[Sunhelp] Slow writes compared to reads with DiskSuite

Magnus Abrante Magnus.Abrante at nocke.Sweden.Sun.COM
Thu Jun 15 10:06:28 CDT 2000


In particular if one of the disks in a software RAID 5 breaks... then 
it will be very very slow since it has to perform even more calculations.

If you really want SDS/RAID 5, dont forget to create a hotpol.

If you want better performance you should use a RAID 0/striping, that 
would improve both read/write performance. Or a RAID 0+1 or 1+0 if you
want some security as well. This of course depends on how much disk you
have, and what kind of configuration you want, if you for example want to
be able to grow a metadevice on the run you shouldn't use RAID 0/stripe.

I dont think you'll be able to add the HW RAID controller for the internal
disks, i can't really see how that would be possible anyway, since the
RAID card wants full control over the disks which are under its
control. At least not AFAIK.

        //Magnus
/* This is not the opinion of my company, its just my own */


> | Assuming that this was a fair test, I guess then it is normal to have 
> | asymmetrical reads v writes with DiskSuite. I'll also buy the 
> | differences between the U1 and U250 mirrors because the U1 has a better 
> | setup (more SCSI controllers) I also knew that a software RAID wouldn't 
> | provide blazing performance. But should the RAID 5 writes be this bad? 
> | Worse than the U5! What would a hardware RAID 5 (like the A1000) give 
> | us? Has anyone heard when Sun's RAID card would be available for the 
> | internal disks of a U250?
> 
> You are running into a natural thing when it comes to RAID 5, especially
> RAID 5 done in software.
> 
> See, on a RAID 5 volume, the DiskSuite RAID5 driver has to calculate
> parity for every block written to the volume. This of course takes much
> more CPU time than normal, and the write write does not return (ie: is not
> fully commited) until the parity is calculated and the data written to
> disk.
> 
> The speed with mirroring is faster because you're just writing data, and
> not having to calculate any parity at all. This is why if you have the
> disk space available, a RAID 0+1 solution is more preferential than a
> RAID5 one (of course, RAID5 is better for situation where you are
> warehousing alot of data and 95% of your time is spent reading from it
> than writing to it. You'll have more disk space this way.)
> 
> This is where hardware-based RAID5 systems come into play. The typical
> RAID box has a cache of some sort (the A1000 you mentioned comes with 24MB
> from the factory) which sits between the server and the RAID controller in
> the RAID box. The server sends data over the SCSI (or FC-AL) bus to the
> RAID box, and into the cache. A processor on the other side of the cache
> takes data from the cache, calculates parity, and writes to the disk
> volume. The cache essentially acts as a buffer for writes and as a cache
> for reads, and the processor removes the overhead of parity calculation
> and volume management from the server.
> 
> 
> HTH
> /dale
> 
> _______________________________________________
> SunHELP maillist  -  SunHELP at sunhelp.org
> http://www.sunhelp.org/mailman/listinfo/sunhelp
> 


	//Magnus Abrante - Sun Service






More information about the SunHELP mailing list