At the OpenVMS Bootcamp 2006, in session A901 Clair Grant & Burns Fisher did a 1 hour session titled "Intermediate Hyperthreads on Montecito and OpenVMS".
I'm not going to be able to summarize all that in 2 minutes.
Unfortunatly I don't think I am supposed to share that slideset here and now. I'll poke them to see if they are willing to reply here.
Bruce Ellis presented slides on hyperthreading in Canada using:
http://www.encompasscanada.com/documents/sem-tor07/tor07%20Ellis_Integrity_server_perf.pdfHoff has a web page:
http://64.223.189.234/node/13Google for +Hyperthreads +Montecito will give a good few docs to check
Here is my take, with some hands-on experience with Xeon hyperthreading.
- Hyperthreading tries to utilize the waits for memory which can be many ( 80+ ) cycles.
- It _should_ be strictly positive... but it is not.
- The extra scheduling and increased contentions slows things down.
- The scheduler needs to learn that all CPUs are NOT create equal... if thread 0 on core 1 is busy, and no thread is busy on core 2, then the scheduler should not pick CPU 5 (thread 1, core 1) but CPU 2. They 'look the same' but they will not behave the same. Yikes.
- Many OpenVMS SMP solutions already suffer from contention (think spinlocks and MPsync) where you have cycles available, but can not use them just yet. Adding more active threads is threat to the general health.
- Hyperthreading has been shown to give up to 30% improvement for just the right cpu / memory load mix, mostly memory bound where the cpu seems busy 100% in usermode but is really waiting for memory to come in. To get to htat 30%, one often needs active, aggresive, CPU affinity settings.
- OpenVMS applications tend to run less than 20% usermode and the inner modes often involve synchronization and serialization.
- Practically Hyperthreading shows -5% to + 10% gain, more towards 5%.
- Hyperthreading cause 100% confusion
So the risk-reward picture isn't great.
Therefor, OpenVMS took the typical cautious stance and decided to initially disable hyperthreading, but at least they did keep it available.
They may have changed the attitude with the update, but I would have expect that to be well documented (was it?), or maybe something slipped through the cracks?!
Ok... that was a little more than 2 minutes.
Hope this helps some,
Hein van den Heuvel (at gmail dot com)
HvdH Performance Consulting
