From: John A Meinel <john@arbash-meinel.com>
To: Dan Harris <fbsd@drivefaster.net>
Cc: pgsql-performance@postgresql.org
Subject: Re: Table locking problems?
Date: Tue, 09 Aug 2005 17:30:17 -0500
Message-ID: <42F92E79.5050905@arbash-meinel.com> (raw)
In-Reply-To: <17797211-4C5F-4410-B38E-53ECC10D0752@drivefaster.net>
References: <52CFA236-FC43-4018-B399-1854240110FF@drivefaster.net>
<1123656547.13362.43.camel@amd64-laptop-spoe>
<2764B744-4CDC-4F2C-99C4-2948D37A9D23@drivefaster.net>
<42F9255A.4080604@arbash-meinel.com>
<17797211-4C5F-4410-B38E-53ECC10D0752@drivefaster.net>
Dan Harris wrote:
>
> On Aug 9, 2005, at 3:51 PM, John A Meinel wrote:
>
>> Dan Harris wrote:
>>
>>> On Aug 10, 2005, at 12:49 AM, Steve Poe wrote:
>>>
>>>> Dan,
>>>>
>>>> Do you mean you did RAID 1 + 0 (RAID 10) or RAID 0 + 1? Just a
>>>> clarification, since RAID 0 is still a single-point of failure even if
>>>> RAID1 is on top of RAID0.
>>>>
>>> Well, you tell me if I stated incorrectly. There are two raid
>>> enclosures with 7 drives in each. Each is on its own bus on a dual-
>>> channel controller. Each box has a stripe across its drives and
>>> the enclosures are mirrors of each other. I understand the
>>> controller could be a single point of failure, but I'm not sure I
>>> understand your concern about the RAID structure itself.
>>>
>>
>> In this configuration, if you have a drive fail on both controllers,
>> the entire RAID dies. Lets label them A1-7, B1-7, because you stripe
>> within a set, if a single one of A dies, and a single one of B dies,
>> you have lost your entire mirror.
>>
>> The correct way of doing it, is to have A1 be a mirror of B1, and
>> then stripe above that. Since you are using 2 7-disk enclosures, I'm
>> not sure how you can do it well, since it is not an even number of
>> disks. Though if you are using software RAID, there should be no
>> problem.
>>
>> The difference is that in this scenario, *all* of the A drives can
>> die, and you haven't lost any data. The only thing you can't lose is
>> a matched pair (eg losing both A1 and B1 will cause complete data loss)
>>
>> I believe the correct notation for this last form is RAID 1 + 0
>> (RAID10) since you have a set of RAID1 drives, with a RAID0 on-top of
>> them.
>>
>
> I have read up on the difference now. I don't understand why it's a
> "single point of failure". Technically any array could be a "single
> point" depending on your level of abstraction. In retrospect, I
> probably should have gone 8 drives in each and used RAID 10 instead for
> the better fault-tolerance, but it's online now and will require some
> planning to see if I want to reconfigure that in the future. I wish
> HP's engineer would have promoted that method instead of 0+1..
I wouldn't say that it is a single point of failure, but I *can* say
that it is much more likely to fail. (2 drives rather than on average n
drives)
If your devices will hold 8 drives, you could simply do 1 8-drive, and
one 6-drive. And then do RAID1 with pairs, and RAID0 across the
resultant 7 RAID1 sets.
I'm really surprised that someone promoted RAID 0+1 over RAID10. I think
I've heard that there is a possible slight performance improvement, but
really the failure mode makes it a poor tradeoff.
John
=:->
>
> -Dan
>
Attachments:
signature.asc (application/pgp-signature, 252B) download
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-performance@postgresql.org
Cc: john@arbash-meinel.com, fbsd@drivefaster.net
Subject: Re: Table locking problems?
In-Reply-To: <42F92E79.5050905@arbash-meinel.com>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox