Skip to contentSkip to conversation

benchmark anobjrmap with 2.5.65-mm2

DISCUSSION [email protected] 1 participant 1 message Mar 22, 2003

lore
Share & tools
lore
#0 lore raw

Anonymous objrmap patches appear to several workloads a little for uniprocessor K6/2 475 mhz with 2 ide disks and 384 MB ram.

Build times for autoconf (a fork test), kernel, and perl.

                           autoconf    kernel         perl
2.5.65                     3845          1641         1348 seconds
2.5.65-mm1                 3898          1646         1326
2.5.65-mm2                 3895          1582         1312
2.5.65-mm2-anobjrmap       3824          1614         1284

autoconf and perl builds were faster with anon obj rmap. kernel build was not. kernel build uses pipe more than autoconf/perl build. Could be the fact there was only one sample too.

Lmbench pipe latency and bandwidth don't provide an obvious explanation why kernel build was slower with anobjrmap.

*Local* Communication latencies in microseconds - smaller is better

kernel                           Pipe   
2.5.65                           15.36 ms 
2.5.65-mm1                       15.80  
2.5.65-mm2                       15.21  
2.5.65-mm2-anobjrmap             14.11  

*Local* Communication bandwidths in MB/s - bigger is better

kernel Pipe

2.5.65                            64.5 MB/second
2.5.65-mm1                        63.3 
2.5.65-mm2                        65.9 
2.5.65-mm2-anobjrmap              65.1 

I'm running 2.5.65-mm3 now and will watch how kernel build goes there.

AIM7 workloads were generally a hair faster with anobjrmap.

More benchmarks on recent kernels at:
http://home.earthlink.net/~rwhron/kernel/latest.html

Irman process load starvation still appears in 2.5.65 and 2.5.65-mm[12].

Signature · 2 lines
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html