cvs commit: src/sys/platform/vkernel/conf files src/sys/platform/vkernel/i386 exception.c src/sys/platform/vkernel/include cothread.h md_var.h src/sys/platform/vkernel/platform cothread.c machintr.c src/sys/dev/virtual/disk vdisk.c
dillon at apollo.backplane.com
Thu Mar 20 09:36:23 PDT 2008
:Do we have some simple before/after numbers for typical operations like
:extracting pkgsrc-2007Q4.tar.gz and/or running some io benchmark (some
:phases of kernel compilation may be affected as well)? I don't have easy
:access to a dragonfly box atm.
The main advantage is that the cpu threads will no longer block when
reading from disk so the vkernel is a lot more responsive under load.
This replaces the junky cpu reservations I was using for ncpus > 2 before
(plus it works for -n 1). I would expect any I/O involving read-ahead
would operate more efficiently but I haven't run any real tests with
regards to I/O throughput. It will likely remain fairly poor from a
pure throughput point of view.
We need to convert the network driver (vke) over to using the new
interface, and improve the bulk operations a bit more.
<dillon at backplane.com>
More information about the Commits