Re: stripe_cache_size ?

[Bill Davidsen <davidsen@tmr.com>]

Neil Brown wrote:

>On Friday December 9, tmrbill@tmr.com wrote:
>  
>
>>On Fri, 9 Dec 2005, Neil Brown wrote:
>>
>>    
>>
>>>On Friday December 9, kylewong@southa.com wrote:
>>>      
>>>
>>>>Hi,
>>>>
>>>>I found that there's a new sysfs "stripe_cache_size" variable. I want to
>>>>know how does it affect RAID5 read / write performance (if any) ?
>>>>Please cc to me if possible, thanks.
>>>>        
>>>>
>>>Would you like to try it out and see?
>>>Any value from about 10 to a few thousand should be perfectly safe,
>>>though very large values may cause the system to run short of memory.
>>>
>>>The memory used is approximately
>>>   stripe_cache_size * 4K * number-of-drives
>>>      
>>>
>>What??? I hope that's a typo...
>> 1 - there's no use of the sysfs variable?
>>    
>>
>
>'stripe_cache_size' is the sysfs variable.  Yes, it is used.
>
>  
>
>> 2 - that's going to be huge, 128k * 4k * 10 = 5.1GB !!!
>>    
>>
>
>That is why I warned to limit it to a few thousand (128k is more than
>a few thousand!).
>  
>
Sorry, for some reason I read that as being in stripes instead of bytes, 
which would make it 128k for size only 2. My misread.

>I just ran bonnie over a 5drive raid5 with stripe_cache_size varying
>in from 256 to 4096 in a exponential sequence. (Numbers below 256
>cause problems - I'll fix that).
>
>Results:
> 256   cage,8G,42594,93,151807,38,50660,18,38610,91,172056,38,912.8,2,16,4356,99,+++++,+++,+++++,+++,4389,99,+++++,+++,14091,100
> 512   cage,8G,42145,92,186535,44,60659,21,42249,96,172057,37,971.9,2,16,4407,99,+++++,+++,+++++,+++,4452,99,+++++,+++,13909,99
>1024   cage,8G,42250,92,210407,50,61254,21,42106,96,172575,37,903.1,2,16,4370,99,+++++,+++,+++++,+++,4395,99,+++++,+++,13809,100
>2048   cage,8G,42458,92,229577,55,61762,21,41965,96,168950,36,837.9,2,16,4373,99,+++++,+++,+++++,+++,4460,99,+++++,+++,14084,100
>4096   cage,8G,42305,92,250318,62,62192,21,42156,96,170692,38,981.8,3,16,4380,99,+++++,+++,+++++,+++,4426,99,+++++,+++,13723,99
>
>Seq Write speed           ^
>Increases substantially.
>Seq Read                                             ^
>Doesn't vary much.
>Seq rewrite                         ^
>improves a bit
>
>So for that limited test, write speed is helped a lot, read speed
>isn't.
>
>Maybe I should try iozone...
>
>NeilBrown
>
>  
>


-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979