[Pvfs2-users] collective IO performance

[Rob Ross]

Hi,

What version of MPI are you using? What are the collective buffer sizes 
set to? How many aggregators?

Thanks,

Rob

Xinhua Gao wrote:
> Hello,
> 
> I am using IOR to write/read a shared file with up to 16 processes, each
> writing data of 128M in units of 256K. Here is the data I got with
> collective IO and independent IO.
> 
> # collective IO 
> #procs write read
>  2  74.53 12.43
>  4  36.28 15.85
>  6  20.66 18.61
>  8  20.16 23.57
> 10  22.18 33.15
> 12  24.18 49.36
> 
> # independent IO to a single file
> # proc   write   read (MB/sec)
>  2   51.98   28.55
>  4   65.15   55.73
>  6   83.03   70.01
>  8   98.96   82.06
> 10   124.65   102.35
> 12   126.86   112.36
> 14   139.63   119.50
> 16   147.01   154.90
> 
> It is a little surprising to see collective IO performance is so low
> compared to independent IO. In addition, collective IO does not increase
> with increasing number of clients. The set servers has 5 IO servers can
> support at least 100MB/sec of IO rates directly from disks.
> 
> Anybody can provide some hints? Thanks,
> 
> Xinhua
> 
> 
> 
> _______________________________________________
> Pvfs2-users mailing list
> Pvfs2-users at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
>