Matt Ingenthron wrote:
Gavin Maltby wrote:
On 08/25/05 17:34, Felix Schulte wrote:
8 SF15k running *one* OS instance?
I don't think so, I think it is some form of compute cluster.
I bet I know which customer that is.
The Sun Fire-Link had a high bandwidth, super low latency interconnect
and then through Sun Cluster you had the Remote Shared Memory APIs
(a.k.a. RSM). I recall from reading the cluster whitepaper at the time,
there was also something called xdoors for fast IPC between nodes.
Sun Cluster also brought global device trees (kinda) and global
filesystems, so while you weren't one OS instance, you had a set of
really interesting APIs you could do some serious damage with using Sun
Fire-Link. Cluster also supported the same APIs over cheaper
interconnects.
Sadly, I think the only major use by a complicated commercial program
was Oracle in 9i RAC (they called it "cache fusion", but a search for
that makes it seem like a retread term-- there was a whitepaper
somewhere once), and they eschewed that in favor of a more
cross-platform (and a more pure Oracle software) approach. Now they are
their own volume manager, etc.
I was on the review team for the cluster 3.0 blueprint book,
in which this feature was covered -- it did indeed look very
interesting but sadly mostly unused in a direct sense. There
aren't all that many customers out there running RAC with
more than 2 nodes and SunFireLink.
I believe it to be the case that Oracle did take the knowledge
they gained with cache fusion and have made it work better in
10g. However I'm not too impressed with the clustering (CRS)
stuff -- then again, it's a 1.0 product along with their storage
manager (ASM). Time will tell. Fortunately for Sun, our SunCluster
engineering teams have a head start and are fantastically well
endowed with talent and skill :)
James C. McPherson
--
Pacrim PTS Engineer 828 Pacific Highway
Gordon NSW
Sun Microsystems Australia 2072
_______________________________________________
opensolaris-discuss mailing list
[email protected]
