Figured it was time to send the list an update.
1. Our sysadm said that there are kernel issues and "perf stat" cannot run
on these systems
2. I installed a slob instance on both a fast and slow server. There was no
way to stop our workload from running, but I ran slob multiple times,
trying to keep each run short (~3 minutes) so the underlying workload would
not vary much. Also tried running when the load was relatively low. All of
the run results were pretty consistent with one another. Configuring SLOB
to run strictly LIO, I saw ~4M lio/sec vs. 1.5M lio/sec
3. Our sysadm did get turbostat installed. What I take as the most
important output from that was:
slow: TSC_MHz - 2600 Bzy_MHz - 1200 PKG_% - 100%
fast: TSC_MHz - 2606 Bzy_MHz - 3110 PKG_% - 0%
So the clock speeds are nearly identical, ,but the running speeds when busy
are very different. Also RAPL throttling is on in the slow server. I am
guessing this could account for the throttled cpu speed. I am starting to
google search to see where this can be set from. Anyone out there know any
good resources on RAPL throttling or how to disable it?
Thanks.
Henry
On Fri, Aug 11, 2017 at 2:52 PM, Henry Poras <henry.poras@xxxxxxxxx> wrote:
Stefan,
I agree about focusing on cpu (and perhaps memory). As for perf, from what
I have seen it should be available. I'll check path, but I thought it might
be a module not installed. I was able to start "perf stat" (didn't get a
command not found) but then got some lib error. I'll go back and check.
Henry
On Fri, Aug 11, 2017 at 1:42 PM, Stefan Koehler <contact@xxxxxxxx> wrote:
Hello Henry,
not quite sure what you are referring to as you are using Linux kernel
2.6.32-696.1.1.el6.x86_64 and perf is based on the perf_events interface
and it is exported by Linux kernel >= 2.6.31 - maybe it is just not in your
path ("/usr/sbin/perf").
In addition i may got your description wrong but you wrote "A select
count(*) from dba_objects showed this behavior as did Jonathan Lewis's
kill_cpu script." - so you can completely ignore the I/O or network stack
and focus on CPU (and memory) as kill_cpu (https://jonathanlewis.wordpre
ss.com/2016/10/03/kill-cpu/) has nothing to do with these stacks.
... and if you want to analyze the CPU (and memory caching) behavior -
measure it with perf and you will see the difference.
Best Regards
Stefan Koehler
Independent Oracle performance consultant and researcher
Website: http://www.soocs.de
Twitter: @OracleSK
Henry Poras <henry.poras@xxxxxxxxx> hat am 11. August 2017 um 19:27geschrieben:
available. I've asked the sysadms about this.
Some more minor updates and responses to suggestions:
Stefan - nice articles. Unfortunately neither turbostat or perf stat is
