Charles M. Hannum wrote:

>On Saturday 03 July 2004 01:42, sigsegv@rambler.ru wrote:
>  
>
>>Charles M. Hannum wrote:
>>    
>>
>>>On Saturday 03 July 2004 01:28, sigsegv@rambler.ru wrote:
>>>      
>>>
>>>>I've been trying to measure NetBSD software RAID-0 performance on my
>>>>system, with 2 hard disks arranged for striping, however the numbers
>>>>(bandwidth) don't seem to add up, i.e.
>>>>        
>>>>
>>>There are a number of problems here.  The biggest one is that dd(1)ing
>>>from the "raw" device does not do any read-ahead.  That is to say, it's
>>>completely synchronous, and so it may not keep all the subunits busy.
>>>      
>>>
>>Well reading from a block device makes it even worse:
>>    
>>
>
>"Block" devices are sort of a sad joke that should probably be removed.  How 
>about trying it with a file in an actual file system, and say bs=32k?
>
>  
>
# mount -t ffs -o softdep /dev/raid0a /mnt

# cp /usr/home.tar /mnt
# ls -lh /mnt
total 879M
-rw-r--r--  1 root  wheel  879M Jul  3 02:51 home.tar

# dd if=/mnt/home.tar of=/dev/null bs=64k

Disks: seeks xfers bytes %busy                99 pic0 pin 0            
64 fmin
   fd0                                                                 
85 ftarg
   md0                                                              
19194 itarg
   wd0         717   22M  70.3                                        
136 wired
   wd1         720   22M  
76.2                                            pdfre
   wd2                                                              
77117 pdscn
 raid0         720   45M   100

Looks a bit better, but still, nowhere near 80M/sec for two IDE disks on 
software RAID-0, each capable of 40M/sec. By the way I tried different 
bs values, bs=32k, bs=64k, bs=512k and 64k is the optimum