[Lse-tech] Success with MQ+loadbalance and Andrea's NUMA scheduler

[Peter R. <fr...@zk...>]

Hi all,

    I just thought I would forward along some results of testing with my
8 CPU / 8 GB Alpha GS160 partition.  I have updated the MQ and
loadbalance patches to work on an Alpha system - took all of about five
minutes. :)  Anyway, it withstood two concurrent "make -j 16 boot;make
clean" (with MAKE="make -j 16" set and all output to /dev/null) along
with eight simultaneous bonnie++ processes split over two disks over
this past weekend.  The load average hovered consistently between 100
and 200 depending on how much parallelism was available to the kernel
builds at the time.  Oh, of course - I also added in Andrea's NUMA
Wildfire support patch (but _not_ the NUMA scheduler patch).

    I also decided to time the builds, and all of it looks good.  With
the loadbalance turned on and set to "4 0 5", kernel build times ran
between 1:00 and 1:10.  With loadbalance turned off, kernel build times
ran between 1:10 and 1:20.  On an otherwise identical kernel with
Andrea's NUMA scheduler instead of MQ+loadbalance, build times are
between 1:50 and 2:10.  I also just noticed that the load suddenly
spiked from it's usual 150 up to 550 and is still climbing and the
system is barely responsive.

    Thought people might like to know how the MQ+loadbalance patch was
working.  Now I just need to get Andrea to needle Linus into accepting
the NUMA for Wildfire patch... ;)  BTW - would anyone like to see any
benchmarks on this system (preferably not requiring _too_ much disk I/O
as I only have the one SCSI chain until I get my FC adapter)?  Just let
me know...  Thanks!

 - Pete