Re: AMM experiences
From: Matthias Hoys <matthias.hoys_at_gmail.com>
Date: Wed, 11 Apr 2012 06:28:05 -0700 (PDT)
Message-ID: <4660023.887.1334150885767.JavaMail.geo-discussion-forums_at_vbex14>
On Tuesday, April 10, 2012 5:48:44 PM UTC+2, Mladen Gogala wrote:
> I am reluctant to set AMM on one of my production databases. The machine
> is Dell, 4 quad-core CPU's and 64GB of RAM, 64 bit RH EL 5.8. I was
> allocating a hefty 32GB SGA, with PGA_AGGREGATE_TARGET set to 20GB. 32GB
> for SGA was allocated by defining huge pages.
> Now, my manager challenged me by mentioning memory_max_size and
> memory_target. My answer was that PGA and SGA have very different paging
> characteristics and that they shouldn't be lumped together. I also
> mentioned that the page table for the 32GB shared memory segment that is
> SGA is a whopping 64MB, if 4KB pages are used, which makes it quite likely
> that page tables will be swapped out, sometimes in the future, with
> predictably disastrous consequences for the speed of the system.
> Normally, these two arguments should suffice, but I am now required to
> "show the numbers", which is not at all hard to do. I am organizing a
> little benchmark using swingbench, with 2 types of queries and I am
> configuring it in such a way that the memory will be close to exhaustion.
> I don't have much doubt about the outcome of the benchmark, but I would
> like to hear other experiences.
> Has anybody used AMM on a large box (see the description above) and what
> were the results and impressions? I must confess that AMM looks to me like
> one of those DBA 2.0 features, intended for small to medium instances, not
> large monsters with 32GB SGA size.
> I wouldn't like my SGA to be paged out and I wouldn't like my sort/hash
> work areas in memory not to be paged out at some point. Please, feel free
> to chime in, except by trying to advertise blogs or "sexy chats".
>
>
>
>
> --
> http://mgogala.freehostia.com
Date: Wed, 11 Apr 2012 06:28:05 -0700 (PDT)
Message-ID: <4660023.887.1334150885767.JavaMail.geo-discussion-forums_at_vbex14>
On Tuesday, April 10, 2012 5:48:44 PM UTC+2, Mladen Gogala wrote:
> I am reluctant to set AMM on one of my production databases. The machine
> is Dell, 4 quad-core CPU's and 64GB of RAM, 64 bit RH EL 5.8. I was
> allocating a hefty 32GB SGA, with PGA_AGGREGATE_TARGET set to 20GB. 32GB
> for SGA was allocated by defining huge pages.
> Now, my manager challenged me by mentioning memory_max_size and
> memory_target. My answer was that PGA and SGA have very different paging
> characteristics and that they shouldn't be lumped together. I also
> mentioned that the page table for the 32GB shared memory segment that is
> SGA is a whopping 64MB, if 4KB pages are used, which makes it quite likely
> that page tables will be swapped out, sometimes in the future, with
> predictably disastrous consequences for the speed of the system.
> Normally, these two arguments should suffice, but I am now required to
> "show the numbers", which is not at all hard to do. I am organizing a
> little benchmark using swingbench, with 2 types of queries and I am
> configuring it in such a way that the memory will be close to exhaustion.
> I don't have much doubt about the outcome of the benchmark, but I would
> like to hear other experiences.
> Has anybody used AMM on a large box (see the description above) and what
> were the results and impressions? I must confess that AMM looks to me like
> one of those DBA 2.0 features, intended for small to medium instances, not
> large monsters with 32GB SGA size.
> I wouldn't like my SGA to be paged out and I wouldn't like my sort/hash
> work areas in memory not to be paged out at some point. Please, feel free
> to chime in, except by trying to advertise blogs or "sexy chats".
>
>
>
>
> --
> http://mgogala.freehostia.com
There's an article on MOS stating that AMM is incompatible with huge pages: document ID 749851.1:
"With AMM all SGA memory is allocated by creating files under /dev/shm. When Oracle DB does SGA allocations that way HugePages are not reserved/used. The use of AMM is absolutely incompatible with HugePages."
Matthias Received on Wed Apr 11 2012 - 08:28:05 CDT