AMD64 MP, ULE and threaded applications

[Alexandre "Sunny" Kovalenko]

I am slightly confused whether I should post 5.3 BETA related stuff here
or in stable@, so feel free to chastise me as appropriate.

I have fairly simple program which is heavily CPU bound and could be
easily parallelized. Recently I laid my hands on dual AMD64 no-brand
machine and installed 5.3 BETA1 on it. 

With SCHED_ULE, no number of threads would deliver more then single CPU
performance, and 'top' will report 50% CPU utilization.

With SCHED_4BSD, two threads would deliver ~185% of single CPU
performance and 'top' will report 100% spent in userland.

Test application could be found here:
http://members.verizon.net/~akovalenko/PDFPassword.tar.gz

To build, just type 'make'.

To run, type ./PDFPassword OValue <# of threads to start>

I use this benchmark on variety of MP hardware (Sun, RS/6000, Intel) and
throughput usually follows number of threads up to number of available
CPU.

I will bring this system up to RELENG_5 as of Saturday (9/4) morning EST
some time on Monday and retest. I can also test patches or try oprions
as needed. Unfortunately, I would not be able to provide remote access
to this box.

---
Alexandre "Sunny" Kovalenko.