tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: RAIDframe level 5 write performance (Brian Buhrow) writes:

>NetBSD-4.x with softdep enabled came in with a write speed of about
>NetBSD-5.1 on FFSv2 without softdep or wapbl came in with a write speed of
>about 8mbytes/sec.
>The read speed on both installations was above  200mbytes/sec.

RAID-5 can only write reasonably fast when you write blocks as large
as a stripe and if the writes and the stripes are aligned. Often you
need even multiple stripe sizes to get effective double buffering.

That's hard to do, since the filesystem will not write more than 64k
(MAXPHYS) in a single chunk.

If you do write smaller blocks, the RAID controller will issue read-
modify-write cycles for each strip. And that can easily be slower
by several orders of magnitude than streaming data.

The workaround is then to have a RAID controller with a smart write cache
that can coalesce multiple small writes into stripe sized writes.

                                Michael van Elst
                                "A potential Snark may lurk in every tree."

Home | Main Index | Thread Index | Old Index