[parsec-users] speedup of bodytrack

cbienia at CS.Princeton.EDU cbienia at CS.Princeton.EDU
Mon Jun 16 09:32:44 EDT 2008


Hello Huan,

A number of things limit the speedup of bodytrack in practice:

1. The particle resampling kernel has not been parallelized yet. This 
limitation is included in Figure 1.

2. Bodytrack has to load 4 pictures for every frame which it processes. 
This is handled synchronously, which means the program is blocked and 
has to wait while the images are loaded.

3. Depending on your machine parameters, the feature extraction kernel 
quickly becomes memory bound. On real machines that happens around a 
speedup of 4x.

- Chris



Quoting Huan Fang <huanf at kth.se>:

> Hi,
>
> From the parsec-report.pdf, I found the achievable speedup of 
> bodytrack is quite different than what I get. I measured it in Simics 
> with n*UltraSPARCIII processor at 1GHZ, 4GBytes Memory. The speedup 
> of bodytrack using simlarge is around 3 on 4processors and 6 on 
> 16processors.
>
> As said in the report "Figure 1 show the maximum achievable speedup 
> measured that way." and "We neglect any delays due to blocking on 
> contended locks and load imbalance." Could you explain more on that 
> and is there any way we could quantify the amount of lock-blocking or 
> load imbalance?
>
> Best Regards,
> Huan Fang
> _______________________________________________
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>





More information about the parsec-users mailing list