bombsquad n00b
Joined: 28 Aug 2007 Posts: 12 Location: Coventry, United Kingdom
|
Posted: Tue Aug 28, 2007 7:53 pm Post subject: gentoo-hppa on HP C360 - performance |
|
|
I've been persevering with Gentoo on my C360 - I'd use HP-UX but there seems to be no (affordable) way to get decent RAID support, amongst other things. Add to that my general familiarity with Linux and Gentoo (using it on an i686 laptop for several years now) this seems to be the way to go.
However, bearing in mind the aging processor (367MHz PA-8500) the machine does seem to spend an age doing anything. HP-UX 11i on both 32-bit and 64-bit kernel widths is notably faster although when running several intensive jobs, Linux seems more responsive. top(1) tells me that processes spend a lot of time in kernel mode under Linux. I know memory operations are a performance hit to RISC machines in general Perhaps this is an hppa-sources issue, or even a gcc one, but from a most capable RISC architecture such as PA2.0, massive caches and 1GB of RAM, running 32-bit, I'd expect more. Things that seem to take the longest are python scripts - a bugger as Portage is written in it ;)
Anyway, enough of my subjective opinion. Perhaps PA-RISC gurus can clear up some questions for me. All in all I'm after performance-enhancing information for the old girl. Perhaps you lot know some tricks between you. This is not a slow machine, but it behaves like one under Linux.
1. Are there any recommended CFLAGS that give better performance?
I've tried -O3, -O2 and now -Os to see which extreme of the gcc code generator might improve things.
2. What exactly does -mjump-in-delay do?
I understand the principle of the delay slot in PA-RISC programming, but I'm unsure as to what benefit, if any, this machine flag for gcc would confer.
3. Kernel configuration options?
The machine will only boot with a 4K pagesize. I haven't experimented with the different allocators (SLAB/SLUB/SLOB) but my only guess is to try and simplify the kernel as much as possible. Maybe even optimise for size so more of it fits in the instruction cache during system calls ;)
The benefit of your collective expertise would be much appreciated.
Cheers |
|