Performance results
I did some performance testing today. I compared our build server (P4 1core, 3 GHz, 2GB RAM, linux), against our T1000 (6 cores, 8GB RAM, Solaris 10).I compared two testcases:
1. A cross compile to ARM architecture of the linux kernel. For that I used a custom built arm-linux-gcc 3.4.5.
2. A cross compile to ARM architecture of one of our applications. For that I used a custom built arm-elf-gcc 4.1.0.
I did no optimizing except for passing "-j 32" to make on the T1000.

Results are interesting. Probably the T1000 cannot use it's full force because lots of io has to be done. Performance for our application is worse because at the end of the compile process there is a long linking phase that can run only on one core.
That is a good starting point. I hope there are some screws to tune this. Maybe linux would perform better on the T1000, Colm MacCárthaighs blog suggests this. But linux still does not work, at least not on my T1000. Still no news from Dave Miller or Fabbione on this.

1 Comments:
Have you tried to run that test without any I/O. Note that per default /tmp is a RAM-Disk as long as there's physical memory available in Solaris.
Have a look at "iostat -xtcP 2" during the make run to see if 32 processes are battling for the one poor S-ATA disk...
Post a Comment
<< Home