Re: "big" machines running Debian?
- To: debian-amd64@lists.debian.org
- Subject: Re: "big" machines running Debian?
- From: Goswin von Brederlow <goswin-v-b@web.de>
- Date: Mon, 02 Mar 2009 14:28:04 +0100
- Message-id: <87fxhwjbq3.fsf@frosties.localdomain>
- In-reply-to: <20090301213501.GA3977@samad.com.au> (Alex Samad's message of "Mon, 2 Mar 2009 08:35:01 +1100")
- References: <20090221130032.GX6748@tamarapatino.org> <52d26d930902210535w4262f10bn3107b58b7243b262@mail.gmail.com> <87d4d6eer9.fsf@frosties.localdomain> <20090225215411.GK3212@samad.com.au> <20090225220630.GX23244@csclub.uwaterloo.ca> <20090226003620.GA26041@samad.com.au> <87myc70wu9.fsf@frosties.localdomain> <20090301213501.GA3977@samad.com.au>
Alex Samad <alex@samad.com.au> writes:
> On Sat, Feb 28, 2009 at 09:50:06AM +0100, Goswin von Brederlow wrote:
>> Alex Samad <alex@samad.com.au> writes:
>>
>
> [snip]
>
>> > true, depends on whos rule of thumb you use. I have seen places where
>> > mandate fc drives only in the data center - get very expensive when you
>> > want lots of disk space.
>>
>> The only argument I see for FC is a switched sorage network. As soon
>> as you dedicate a storage box to one (or two) servers there is really
>> no point in FC. Just use a SAS box with SATA disks inside. It is a)
>> faster, b) simpler, c) works better and d) costs a fraction.
>
> The problem I have seen is the person who controls the purse strings
> doesn't always have have the best technological mind. There was a while
> back where have fibre meant fibre to the disk. So managers wanted fibre
> to the disk, so they paid for fibre to the disk.
And now they have to learn that we have new technologies. New
requirements and new solutions. What was good 5 years ago isn't
neccessarily good today. Saddly enough a lot of purse strings seem to
be made of stone and only move in geological timespans. :)
>> And hey, we are talking big disks space for a single system here. Not
>> sharing one 16TB raid box with 100 hosts.
>>
>> > Also the disk space might not be need for feeding across the network, db
>> > aren't the only thing that chew through disk space.
>> >
>> > the op did specific enterprise, I was think very large enterprise, the
>> > sort of people who mandate scsi or sas only drives in their data centre
>>
>> They have way to much money and not enough brain.
>
> I would have to dissagree, some times the guidelines that you set for
> your data storage network mandate having the reliability (or the
> performance) of scsi (or now sas), they could be valid business
> requirements.
Could be. If you build storage for a DB you want SAS disks and
raid1. If you build a petabyte storage cluster for files >1GB then you
rather want 3 times as many SATA disks. An XYZ only rule will always
be bad for some use cases.
> Traditionally scsi drives had a longer warranty period, were meant to be
> of better build that cheap ata (sata) drives.
>
> Although this line is getting blurred a bit.
There surely is a difference between a 24/7, 5 year warranty, server
SCSI disk and a cheap home use SATA disk. But then again there are
also 24/7, 5 year warranty, server SATA disks.
I don't think there is any quality difference anymore between the scsi
and sata server disks.
> Unless we talk about a specific situation, storage as other areas of IT
> are very fluid, and there are many solutions to each problem.
Exactly.
> Look at the big data centers of google and such that use pizza box's
> machine dies who cares its clustered and they will get around to fixing
> it at some point. to 4-8 nods clusters of oracle that are just about
> maxed out, one server goes down and ....
Same here. Nobody builds HA into a HPC cluster. If a node fails the
cluster runs with less node. Big deal.
Saddly enough for storage there is a distinct lack of
software/filesystems that can work with such a lax reliability. With
the growing space requirements and stalling size increase in disk size
there are more and more components in a storage cluster. I feel that
redundancy has to move to a higher level. Away from the disk level
where you have raid and towards true distributed redundancy across the
storage cluster as a whole.
MfG
Goswin
Reply to: