In-house Hosting Options

Peter Bowen peter at bowenfamily.org
Fri Mar 4 11:39:43 MST 2005


Josh Coates wrote:

>>L - O - L!
>>    
>>
>
>heh - you shouldn't laugh so loud - because i think you may be way off on
>your perceptions about shared-nothing pc clusters.
>
>first off, watts is watts, and the power bill on your shared-nothing cluster
>is about the same as your smp.
>  
>
I was pointing out that the super computers are overkill for my family 
server... the LOL was prefacing a joke!  Lighten up! 

>>Same thing - sure a cluster of 300 PCs is computationally faster, than an
>>    
>>
>E10K,but I stand by my statement... every job has a tool.
>
>yep, every job has it's tool - but the point was that the shared-nothing
>cluster continues to prove itself a better too than an SMP.  i bet the
>reason you run an E10K is because it's much easier for your particular
>application to run w/ shared memory - and i bet that's the only reason.  you
>probably overpaid and overprovisioned and are now attached to a single
>vendor because your app wasnt engineered to scale.
>
>  
>
I'm not running an E10K - If I needed one, I would use one. (Although I 
don't anticipate that anytime soon)   If I needed a PC cluster, I would 
run one of those.  (I want to set one of tose up) Every job has a tool.  
Sometimes I run MySQL, sometimes I run Postgres, and sometimes I run 
Oracle.  Are you seeing a theme.  There is no magic bullet EVER - but 
after looking at your site, it looks like I'm preaching to the choir here.

>>I/O was my point not processing power.
>>    
>>
>
>TFLOPS is just an over-simplified way of measuring power - IOPS is another.
>i assume you are thinking of network IO?  ever done benchmarks SUN vs PC w/
>for an HTTP proxy?  think slashdot or google uses an E10K?  i assure you,
>shared-nothing clusters open a can of whoop-you-know-what on SMPs for
>embarrassingly parallel IO.
>
>and if you are talking about disk IO, seeks are constant, and bus and memory
>management kicks in, and smp v shared-nothing both have their pros & cons
>there (but you should take a gander at SPECSFS benchmarks - you will note
>that PCs dominate for disk IO as well...)
>
>i suggest you do some homework - and consider the possibility that your big
>iron is more akin to the dinosaurs than you think, or at least get educated
>about shared-nothing pc clusters.  nothing personal, and i dont mean to be
>obnoxious - i just want to keep the record straight.
>  
>
A 2x sun box vs a 2x PC has in my experience gone to the sun box.  It's 
hard to compare a 4x sun box since it's hard to find a 4x PC.  There are 
no magic bullets and to say that one architecture is in every way better 
is absolutely wrong - - Unless of course,  you're talking to MS because 
they have documented proof of how much better and less expensive they 
are.  (Sarcasm)

I was objecting to the blanket statement the Sun is always wrong.  I was 
NOT saying that Sun is always right... and many times we all forget to 
perscope up and have a look around and make sure that we're really 
running on the right stuff.  It's like the start of this thread... Eric 
asked a great question, and we're all better off for it.

