FreeBSD + ZFS on a production server?

[Wojciech Puchar]

> Wojciech Puchar wrote:
>
>> 2) ZFS RAID-z turns your X drives to single drive performance both on read 
>> and write. every normal RAID-5 implementation will give you random read 
>> speed of X-1 times single drive speed, while slow random write speeds (but 
>> still at least half of single drive). but this is advertised as a feature
>
> Is this because of checksum verification (the need to read all components) or

nothing to checksum. sit down for a while and think.

you have say 100 MB file linearly placed on position A on disk.
and your program requests 100kB writes to possitions like 200kB*(between 0 and 499) randomly.

ZFS will cache all it then blow all this as 50MB linear write" at 
position B on disk.

then after some time (data not in cache) you like to read file linearly.

what you get:

read 100kB from position B, seek to position A to read 100kB, then to 
position B reading 100kB, then to position A etc. etc.

LOTS of seeking.

while reads are more common than writes on most cases ZFS make things 
worse.

of course i told about good case where ZFS could find large continous 
space. if your drive is well filled it's unlikely.


with UFS disk is divided for cylinder groups. so too - it's unlikely you 
will find large continuous space BUT there are very likely you will find 
large chunk of fragments withing same cylinder group which requires 
much shorter head movement. while really big blocks are forcibly splitted 
to different cylinder group, as having long seek every few megabytes isn't 
a problem.

that's what UFS does for 20 years. the major improvement then was soft 
updates, now it is really fast even with small files.



i just skipped talking about memory and CPU usage, because there is 
nothing to talk much. it's just make ZFS crap and nothing else.

Today most people's problem are TOO FAST CPU and TOO MUCH MEMORY.Sun found 
the solution, just like windows.