P4 (4.10) kernel build performance with/without SMP/APIC_IO/hyperthreading

Dewayne Geraghty dewayne_geraghty at yahoo.com.au
Fri Jun 11 21:51:43 GMT 2004


Re an earlier em thread, "HTTP and SMP" April 2004, I
decided to test my out-of-the-box P4 (2.8E) HT
threading enabled with SATA HD.

Test was to build identical kernels using FreeBSD 4.10
taking note of the compile times. Then reboot with:
SMP, APIC_IO in kernel, and machdep.hlt_logical_cpus:
0.  In all cases CPU_ENABLE_SSE was also enabled.

Summary for two clean kernel builds, using csh time
---------------------------------------------------
Kernel build without SMP, APIC_IO, (no variable 
machdep.hlt_logical_cpus)
1) 143.612u 23.095s 2:52.19 96.8%  1487+2195k
1872+2748io 0pf+0w
2) 143.831u 22.830s 2:55.61 94.9%  1488+2194k
3249+2747io 130pf+0w

Kernel build with SMP, APIC_IO,
machdep.hlt_logical_cpus=0
3) 145.106u 33.345s 2:57.38 100.5% 1451+2093k
3205+2746io 128pf+0w
4) 146.284u 32.481s 2:50.61 104.7% 1446+2093k
207+2745io 0pf+0w

An interesting result - the "run (3)" time was
reproduced and revealed that files were being deleted
via softupdate, and should be discarded. Experimental
error on my part.  This didn't affect runs 1 & 2.    

I repeated run 4 redirecting io to a file to reduce
screen delays, but with little difference.
# cd ..; rm -R TEST3; cd /sys/i386/conf; config TEST3;
cd ../../compile/TEST3; 
# time make depend all >/tmp/build.lis
146.155u 32.623s 2:50.59 104.7% 1447+2090k 185+2755io
0pf+0w

In this test of kernel builds, you'll save around 1.1%
of your time using SMP with logical HT CPU's on
FreeBSd 4.10.  

Regards, Dewayne.

PS: Of course I attempted to "make -j2 depend all" but
this failed as a previous module hadn't completed when
needed. e.g. bioscall.s


Find local movie times and trailers on Yahoo! Movies.
http://au.movies.yahoo.com


More information about the freebsd-stable mailing list