[wellylug] preliminary SATA Linux software RAID0/RAID1 comparison

[Brent Wood]

--- sam at plaz.net.nz wrote:

> Quoting Cesare D'Amico <cesare at ngi.it>:
> 
> > Alle 01:22, mercoledì 13 ottobre 2004, Brent Wood ha scritto:
> > > I just need a few days to play & see if I'm memory, cpu or I/O bound.
> >
> > Can you please explain what are you going to do to check this?
> 
> vmstat will display this kind of information ... have a look at its manpage.

I use vmstat, as above, & also just watch top while things are working to see
memory/cpu useage by process. I also tend to assume that if cpu load & memory
use are not high while the drives are going crazy that I have an I/O
bottleneck.
Sometimes due to wapping, but at least I know if this is happening.

I just want a good feel for which aspects are the limitations of this setup.

I have done some prelim comparisons, using QGIS to plot about 5Gb of shapefile
data, with the data replicated on the RAID0, RAID1 & single drive partitions.
Also ran the bonnie++ checks.

I'll summarise in more detail shortly if anyone is interested, but it is hardly
a rigorous benchmark. One issue is that the work is very cpu intensive, which
suggests that it may impact on a software RAID implementation's performance.

Basically there is a reasonable but not huge performance benefit from RAID0 & a
slight performance loss from RAID1 compared with no RAID. Overall it seems to
be about 2x faster than a P4 2600 with 1Gb memory & single WD 120Gb SE UATA
drive.
(different but both 2.6 kernels)
 
I figure this increase is mostly memory/cpu based, bonnie++ & hdparm suggest a 
20% improvement in the basic HDD performance at best. 

X & qgis are the heavy users, close to 100% cpu between them.

I also want to get Postgres & PostGIS running, load the shapefiles into RDBMS
tables & try some work with the DBMS on various partitions. The Postgres
regression tests are of limited use as they don't involve any POstGIS queries,
& spatial queries can involve very different loads: eg: is a point inside the
North Island? With a million point coastline (in one record) it is not
something the regression tests tell you much about :-) 


Brent