Re: RAC Experiences
- From: Mogens Nørgaard <mln@xxxxxxxxxxxx>
- To: oracle-l@xxxxxxxxxxxxx
- Date: Sat, 22 May 2004 11:40:48 +0200
Hi Ryan,
The reason I bought 200 copies of James' book recently was not that it
was cheap due to the "8" in the title, but because I still find the
advise given valuable, also with respect to OPS/RAC problems.
The pings are faster when they go via the HSI, there are certain
situations where pings can be avoided due to new, smart algorithms, but
the ways to obtain good performance on multi-instance systems are still
the same: Mainly data partitioning, of course, since it reduces pinging
a lot, but please note that there are many ways of obtaining this cache
affinity thing.
The Full Disclosure Report for the 16-node RAC TPC-C benchmark shows you
how to get extreme performance out of a RAC system. Among other things,
all extents are allocated using the "alter table/cluster allocate extent
size XXX instance YYY" command. However, for some reason, in the
benchmark YYY is set to 10000, which, as far as I recall, will then map
all the extents to instance mod(10000,16). Don't know if this is a very
special and unknown feature being used, or if the benchmark guy simply
didn't know what he was doing. I'd like to know what really happened :-).
I think the 7.3 OPS manual, the 9.0.1 Deployment manual and James' book
is the very best basis one can have for understanding and managing
multi-instance databases. All the new stuff, and all the books about it,
is fine and good, especially if you have read and understood the basis
for it all, so that you don't believe this is anything new or revolutionary.
Mogens
Ryan wrote:
> Its pretty much only the federal government who can afford Sun E10Ks...
> which one is spending my money on this stuff? Those are like $10 million
> each aren't they?
>
> I thought the federal government was moving away from using 'big 5'
> consulting due to the high rates. You can get the same quality people(who
> will typically make higher salaries than the ones at the top 5 anyway...)
> for far less money.
>
> $200/hour? Less than half of that would go to the developer. What a waste of
> money. Which consulting company?
>
> BTW, what are some bad designs for RAC? I have James Morle's book, but all
> his tips are from before Cache Fusion, when Parallel Server had to force
> writes to datafiles so another node can get the block. So you had to
> partition your application.
>
>>Don Granaman wrote:
>>
>>
>>>BTW: Executive management's preferred "solution" to this was bigger
>>>hardware. Only after getting to a cluster of fully-loaded Sun E10Ks and
>
> the
>
>>>biggest Symmetrix that EMC could offer, with still poor performance, did
>>>they *really* push the $200/hour (each) outsourced (USA)
>
> designer/developers
>
>>>to change the code - and the way the system worked.
>>>
>>>-Don Granaman
>
>
> ----------------------------------------------------------------
> Please see the official ORACLE-L FAQ: http://www.orafaq.com
> ----------------------------------------------------------------
> To unsubscribe send email to: oracle-l-request@xxxxxxxxxxxxx
> put 'unsubscribe' in the subject line.
> --
> Archives are at //www.freelists.org/archives/oracle-l/
> FAQ is at //www.freelists.org/help/fom-serve/cache/1.html
> -----------------------------------------------------------------
>
----------------------------------------------------------------
Please see the official ORACLE-L FAQ: http://www.orafaq.com
----------------------------------------------------------------
To unsubscribe send email to: oracle-l-request@xxxxxxxxxxxxx
put 'unsubscribe' in the subject line.
--
Archives are at //www.freelists.org/archives/oracle-l/
FAQ is at //www.freelists.org/help/fom-serve/cache/1.html
-----------------------------------------------------------------
Other related posts:
