>>> On 1/25/2014 at 01:09 AM, David Crayford <[email protected]> wrote: 
> On 25/01/2014 1:57 PM, Mark Post wrote:
-snip-
>> Given who I work for, I would truly like to believe that, but I have grave 
> doubts about such statements unless the sources are cited, etc.  I know for a 
> fact that Java on Linux for System z can be as much of a resource hog as on 
> any other platform.  It all really boils down to whether there are good 
> performance tools to work with, application programmers that are willing to 
> not treat hardware as an infinite resource and all groups involved able to 
> work together.
> 
> http://www.dovetail.com/docs/coz/zos_hybrid_batch_zes30_ibm.pdf page 35.

I'm looking at that document, and I see the system involved was a 2094-405.  
So, a reduced-capacity system to start with.  The improvement with the addition 
of a zIIP (in zAAP) mode increased throughput because specialty processors run 
at full speed, but they only process the part of the workload that's eligible.  
The IFL, on the other hand, will be running _everything_ at full speed, which 
is definitely going to skew the results.  So, while I support the idea of 
offloading cycles from z/OS to Linux where it makes sense, I am not going to 
say the z/OS software stack is a performance bottleneck.  More likely, in my 
mind, a combination of capacity capping, system load, and whatever percentage 
of the workload being eligible for the zIIP limiting the throughput.  One thing 
to keep in mind is that z/OS and z/VM are written in a language far closer to 
assembler code than Linux, with decades of effort into optimizing the code for 
performance.  That's why z/VM could support thousands of interactive CMS users 
on pre-System z hardware, but Linux ran like a pig on the same hardware.  It 
wasn't until IBM began cranking up the CPU speeds with System z and adding 
hardware instructions that were specific to the workload that Linux got to a 
really good level of performance on the mainframe.


Mark Post

----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN

Reply via email to