IDS Forum

Re: Large stripe sizes

Posted By: darren_jacobs@carmax.com
Date:

A response from our unix admin on the striping question:

If you need clarification on anything please feel free to ask.



----- Forwarded by Darren Jacobs/7001/Carmax on 01/21/04 10:49 AM -----

Tom Rectenwald
To: Darren Jacobs/7001/Carmax@Carmax
01/21/04 10:47 AM cc:
Subject: Re: Large stripe sizes [2463](Document link: Darren Jacobs)




My thoughts on it are:

Striping used to matter. It still does to a small degree, but not so much
anymore. EMC arrays and things like the XP512 push everything into a
cache, shared memory, ACPs, Intel CHIP processors, and finally into RAID
groups (think MAGS).

The end result is that the data is already balanced every which way
imaginable. We carve out data in logical units (LUNS) which have virtually
no resemblance left to the actual physical disk.

In other words, what used to be done with striping is now handled by the
internal logic of modern disk arrays. I stripe data on the XP512 to:

a) load balance across multiple fibre interfaces (in the case of Rolls)
b) load balance amongst CHIP processors (the entry point to the array)

However, it really isn't striping in terms of the old days where you could
count spindles, etc. Those days are rapidly going away. It leaves a
difficult environment because striping and spindle counts used to be an
excellent method of gaining system performance, especially with RDBM's. It
just isn't like that much more, and focusing on striping on things like the
XP512 does nothing but add complication.

As for 1MB across 100+GB spindles, we tried that on the old EMC 3500 back
in the day. End result: literally a week of wasted sys. admin. time (Jim
Camp was doing it), and no performance increase at all. In fact, I think
things slowed down a bit because it drove the LVM daemons into overdrive on
the OS level.

Counting spindle days are rapidly fading. In fact, IMHO not allowing data
to cross MAGs on the XP512 hurts performance slightly because you cannot
balance data across ACPs (though this is probably not noticeable), and,
more importantly, makes administration of the array very difficult and
complicated. Think of the recent expansion of the /psoft filesystem. We
had to take it offline to accomplish the extension. If we weren't stuck on
the same MAGs it could have been done online with no downtime.

Modern technology reduces complication when used appropriately. Otherwise,
when things like Business Copy, MAG sharing, and just letting the logic in
the hardware do what we paid it to do take a back seat, downtime and wasted
time incure.

Time to step off of my soap box and get more coffee.






Darren Jacobs
To: Tom Rectenwald/7001/Carmax@Carmax
01/21/2004 10:19 cc:
AM Subject: Re: Large stripe sizes [2463]





Your a disk expert.....What's you feelings on the striping question
below....sounds like its a little flaky going with a 1 meg strip across 4
spindles at 181 gig.
----- Forwarded by Darren Jacobs/7001/Carmax on 01/21/04 10:17 AM -----

"Jack Parker"
<vze2qjg5@verizon To: ids@iiug.org
.net> cc:
Sent by: Subject: Re: Large stripe sizes [2463]
forum.subscriber@
iiug.org


01/20/04 11:32 PM







I see you don't have a lot of responses. So I'll pitch in.

Some time ago I was driven to the striping thing by managerial decree on a
1.2TB warehouse. I worked with the sysadmin to test out different stripe
sizes vs DBA controlled spindle allocation. The end results were
inconclusive - we got roughly the same performance out of the disks no
matter what striping size we used vs direct spindle control by the DBA. We
wound up with 4 spindles and I forget what size stripes on an EMC SAN.
Mind
you what we were doing with that disk was sequential scans - NOT OLTP. It
only took us 2 or 3 days to try out the different mixes - I would recommend
you do the same. Develop a benchmark, and then slice off some disk and try
it 12 ways to sunday.

cheers
j.

----- Original Message -----
From: "geoff.poole...." <geoff.poole@verizon.com>
To: <ids@iiug.org>
Sent: Tuesday, January 20, 2004 12:17 PM
Subject: Large stripe sizes [2462]


> I'm looking for everyone's experience rather than theory, we are about to
> be swallowed by a corporate (Oracle/DB2 driven) standard of 1M stripe
size
> across 4 spindles (fragments of 181GB drives). Has anyone done something
> similar on any hardware ? We have every version of Informix with huge
and
> small databases alike, most platforms (Sun/HP/AIX) and mostly EMC. We are
> about to convert our biggest OLTP system (batch at night) from contiguous
> storage (18GB drives with 8GB storage cache) to 1Meg stripe (newer 181GB
> drives with 32GB storage) we just was wondering whether there is anyone
out
> there that's done something similar ? Does anyone know of a white paper
on
> this subject even if it was written from a different DB ?
>
> Geoff Poole
> Verizon
>
>
>
>

Messages In This Thread

• Large stripe sizes
  geoff.poole@verizon.com -- Tuesday, 20 January 2004, at 3:17 p.m.
  • Re: Large stripe sizes
    Jack Parker -- Tuesday, 20 January 2004, at 11:32 p.m.
    • Re: Large stripe sizes
      darren_jacobs@carmax.com -- Wednesday, 21 January 2004, at 10:56 a.m.

IDS Forum is maintained by Administrator with WebBBS 5.12.