Date: Sat, 23 Nov 1996 17:28:10 -0800 (PST) From: Tony Li <tli@jnx.com> To: smp@csn.net (Steve Passe) Cc: freebsd-smp@freefall.freebsd.org, Peter Wemm <peter@spinner.dialix.com> Subject: Re: SMP -current merge Message-ID: <199611240128.RAA28719@chimp.jnx.com> In-Reply-To: smp@csn.net's message of 24 Nov 96 00:43:09 GMT References: <199611232323.PAA02050@root.com> <199611240043.RAA20563@clem.systemsix.com>
next in thread | previous in thread | raw e-mail | index | archive | help
I have tested the times reported by "time make" against a stopwatch, and they are accurate to within 0.5 second. I was surprised to see increasingly better times beyond "-j 2". I guess this is from parallelism on the disk accesses, ie when both CPUs are waiting for disk IO on a particular object build, there is another they can work on if jobs > 2. The point of diminishing returns seems to be around 8 jobs. That's entirely consistent with the behavior of other SMP systems (e.g., Solbournes). The optimal point increases with your compute/IO ratio. For an 8 processor Solbourne, I normally used -j 11. Note that the -l flag to Gnu Make is what one would _like_ to use to insure a consistent load level. Unfortunately, it's implementation doesn't exhibit enough damping, so you get feast or famine. ;-( Tony
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199611240128.RAA28719>