Re: 5.1 RAID5 write performance

["Michael L. Hitch" <mhitch%lightning.msu.montana.edu@localhost>]

On Thu, 5 May 2011, Edgar Fuß wrote:

> > The block device writes through the buffer cache which has its own
> > idea of block sizes which is either the filesystem block size stored
> > in the disklabel for that partition or 2048 bytes as a default.
> Yes, I'm aware of that. But that shouldn't lead to a hundredfold loss of 
> throughput, should it?

  Well, depending on how the raidframe code actually works, it could certainly 
slow things down quite a bit.  Assuming you are trying to write 64k blocks to 
the block device and a 2K buffer cache size, that means the one write of 64k 
blocks will be split into 32 buffers, with the data copied from the user 
address space to the buffers.  Then each of those buffers would be written to 
the raid device.  The raidframe driver might them be reading the full data 
strip into buffers, updating the 2K portion corresponding to the request, and 
then writing the full data strip back to the disk.  [Note that I don't really 
know if the raidframe driver actually works that way, but it would certainly 
account for a massive slowdown].

  Writing to the raw device will map the 64k buffer being written to the raid 
device and pass that entire buffer to the raidframe driver.  If the data is 
aligned on a strip boundary, then the raidframe driver could write that 
directly to the disk (or perhaps just a copy to a raidframe buffer which would 
then be written directly to the disk).

--
Michael L. Hitch                        mhitch%montana.edu@localhost
Computer Consultant
Information Technology Center
Montana State University        Bozeman, MT     USA