Re: raid10 hard disk choice

On 5/21/09 3:05 PM, "Robert Haas" <robertmhaas(at)gmail(dot)com> wrote:

> On Thu, May 21, 2009 at 5:41 PM, Scott Marlowe <scott(dot)marlowe(at)gmail(dot)com>
> wrote:
>> On Thu, May 21, 2009 at 2:29 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>>> On Thu, May 21, 2009 at 8:59 AM, Matthew Wakeling <matthew(at)flymine(dot)org>
>>> wrote:
>>>> On Thu, 21 May 2009, Linos wrote:
>>>>>
>>>>>        i have to buy a new server and in the budget i have (small) i have
>>>>> to select one of this two options:
>>>>>
>>>>> -4 sas 146gb 15k rpm raid10.
>>>>> -8 sas 146gb 10k rpm raid10.
>>>>
>>>> It depends what you are doing. I think in most situations, the second
>>>> option
>>>> is better, but there may be a few situations where the reverse is true.
>>>
>>> One possible case of this - I believe that 15K drives will allow you
>>> to commit ~250 times per second (15K/60) vs. ~166 times per second
>>> (10K/60).  If you have a lot of small write transactions, this might
>>> be an issue.
>>
>> But in a RAID-10 you aggreate pairs like RAID-0, so you could write
>> 250(n/2) times per second on 15k where n=4 and 166(n/2) for 10k drives
>> where n=8.  So 500 versus 664... ?  Or am I getting it wrong.
>
> Well, that would be true if every write used a different disk, but I
> don't think that will be the case in practice. The WAL writes are
> very small, so often you'll have multiple writes even to the same
> block. But even if they're to different blocks they're likely to be
> in the same RAID stripe.

Disk count and stripe size don't have much to do with it, the write cache
merges write requests and the client (the wal log write) doesn't have to
wait on anything. The RAID card can merge and order the writes, so it can
go nearly at sequential transfer rate, limited more by other concurrent
pressure on the raid card's cache than anything else.

Since WAL log requests are sequential (but small) this provides huge gains
and a large multiplier over the raw iops of the drive.

>
> ...Robert
>
> --
> Sent via pgsql-performance mailing list (pgsql-performance(at)postgresql(dot)org)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-performance
>