Hello everybody,

following scenario: I'm analyzing the performance of a DB2-VM (VMware) while running on Intel and on AMD. The VM is the same, the state of the DB2-DB also (it's being restored everytime before each test-run). Storage is also the same (raw devices/direct LUNs). On both cases the VM runs in a single NUMA-Node.

I'm facing an enormous runtime-gap in the offline reorg-field, clearly noticeable on a very large table. On the Intel CPU it takes 16 hours; on the AMD 23 hours.

Here the details:

Intel:  Intel(R) Xeon(R) Platinum 8462Y+ 2.8 GHz 32-Core Processor
AMD: AMD EPYC 9354 3.25 GHz 32-Core Processor
VM: 24 cores

My question to you: Does anybody have experience on these sort-heavy operations and is able to explain this runtime difference? Is the AMD-CPU less suitable for this kind of task?

Thank you in advance for any hint

------------------------------
Miguel Sanjurjo
Chief Database Architect
BITMARCK Technik GmbH
Pinneberg
------------------------------

Miguel,

Sort operations require a lot of main memory access which is a big source of latency; while the AMD processor has a faster clock speed you are unable to take advantage of this (and may see a penalty) due to processor pipeline stalls.  Because of this, at least during the sort phase, the reorg process on this system will perform slower.

System profiling tools like 'perf' can be used to compare CPU profiles between the two scenarios.  If the data points to specific functions performing poorly, L3 support would be interested in the data.

------------------------------
Matthew Emmerton
Db2 Development
------------------------------

Matt,

thanks for your very insightful answer. Processor pipeline stalls is one of those aspects I would have never thought about in this context.

Because of the long runtime of these reorgs I'm analyzing, I have to think about how and when to run a tool like perf. I will get back to you about this.

We will for now stay on Intel CPUs

------------------------------
Miguel Sanjurjo
Chief Database Architect
BITMARCK Technik GmbH
Pinneberg
------------------------------

Original Message:
Sent: Fri July 04, 2025 10:11 AM
From: Matthew Emmerton
Subject: reorg performance on Intel vs. AMD

Miguel,

Sort operations require a lot of main memory access which is a big source of latency; while the AMD processor has a faster clock speed you are unable to take advantage of this (and may see a penalty) due to processor pipeline stalls.  Because of this, at least during the sort phase, the reorg process on this system will perform slower.

System profiling tools like 'perf' can be used to compare CPU profiles between the two scenarios.  If the data points to specific functions performing poorly, L3 support would be interested in the data.

------------------------------
Matthew Emmerton
Db2 Development
------------------------------

Hi, 

 Intel(R) Xeon(R) Platinum 8462Y has 60MB of L3 cache

AMD EPYC 9354 has 256MB of L3 cache.

it is odd, AMD should be faster, non ?

------------------------------
malek shabou
------------------------------

Original Message:
Sent: Fri July 04, 2025 10:11 AM
From: Matthew Emmerton
Subject: reorg performance on Intel vs. AMD

Miguel,

Sort operations require a lot of main memory access which is a big source of latency; while the AMD processor has a faster clock speed you are unable to take advantage of this (and may see a penalty) due to processor pipeline stalls.  Because of this, at least during the sort phase, the reorg process on this system will perform slower.

System profiling tools like 'perf' can be used to compare CPU profiles between the two scenarios.  If the data points to specific functions performing poorly, L3 support would be interested in the data.

------------------------------
Matthew Emmerton
Db2 Development
------------------------------