Subject: Re: slowdown with MI scsi on pmax (ds5000/200)
To: Toru Nishimura <nisimura@itc.aist-nara.ac.jp>
From: Chuck Silvers <chuq@chuq.com>
List: tech-kern
Date: 05/21/2001 05:20:35
hi tohru,

you're talking about filesystem performance, whereas I'm talking about
raw disk performance (ie. writing to /dev/rsd0h).

about a month ago I had filesystem writes running within 1% of the NetBSD 1.5
speed.  I'm getting close to being ready to check in those changes, but now I
discover that raw disk access has gotten much slower in the meantime, which
slows down filesystem performance as well.

-Chuck


On Mon, May 21, 2001 at 05:03:57PM +0900, Toru Nishimura wrote:
> Chuck Silvers <chuq@chuq.com> wrote;
> 
> > hi,
> >
> > I went back to run some benchmarks on my decstation 5000/200 again
> > first the first time in about a month, and I find that scsi performance
> > has dropped considerably.
> 
> I'm experiencing similars on my DECstation.
> 
> My 1.5V kernel for 33MHz R3000A DECstation brings strange SCSI
> performance figures I've never seen before.
> 
>   NetBSD 1.5V (MAXINE) #0: Tue May  8 13:28:54 JST 2001
>       nisimura@build:/usr/src/sys/arch/pmax/compile/MAXINE
> 
> iozone consistently shows WRITE runs half than READ.
> 
>   IOZONE performance measurements:
>         869285 bytes/second for writing the file
>         2111935 bytes/second for reading the file
> 
> The READ performance is restored as good as the one when I made pmax
> MI SCSI driver years ago, but WRITE is the worst as long as I know.
> iostat and vmstat indicate the kernel beats SCSI hardware par with the
> data rate.
> 
> $ /usr/sbin/iostat -D 1
>    KB xfr time 
>   ...
>   854  15 0.52 
>   743  16 0.46 
>   760  18 0.55 
>   786  15 0.48    <------ write has finished here
>  1311  33 0.63    <------ starts reading
>  2039  34 0.70 
>  2091  34 0.70 
>  1904  34 0.72 
>   ...
> 
> $ vmstat 1
>  procs   memory     page                       disk faults      cpu
>  r b w   avm   fre  flt  re  pi   po   fr   sr s0   in   sy  cs us sy id
>  ...
>  1 0 0 20860  1192    0   0   0  215    0    0 15    0    0   0  1 99  0
>  1 0 0 21708   344    0   0   0  213    0    0 16    0    0   0  0 98  2
>  1 0 0 21436   616    0   0   0  213    0    0 14    0    0   0  1 99  0
>  1 0 0 21048  1008    0   0   0  182    0    0 16    0    0   0  1 92  7 <- W
>  1 0 0 21572   464    0   0   0  213    0    0 38    0    0   0  1 83 16 <- R
>  1 1 0 21276   780    0   1   0  257    0    0 34    0    0   0  2 91  7
>  1 0 0 20960  1036    0   2   0  258    0    0 33    0    0   0  3 90  8
>  1 0 0 20704  1292    0  14   0  393    0    0 32    0    0   0  2 66 32
>  1 0 0 20768  1232    0   1   0  280    0    0 33    0    0   0  2 92  6
>  ...
> 
> The NetBSD/pmax kernel has, UBC, "UBC fix for throttling", SCSIPI
> integration and kthread fix.
> 
> Tohru Nishimura