On Mon, 2006-03-13 at 19:10 +0100, Adrian Bunk wrote:
> I'm still a bit surprised, since in the rest of the kernel we are even
> going from -O2 to -Os for getting better performance.
>
> Robert said he wanted to post some numbers showing that -O3 is
> measurably better for you [1], but I haven't seen them.
I just ran some numbers. At large packet sizes, it doesn't matter what
options we use, because we spend all of our time in __iowrite_copy32,
which uses the string copy instructions.
For small packets, my quick tests indicate that -Os gives about 5%
*better* performance than -O3 (using gcc 4 on FC4). This is in line
with what people have been finding in the kernel in general recently.
So if I change that CFLAGS line from -O3 to -Os, are we in OK
shape? :-)
> > +_ipath_idstr:="PathScale $(shell date +%F)"
> > +EXTRA_CFLAGS += -DIPATH_IDSTR='$(_ipath_idstr)' -DIPATH_KERN_TYPE=0
> >...
>
> UTS_VERSION is already available and printed at the top of dmesg.
> I don't see the point in printing it a second time.
Good point. The idstr stuff is for our out-of-tree drivers.
Thanks,
<b
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
