Re: Numa Config Effekt in 19c?

On these large 8+ socket machines, you'd very likely want to have NUMA awareness enabled, at very least at OS level, but likely at Oracle level too (if you're running a large instance that has to span all these nodes and can't separate/isolate separate instances to individual nodes with local memory access). But in Oracle you might hit some additional bugs coming from NUMA-awareness complexity.

--
Tanel Poder
https://learn.tanelpoder.com


On Fri, Jun 24, 2022 at 1:40 PM Mladen Gogala <gogala.mladen_at_gmail.com>
wrote:



> On 6/23/22 15:30, Tanel Poder wrote:

>

> If you have a modern 2-socket server (where remote memory access latency


> is not multiple *times* higher than local), keep NUMA awareness off at OS

> and Oracle level). But, yeah with some (old?) 8 socket server, the

> multi-hop remote access latency might be so slow that you need either NUMA

> placement or try to constrain your entire workload (memory and processes)

> into a single NUMA node, or 2-3 NUMA nodes that have the lowest latency

> between each other. You might find out that most of your current CPU

> "usage" is actually stalled CPU empty cycles, waiting for remote memory

> access...

>

> What should be done with large NUMA machines, like HPE SuperDome Flex or

> modern IBM large Power 10 servers, which are all NUMA? Those machines are

> expensive because they need global directory, which is implemented as fully

> associative RAM. Fully associative RAM is rather expensive. There are many

> nodes and you need a "phone book" to figure out which node houses your

> desired memory page. Those large NUMA boxes are direct competitors to RAC.

>

> Regards

>

> --

> Mladen Gogala

> Database Consultant

> Tel: (347) 321-1217https://dbwhisperer.wordpress.com

>

> -- http://www.freelists.org/webpage/oracle-l


--
http://www.freelists.org/webpage/oracle-l