Re: [simspark-devel] Simulator Speed: Suggestion

[Joschka B. <jos...@am...>]

Hi Mahdi,

CC Hedayat, Carlos, and the developer's list

On Jul 31, 2008, at 5:50 AM, Zigorat Team wrote:

> Hi Joschka!
>
>   Hope you're well.
>

Thanks, I'm doing okay. Hope you're well, too.

> There is a -thing- in my mind about speeding up the simulator, I  
> wanted to ask you about that before it goes to the mailing list:
>   You know, the typical frame rate of a realtime 3D graphic  
> application is about 30. It means that every 33ms, a frame is drawn.  
> In this way there is absolutely no difference from reality in a  
> human's eye.
>   In our simulator, instead of graphic parts, the burden is on the  
> physics calculation section.

To be honest, I think we first should make sure this is really the  
case. I mean, I suspect it, too, but we should do some profiling under  
competition conditions (i.e. in a network setting with external  
monitor and realistic agent binaries). This is the only way we can  
find out where the biggest amount of computation is really used. In  
the past, for instance, Zeitgeist was to a large extent responsible  
for slowing down the server. We addressed these problems in a number  
of ways, but we should make sure they are really fixed now. Another  
candidate for needing a lot of processing power is the generation,  
parsing, sending, and receiving of network messages. We have to look  
at all of these.

We also should try to determine how well the server runs with the  
multi-threaded timer and also using Hesham's ODE improvements which  
weren't used in Suzhou (as far as I know). These two things have  
potential to speed up the server considerably.

> Currently the cycle rate of the simulator is 50 per second => 20ms.  
> I think that (I have no idea) that even in a real humanoid robot,  
> processings are more time consuming than 20ms.

It depends on what part of the processing in a humanoid you mean. If  
it comes to control, 20ms is probably close to the upper limit I  
guess, otherwise your movements might get jerky. I remember in our old  
HOAP-2 robot, we had a dedicated microprocessor for the control that  
had a timestep of 1ms!

> I had a brief discussion with carlos and hedayat, and we agreed that  
> 40 ms won't do much harm to the simulations. How do you think of  
> changing the simulator's cycle rate to 25? I only mean the  
> simulators, not the monitor's frame rate.
>

I can't really predict what the effects will be, but let's try ;-)

Thanks a lot for the input!

Best regards,
Joschka