Re: asm on san

From: joel garry <>
Date: Fri, 12 Dec 2008 15:47:30 -0800 (PST)
Message-ID: <>

On Dec 12, 12:09pm, Mladen Gogala <> wrote:
> On Thu, 11 Dec 2008 13:16:01 -0800, joel garry wrote:
> > I have to
> > admit that it works for my employer, who not only is not doomed but
> > keeps me busy with expansion projects. Business needs uber alles and
> > all that. The only time it is a problem is when it is operating in
> > degraded mode or some rare thing overwhelms the buffering. The more
> > usual case is akin to a mostly empty municiple bus, plenty of room for
> > quite a bit more riders. I'd like to take the thanks for that, but
> > really it's just a reasonable configuration for what it does (ERP/MRP
> > OLTP, mostly). If I had to choose between Itanium hp-ux RAID-5 and
> > Itanium Windows RAID-10, I wouldn't pick the Windows.
> Joel, I used to share your bias against the Windows boxes but Windows
> Server 2008 is actually an excellent OS, with a great shell. It definitely
> has a potential. I am not sure whether there is a version of "vi", native
> perl installation and some utilities like grep, awk and find, but the OS
> holds a promise. It can definitely give Unix a run for its money.
> As for RAID 1+0 vs RAID 5 debate, RAID 5 can perform well when equipped
> with the copious amounts of NVRAM but that makes it expensive.
> Unfortunately, if I give consent to RAID 5, I will probably get a cheapo
> version with 2GB of NVRAM and an explanation from the CIO that the sales
> person told him that he doesn't need more than that. Low level RAID 1+0
> will always beat low level RAID 5, at least in my experience. If you have
> enough money for the premium HW, then it doesn't really matter. The small
> difference in price will be decided by the current ratio in memory prices
> vs. disk prices. You and me both know that RAID 5 must distribute things
> over 5 disks, using a complex and compute intensive algorithm. The block
> is not written until it's distributed across all 5 disks. That means that
> write can only as fast as the slowest of the 5 disks. Granted, ample
> memory configuration will alleviate the pain but I am unwilling to rely
> on that in case of multi TB databases where a sudden query from Crystal
> Reports or Business Objects can attempt to summarize an enormous table
> and flush your cache. My experience tells me that RAID 5 usually means
> trouble.
> --

Interesting you should mention CR, that will be implemented soon. Long ago when I was an independent contractor I pointed out these same arguments that you and Dan have made (it is my viewpoint, too), but the only trouble has been during testing, either hardware bogus enough to immediately fail or human failings. But it's not multi-TB, either. And I did mange to get at least the redo on 10, IIRC (don't have time to check just now, went through several iterations). The CR and app will be bottlenecked by IIS on W2003r2, I expect. The analytics are coming from some other MS+3rd party thing on another W2003 server, I haven't yet seen a problem on the db side, but I don't think it's being used in anger, either, yet.

When it comes down to it, Cary Millsap is right - you can't prejudge, you have to look and see. If your users don't have to wait on the spinning rust, why bother with faster rust? Besides that, sometimes PHB's guess right, even if for all the wrong reasons, hardware vendors packages sometimes make sense even if salesmen don't.

So I see 14 actual drives, looking at the thing in the rack. IIRC, one is unused, one is a hot spare, one is a parity, leaving 11 effective spindles (+IIRC the redo is on internal drives). If it were RAID-10, wouldn't it be 6 effective spindles? Somehow I don't recall anyone addressing that, though I might just not remember (or ignored a RAID-F person...). I know I've pointed out in the past there is a discontinuity in cost once you fully populate these things. You have to buy another one, and more controllers, too.

I'm under the impression spindles=performance.

I have a great bias against latest and greatest in production, whether unix, windows, or oracle, unless demonstrated necessary. But I do appreciate those who, like you and Dan, check them out and let us know.


-- is bogus.
"I just clicked and turned the f/stop. I'm certainly not a
Received on Fri Dec 12 2008 - 17:47:30 CST

Original text of this message