iozone-ing an SSD (Re: Using an SSD "disk" for /)
Mikhail T.
mi+thunw at aldan.algebra.com
Sat Nov 6 00:14:06 UTC 2010
Hello!
So, after an earlier inquiry, I went ahead and purchased an SSD
(Crucial's CTFDDAC128MAG-1G1) and put it to some testing today.
The computer is Dell Poweredge 2900, running FreeBSD-8.1/amd64 (the
October 10th snapshot). Generic kernel. The system drive (for now) is
traditional "real" HD -- a 15K RPM by Fujitsu (MAX3073RC), I ran `iozone
-a' 4 times:
1. On /var/tmp -- freshly newfs-ed by the sysinstall on the Fujitsu
drive (/dev/da0).
2. On the SSD (/dev/ad4) freshly newfs-ed by me without ANY options
(no softupdates).
3. On the SSD (/dev/ad4) freshly newfs-ed by me with very large -e
and -a options. Reading the man-page, I figured, any parameters
mentioning "cylinders" can be set to very large values...
4. On the SSD (/dev/da1) connected to the server's mpt-controller,
rather than the plain SATA port -- using the same filesystem created in
3. above (no reformatting). (The 2.5" can't be secured in the 3.5" slot
and is simply hanging in the air on the SATA/SAS connectors.)
The results can be found in 4 HTML files found at:
http://aldan.algebra.com/~mi/io/ (The original iozone-created Excel
files are there too.)
They puzzle... Fujitsu, for example, is not an OBVIOUS loser -- it beats
the SSD in a number of file-size record-length combinations. I also
can't explain, the differences between different takes on the SSD.
And, lastly, there is a surprising (to me) spike in "Record Rewrite"
throughput -- for both SSD and HD -- for large files when the reclen is
64. Using reclen of 128 results in much worsening throughput --
especially for the Fujitsu.
I wonder, if these data can be exploited to come up with better newfs
parameters for the modern disks (SSD and not)... Comments? Thanks!
-mi
More information about the freebsd-fs
mailing list