Wintel Big Iron
I went to a SQL Server scalability seminar at the local Microsoft office
yesterday. It was sponsored by Unisys, who was selling their 32-CPU
server consolidation solution. It was pretty cool. They've implemented
something called slot-PCs, that is a cross between VMWare and blade
computing. You have a computer on a PCI board which has it's own NIC
and CPU, but no hard drives. The hard drives are virtualized from disk
image files on the main server. They said that for many applications
they suggested using VMWare instead of or addition to their native
clustering. How much sense does that make for web farms? Slide 5
backplanes in place, copy an system image five times, you've got a
farm. Awesome.
At the seminar, they talked a lot about SQL Server clustering. It was
certainly interesting to hear about, but I'm really against clustering
the database. Although I know that situations exist which demand
clustering, I haven't personally seen instances where RAID/SANs, warm
spares and log shipping weren't a better approach from a
cost/maintenance perspective. It probably makes more sense in 2-tier
client server deployments, where you're much more likely to hose the db
tier than when using multiplexed connections via a web farm. I'd really
like to know what percentage of application failures do not result from
disk file corruption - which MS clustering is powerless to prevent.
With their shared disk architecture, any file problems will be reliably
mirrored to the failover server, which is not really what you want.
It's sad to think a company could spend 6 figures clustering and be just
as vulnerable to bits going astray on a disk.
They also mentioned the problems with doing server consolidation and
running out of drive letters (i.e. A-Z). Drive letters!???! (shakes
head and sighs)
Virtualization for the masses
Also, here's a local company hoping to bring Unisys-like virtualization
features to commodity systems:
http://www.leostream.com/
very cool.
1:14:32 PM 
|
