Performance problems during parallel processing

[Former Member]

Hello,

since the migration of our financial solution system (FS-CD) from AIX/Oracle to RHEL6/DB2/vSphere4.1 two weeks ago we have been facing massive performance problems during parallel processing in the system.

We are now able to reconstruct this behavior while running a SGEN in other systems in our new DB2 LUW landscape:

The higher the parallelization, the less CPU usage!

A SGEN of 44,380 objects with 17 parallel jobs in a sandbox system with 8 vCPUs consumes a max. CPU utilization of just 42%.

A SGEN in the same system with 6 parallel jobs leads to a CPU utilization of about 70% (further analysis over longer periods still need to be performed).

Same behavior in another quality assurance system (BW, 4 vCPUs):
SGEN 13x parallel: max. 50% CPU usage.
SGEN 5x parallel: max. 80% CPU usage.
There's nearly the same throughput of objects per second, whether we are working with 5 or 13 parallel jobs.

Important:
In last year's proof of concept for the platform change, this behavior didn't occur!
Back then a 15x parallel SGEN led to a nearly constant CPU utilization of 70-80%.

Differences between last year's PoC and the current production (apart from altered DB2 parameter recommendations by SAP):
PoC: DB2 9.7 FP3 / FP4, RHEL 5.7
Current production: DB2 9.7 FP5, RHEL 6.1

Until now no explanations were found for VMware, DB2 or Linux which led to a significant improvement of processing times, so we would like to ask if there's anybody who can confirm or even reconstruct this behavior during parallel processing in his system landscape?

Kind regards,
André Richter

[Former Member]

There's nearly the same throughput of objects per second, whether we are working with 5 or 13 parallel jobs.

This indicates you are hitting a bottleneck with 5 processes already. Could be the database for example. SGEN is typically quite cpu intensive on the application server, but on todays cpu's the database becomes a bottleneck very easily.

Cheers Michael

[Former Member]

Hi Andre,

Please try with the following you might get better throughput

- Runstat using DB20 for the following tables for start of SGEN

GENSETC, REPOTEXT, DYNPSOURCE

- Turn off archive log

If you have resource (h/w) use multiple App Servers .

I used CI+ five App Servers ( with 10 Dialog proceses on each server) for 3.5 TB database and SGEN completed in 3 hr 40 minutes.

Regards,

Dipam