On 6/3/2010 5:06 AM, Edward Martinez wrote:
03.06.2010 14:01, Edward Martinez пишет:
On 06/ 1/10 11:59 PM, Edward Martinez wrote:
I just read AMD opterons and Linux is powering
the
worlds fastest supercomputer. If the x86 platform
and
Linux now has the capacity to produce this type
of
results, where does this leave Power and SPARC
platfroms?
http://www.marketwatch.com/story/amd-opterontm-process
or-again-dominates-top500-2010-05-31?reflink=MW_news_s
tmp
The graphics here are interesting:
http://news.bbc.co.uk/2/hi/technology/10187248.stm
--
Ian.
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
They are indeed interesting, but what I find more
interesting is that there are more X86 platforms
running linux then there are POWER platform running
AIX and SPARC platforms running Linux/Solaris. when
looked under "By Processor" from the top tabs" SO,
this makes me ask if POWER and SPARC platforms
running AIX and SOLARIS are more superior then x86
platform running Linux, then why are they not being
used more to build supercomputers?
http://news.bbc.co.uk/2/hi/technology/10187248.stm
elieve reason is simple it is acceptable
performace/price ratio for
x86 hardware with regards to math calculations. You
would not need all
this visualization support provided by POWER and
SPARC hardware and
corresponding to them operating system for HPC. So
why pay for something
you wouldn't need?
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
I'm beginning to get it.
Here are few quotes from article
quote:
The announcement of high-end servers by IBM and Hewlett-Packard this week won't
halt declining Unix server sales as the onslaught of x86 servers continues,
analysts said on Tuesday.
Once again, they're missing the point - volume sales of UNIX systems are
in a slow decline (as are systems like Mainframes). However, per-sale
*revenue* is generally up (Q1Y2010 was an anomaly, as discussed before).
Where UNIX systems are losing to x64 is the same place they lost to
Windows a decade ago: workstations and department/workgroup servers,
plus the add-on retirement of web servers. They're not losing in the
back-end server segment at all.
quote:
But even the new chips will have little effect on reviving the declining sales
of Unix servers, analysts said. Customers are increasingly opting for servers
based on x86 chips, which are getting more powerful and entering markets
traditionally dominated by Unix servers.
No, they're really NOT. The transition from SPARC/POWER/Alpha/etc to
x86 for smaller servers started over a decade an a half ago. But,
fundamentally, even x64 hasn't made much of any dent in the Big Iron
UNIX sales, just as UNIX systems can't make much of a dent in the
Mainframe market. Sure, nobody buys a 2-CPU Sparc or Power system
anymore, but that's been true for almost a decade. Nothing new there
(and, darned little change in sales outcome).
I don't see 50-way x64 boxes. I don't see x64 boxes with all the nice
RAS features of a SPARC or POWER system. I don't see financial
institutions replacing their NonStop boxes with x64. And, x64 still
has a horrible performance/watt compared to other designs.
UNIX systems lost the small server market in the 1990s. More recently,
it has been shown that x64 is sufficient for massively-scalable cluster
systems. But they're holding on strong in their back-end server arena,
just like Mainframes are holding onto their batch-processing crown. And
the T-series SPARC chips showed that certain classes of problems that
were "owned" by the x64 weren't so dominated as everyone though...
quote:
A lot of customers are switching to x86 servers because of lower hardware and
software costs attached to acquiring and maintaining the systems, said Jim
McGregor, technology strategist at In-Stat.
http://www.pcworld.com/businesscenter/article/188951/ibm_hp_servers_wont_stop_x86_onslaught_on_unix.html
I guess this would also apply to supercomputers?
Nope.
The reason Linux/x64 dominates "supercomputers" these days is that we've
re-defined what a supercomputer is. Supercomputing *used* to be
considered single-image systems which were amazingly fast at serial
computions (think all the CRAY systems), or had enormous parallel
throughput (ThinkingMachines stuff). However, in the mid-90s, it was
discovered that many modern problems are analyzable by massive
parallel-processing. NASA's Beowulf system was a proof-of-concept that
you could stitch large numbers of low-powered clients into a giant
parallel cluster. But's they're NOT single-image systems, and they're
completely unsuitable for certain classes of problems. Nevertheless,
they are *very* useful for other classes of problems.
So, we've broadened the "supercomputer" category to include this new
class of problems, which is dominated in cost/performance by x64.
Also, acquisition costs for hardware are but a fraction of the TCO.
Also, if that analyst thinks that it's cheaper to buy Windows 2008
server or Redhat Enterprise Linux than to get an IBM or Oracle service
contract for AIX/Solaris, well, he's in for a rude surprise. Not to
mention that in many areas, it actually is SIGNIFICANTLY higher cost to
purchase x64 over the long-haul.
Ever wonder why Mainframes are still around, even though a modern
desktop PC has more computational power than a room-sized mainframe?
Turns out that raw computation isn't the end-all-be-all of the computing.
As usual, people ignore ownership costs, and look only at acquisition
costs. x64 is cheaper generally only if you forgo any enterprise-class
OS or ongoing support. Which is fine as long as either (a) you're HUGE,
and can provide your own software/support (ala Goggle) or (b) don't care
about your data or systems (which, fundamentally, boils down to small
businesses - they don't understand how much their data means to them, so
they undervalue it, and consequently make purchasing decisions as if
data security/protection/reliability didn't matter).
I think it's also the reason behind Micrsoft decision to drop Itanium support
Itanium was a failure from the git-go. It had too many problems, and no
real path to success. Realistically, the AMD x86_64 architecture
introduction in 2001 signaled the death-knell for Itanium, which has
never recovered (or, more realistically, never really made it out of the
gate).
quote:
Why the change? The natural evolution of the x86 64-bit (’x64′) architecture
has led to the creation of processors and servers which deliver the scalability
and reliability needed for today’s ‘mission-critical’ workloads,’” [Dan] Reger
wrote. “Just this week, both Intel and AMD have released new high core-count
processors, and servers with eight or more x64 processors have now been
announced by a full dozen server manufacturers. Such servers contain 64 to 96
processor cores, with more on the horizon.”
http://insidehpc.com/2010/04/06/microsoft-drops-itanium-support/
I use the latest AMD and Intel systems everyday - we're on their
prototype programs. They're nice - the new Nehalem-EX and Mangy-Cours
architectures are heavily threaded, many-cored systems with lots of
speed and reasonable features.
But, obviously, Mr Reger doesn't understand what "scales" means. Or
"reliability".
It turns out that it's often far cheaper to put all your eggs in one
basket, after making sure that it's a *really good* basket.
The problem with these new x64 systems is that they:
(1) are only just beginning to have the RAS functionality of existing
UNIX systems, which is actually far more important that raw performance
for many market segments
(2) really only have Solaris as an available Enterprise-ready OS to run
on them. (which, is good for Oracle. :-)
The reason why UNIX and Mainframes are safe in their core markets boils
down to design decisions: Mainframes go for raw throughput and
incredible reliability, UNIX systems go for flexibility/high
reliability/managability, and x64 goes for raw performance. Each has
their own niche (and people that value each differently). UNIX "lost"
market segments when it was no longer the performance king, and those
segments didn't care about the other features of UNIX. But x64 has made
little headway in the segments where performance isn't the only metric.
--
Erik Trimble
Java System Support
Mailstop: usca22-123
Phone: x17195
Santa Clara, CA
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
