Page 1 of 1
PVU Comparability of AIX vs Linux on 11.3.1 (or any version)
Posted: Wed Aug 19, 2015 11:01 am
by AGStafford
We are in the process of upgrading to 11.3.1 from 8.5 and management has asked us to consider moving from AIX to Linux.
Our server staff has indicated that similar CPU PVU on AIX vs Intel on Linux, the performance of AIX is 1.5 to 2 times faster.
Has anyone else moved from AIX to Linux and noticed performance differences with the same PVU count?
Andrew
Posted: Wed Aug 19, 2015 12:05 pm
by qt_ky
Each type of server and processor is assigned its own PVU rating, so it depends on which hardware you run AIX on vs which hardware you run Linux on. What are your processor types and PVU ratios for each?
Posted: Wed Aug 19, 2015 5:02 pm
by ray.wurlod
The PVU for AIX is higher than that for Intel/AMD exactly because the POWER processors are more powerful. To get the equivalent processing capability you would need more Intel/AMD cores.
Posted: Thu Aug 20, 2015 5:21 am
by priyadarshikunal
Do check the PVU calculator on IBM website on how the PVU calculated per core of different type of processors.
Posted: Thu Aug 20, 2015 7:36 am
by chulett
Re: PVU Comparability of AIX vs Linux on 11.3.1 (or any vers
Posted: Thu Aug 20, 2015 2:18 pm
by rkashyap
AGStafford wrote:Our server staff has indicated that similar CPU PVU on AIX vs Intel on Linux, the performance of AIX is 1.5 to 2 times faster.
DSXchangers, What are your thoughts on quoted statement?
Pardon my ignorance ... If for same PVU license, an AIX(/RISC) server is 1.5/2 times faster that a Linux server, then most of the DataStage installations would be migrating from Linux to AIX.
Posted: Thu Aug 20, 2015 2:36 pm
by qt_ky
There are too many variables at play, in addition to the OS and the processor type and speed, to say anything meaningful. I think you need a variety of benchmarking jobs that exercise the CPU, memory, disk, network, etc., that represent your typical workload and degree of parallelism, then run each set of jobs on the same data and then compare the results.
Posted: Thu Aug 20, 2015 5:07 pm
by ray.wurlod
Start by getting them to define what they mean by "performance".
Performance in an ETL environment is not about raw cycles/second. There's always heavy demand for memory and I/O to factor in, and the efficiency with which context switching is performed.
Posted: Thu Aug 20, 2015 6:18 pm
by rkashyap
Agreed ... definition of performance needs to be clarified and there are too many variables at play, but the statement about AIX's performance being 1.5/2 times that of Linux for same PVU licensing, sounds too good to be true.
I feel that even AIX sales reps would be circumspect in making this statement.
Posted: Thu Aug 20, 2015 6:25 pm
by ray.wurlod
They may be being simplistic, comparing only the GHz rating of the i386 and the POWER chips.
That is not the same as "performance" except in an extremely restricted sense.
Posted: Fri Aug 21, 2015 8:10 am
by PaulVL
It boils down to "Is your environment CPU constrained or IO constrained?" I do not factor in memory because that's a very cheap item to purchase and should not be a hinderence to a well maintained environment.
I have found that DS jobs are mostly travaling at the speed of IO.
I have also found that the shops I work(ed) in also tend to favor smaller jobs (20 stages or less).
Thus I tend to prefer Linux Grids for my environments. Mainly because the application teams never know how much they are going to grow their usage.