[OpenAFS] Extremely poor write performance.

[jon+openafs@silicide.dk]

On Fri, Jan 17, 2003 at 01:11:20AM +0100, Rubino Gei? wrote:

 
> ############### daemons = 3
> 
> Version  1.03       ------Sequential Output------ --Sequential Input-
> --Random-
>                     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
> --Seeks--
> Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec
> %CP
> i44pc56.info.u 128M  1981  54  4211  12  3494  14  2917  78 38557  50 676.5
> 6
>                     ------Sequential Create------ --------Random
> Create--------
>                     -Create-- --Read--- -Delete-- -Create-- --Read---
> -Delete--
>               files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec
> %CP
>                  16    97  45   840  89   187  42   103  47   531  72   177
> 24
> i44pc56.info.uni-karlsruhe.de,128M,1981,54,4211,12,3494,14,2917,78,38557,50,
> 676.5,6,16,97,45,840,89,187,42,103,47,531,72,177,24
> 
> ############### daemons = 300
> 
> Version  1.03       ------Sequential Output------ --Sequential Input-
> --Random-
>                     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
> --Seeks--
> Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec
> %CP
> i44pc56.info.u 128M  2128  58  4282  12  3124  13  3265  86 29197  45 495.0
> 5
>                     ------Sequential Create------ --------Random
> Create--------
>                     -Create-- --Read--- -Delete-- -Create-- --Read---
> -Delete--
>               files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec
> %CP
>                  16    99  41   812  84   167  36    90  37   344  47   159
> 22
> i44pc56.info.uni-karlsruhe.de,128M,2128,58,4282,12,3124,13,3265,86,29197,45,
> 495.0,5,16,99,41,812,84,167,36,90,37,344,47,159,22
> 
> >>>>>>>> Conclusion: No effect.

You cant conclude from a single bonnie? run.
You need to thread it. You need to have more concurrent read and writes.
You need to read and write different file sizes.
Why, because thats usualy how people access their data.
I'll suggest you use tiobench, og iometer from intel.

Of course a benchmark can never really substitute trying with real data on
real servers, under a real load.



JonB