Date: Mon, 31 Jul 2006 17:09:23 -0700 From: Paul Allen <nospam@ugcs.caltech.edu> To: Julian Elischer <julian@elischer.org> Cc: Peter Jeremy <peterjeremy@optushome.com.au>, freebsd-current@freebsd.org, Brian Candler <B.Candler@pobox.com> Subject: Re: vmstat's entries type Message-ID: <20060801000923.GA308@mark.ugcs.caltech.edu> In-Reply-To: <44CE5634.9080205@elischer.org> References: <200607251254.k6PCsBef092737@lurza.secnetix.de> <20060728134701.GA45273@uk.tiscali.com> <20060728210154.GC748@turion.vk2pj.dyndns.org> <200607311437.57273.jhb@freebsd.org> <44CE5634.9080205@elischer.org>
next in thread | previous in thread | raw e-mail | index | archive | help
>From Julian Elischer <julian@elischer.org>, Mon, Jul 31, 2006 at 12:12:52PM -0700: > but you do want to disable interrupts so that you don't get switched to > another cpu in the middle of it. > That would result in the possibility of 2 threads writing to the same > stats at the same time. Or just setup the sched to pin the kernel thread to the CPU and then use N_CPU threads for your packet processing + a load balancing algorithm to allocate flows between threads. I'm not saying this always the best solution, but certainly it is worth consideration--especially in cases where the word 'flow' can be omitted from my earlier statement without making anyone quesy. (Parallelism versus pipelining: the difference between Time-Division-Multiplexing each step in a packet's lifetime and Spacial-Division Multiplexing the flows versus Time-Division-Multiplexing the flows and Spacial-Division Multiplexing the packet's lifetime.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20060801000923.GA308>