for Ingo's O(1)-scheduler

[Dieter Nützel]

On Tuesday, 22. January 2002 00:03, you wrote:
> Here are some results from running VolanoMark on different
> versions of O(1)-scheduler based on 2.4.17.
>
> VolanoMark 2.1.2 Loopback test,
> 8-way 700MHZ Pentium III,
> 1GB Kernel,
> IBM JVM 1.3. (build cx 130 -20010626)
> Throughput in msg/sec
>
>
> KERNEL              UP          4-way       8-way
> =========  ======      ======      ======
>
> 2.4.17              11005       15894       11595
[-]

Would you be so kind to redo it with 2.4.18pre2aa2 (Andrea's 10_vm-22) stuff?

Thank you very much.

	Dieter

-- 
Dieter Nützel
Graduate Student, Computer Science

University of Hamburg
Department of Computer Science
@home: Dieter.Nuetzel@hamburg.de

Re: Performance Results for Ingo's O(1)-scheduler

[Ingo Molnar]

On Tue, 22 Jan 2002, Dieter [iso-8859-15] Nützel wrote:

> > KERNEL              UP          4-way       8-way
> > =========  ======      ======      ======
> >
> > 2.4.17              11005       15894       11595
> [-]
>
> Would you be so kind to redo it with 2.4.18pre2aa2 (Andrea's 10_vm-22) stuff?

prediction: the results wont change a bit.

VolanoMark has no sensitivity on VM issues at all.

	Ingo