[rescue] Agan, I seek the wisdom of the Collective ...

Skeezics Boondoggle skeezics at q7.com
Fri Aug 12 18:17:22 CDT 2005

On Fri, 12 Aug 2005, Patrick Finnegan wrote:

> On Friday 12 August 2005 15:58, Phil Stracchino wrote:
> > Speculation, on which I seek advice and experience:  Is my problem
> > simply that a single USII-250 isn't enough CPU for RAID5 on that
> > hardware?  It's run a six-drive RAID0 for about four years without
> > once falling down like this.
> That's very unlikely, unless the RAID layer on Solaris sucks a lot more 
> than I think it does.  You shouldn't have any trouble running the same 
> RAID array on a 70MHz SPARC5, you just might get slower rebuild times.

My experience is that RAID5 under DiskSuite[0] (or whatever the hell it's
called these days) is just like OpenWindows - it *must* have timing loops
inside so that it runs at the same speed regardless of the hardware.  I'm 
not kidding... 

In all these configurations (and more):

12-way 85Mhz SC2000E w/8 fast/narrow SCSI buses, 12x 5400RPM drives
4-way 125Mhz SS20 w/2 fast/wide SCSI buses, 8x 7200RPM drives
8-way 400Mhz E4500 w/2 FC-AL attachments, 6x 10000RPM drives
4-way 450Mhz E420R w/2 wide/Ultra2 SCSI buses, 8x 15000RPM drives

I've never seen more than ~5-10MB/sec write speeds, ever, using software 
RAID5.  Seriously.  And the 2000E and SS20 configs had Prestoserve NVRAM 
to help with write coalescing.

On those same configurations with software RAID0+1 you can easily get 3-4x
the write throughput, and in all cases the DiskSuite read performance is
pretty good, actually... but if you can't afford the increased
cost/power/space/noise/heat of running double the drives and need to use
RAID5, you might consider a cheap NetApp (their RAID4 and WAFL filesystem
implementation kick major ass) or just get a hardware RAID5 controller.  
'Cuz the software implementation is painfully slow.

I'd be very surprised if the current ODS/SDS/VolumeMgr/WTFEVER that's
"rolled in" (finally) with Solaris 10 isn't fully threaded; I mean, the
way you get performance out of Suns is through lots of threads in a "long
and wide" kind of application... is the code released through OpenSolaris
yet?  Might be worth checking out. :-)

-- Chris

[0] Another Sun product that suffers from Sun Renaming Syndrome, where it
seems that jackasses in marketing are compelled to rename the damned thing
every three weeks, like the compilers, the management tools, the
zones/containers/N-this,ONE-that,Java-everything, etc.  Argh!

More information about the rescue mailing list