On 5/09/2023 9:19 pm, Scott Chapman wrote:
It's more complicated than that. Although I would agree that if an LPAR has
only a single zIIP, likely SMT would be a good idea. But B is not true for
intervals that people usually consider when looking at utilization levels
because at the level of dispatch intervals, it's much more likely there's at
least some intervals where the zIIPs are 100% busy. It really depends on
arrival patterns and how much you care about very short transactions that may
be running on the zIIPs.
I had a look at some RMF data, and I think that zIIP velocity (from zIIP
using and zIIP delay in the SMF 72 records) might be a good indicator.
It hopefully gives an indication of the real delay much better than the
overall utilization.
I saw that the velocities on my system were quite low i.e. lots of delay
even with very low utilization. I suspect (guess) that this is due to
short running Java jobs where the JVM is running multitple threads
during JVM initialization. This would probably benefit from SMT2. On the
other hand if your zIIP work is single threads arriving randomly you may
not see delay until higher utilization. Does DDF run multiple threads
for a single unit of work?
I wrote a SMF real time interface sample to read type 72 records and
issue a message or even a command to change the SMT setting based on
theSMF 72 zIIP velocity across all service classes. You can see the
program here:
https://github.com/BlackHillSoftware/easysmf-samples/tree/main/easysmf-rti/rti-smt-switch
My theory is if the zIIP velocity is low, the work would benefit from
SMT2 because there are lots of threads waiting. If the velocity is high,
SMT might be hurting due to multiple threads on a processor when there
is no work waiting. Obviously what is considered "high" and "low" will
depend on your environment. Martin's question about a WLM reset also
applies.
--
Andrew Rowley
Black Hill Software
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