-Peter

>Josh Coates
>http://www.jcoates.org
>
>-----Original Message-----
>From: plug-bounces at plug.org [mailto:plug-bounces at plug.org]On Behalf Of
>Peter Bowen
>Sent: Friday, March 04, 2005 10:27 AM
>To: Provo Linux Users Group Mailing List
>Subject: Re: In-house Hosting Options
>
>
>L - O - L  I wonder how many users my fam server would support with 70
>Tflops of computing power.  I think with 32K procs I would have to
>definately co-locate that baby.  I'll bet that the monthly power bill
>would be more than any computer I'll ever own.  L - O - L!
>
>    BTW - I/O was my point not processing power.  When I was a kid...
>pre-pc we used to go to Radio Shack and load the cheapo computer with a
>basic script that counted to 10K.  Well the cheapo computer could count
>faster than the expensive computer.  So it was faster right.... well at
>counting it was, but it wasn't as powerful with actual math and
>calculation and it had a ton less memory.  Same thing - sure a cluster
>of 300 PCs is computationally faster, than an E10K,but I stand by my
>statement... every job has a tool.
>
>-Peter
>Josh Coates wrote:
>
>  
>
>>>Like Sun hardware - come now... High end sun hardware kills PCs
>>>
>>>
>>>      
>>>
>>right - that's why there are so many high-end sun machines in the top500
>>supercomputers (top500.org).  oh wait, um...there are hardly any suns
>>listed.
>>
>>hey - wait a second..what's this?  a whole bunch of PC clusters in the
>>top500?  gee, that's weird...  ;-)
>>
>></sarcasm>
>>
>>the big-iron SMP vs shared-nothing cluster debate rages on, but high-end
>>    
>>
>sun
>  
>
>>hardware, along with all the other monolithic smp machines have been
>>trailing on price/performance for quite a while now.  shared-nothing pc
>>clusters simply dominate raw performance as well as price/performance - not
>>to mention ease of provisioning.
>>
>>Josh Coates
>>http://www.jcoates.org
>>
>>-----Original Message-----
>>From: plug-bounces at plug.org [mailto:plug-bounces at plug.org]On Behalf Of
>>Peter Bowen
>>Sent: Friday, March 04, 2005 9:07 AM
>>To: Provo Linux Users Group Mailing List
>>Subject: Re: In-house Hosting Options
>>
>>
>>Sasha Pachev wrote:
>>
>>
>>
>>    
>>
>>>>So I take that's what the advantage of T1s are?  I never understood
>>>>
>>>>
>>>>        
>>>>
>>>that either as DSL can go faster then a T1 and significantly cheaper,
>>>but I haven't seen a way to combine the bandwidth for one IP.  I've
>>>also seen DC3 connections from a few ISPs that are up to 20,000 a
>>>month.  But still at a very poor bandwidth.  I just don't understand
>>>what makes these different connections so special.
>>>
>>>
>>>Perceived value :-) Kind of like Sun hardware. Sometimes people feel
>>>more comfortable when they have paid more, and the market is quite to
>>>respond for the need for this warm and cozy feeling.
>>>
>>>For the bandwidth solution - the first thing that comes to mind. If
>>>collocation is too expensive, get as many cheap 1.5 MBit/s lines as
>>>needed with different IP addresses and use iptables + round-robin DNS.
>>>
>>>Near-perfect uptime is expensive, the closer you try to get to
>>>perfect, and in many cases overvalued. I would venture to say that for
>>>a regular web application, if your site beats the reliability of their
>>>desktop , most of your clients will be satisfied. Once you reach a
>>>certain point, it is wise to spend your resources on the things that
>>>matter more especially when those resources are limited.
>>>
>>>
>>>
>>>      
>>>
>>Like Sun hardware - come now... High end sun hardware kills PCs - On the
>>other hand who wants to run the family site on an E10K. :)  Every tool
>>has a job.
>>
>>-Peter
>>.===================================.
>>| This has been a P.L.U.G. mailing. |
>>|      Don't Fear the Penguin.      |
>>|  IRC: #utah at irc.freenode.net   |
>>`==================================='
>>
>>.===================================.
>>| This has been a P.L.U.G. mailing. |
>>|      Don't Fear the Penguin.      |
>>|  IRC: #utah at irc.freenode.net   |
>>`==================================='
>>
>>
>>    
>>
>
>.===================================.
>| This has been a P.L.U.G. mailing. |
>|      Don't Fear the Penguin.      |
>|  IRC: #utah at irc.freenode.net   |
>`==================================='
>
>.===================================.
>| This has been a P.L.U.G. mailing. |
>|      Don't Fear the Penguin.      |
>|  IRC: #utah at irc.freenode.net   |
>`==================================='
>  
>




More information about the PLUG mailing list