Date: Mon, 15 Apr 1996 12:50:07 -0700 (MST) From: Terry Lambert <terry@lambert.org> To: gpalmer@FreeBSD.org (Gary Palmer) Cc: scrappy@ki.net, nash@mcs.com, hackers@FreeBSD.org Subject: Re: Unices are created equal, but... Message-ID: <199604151950.MAA09397@phaeton.artisoft.com> In-Reply-To: <1061.829523630@palmer.demon.co.uk> from "Gary Palmer" at Apr 15, 96 00:13:50 am
next in thread | previous in thread | raw e-mail | index | archive | help
> Writing the 32 Megabyte file, 'iozone.tmp'...15.367188 seconds > Reading the file...9.437500 seconds > > IOZONE performance measurements: > 2183511 bytes/second for writing the file > 3555436 bytes/second for reading the file > > In both cases, it was at least 42% faster for reading than writing... This is expected... a partial FS block write requires a read. This is actually bogus, down to the device block level, because the page size is 4k, and there is an 8 bit bitmap, capable of taking the sparse bitmap granularity to 512k (one device block for most modern non-optical devices). For writes on a power of two block boundry for an area not spanning a second exp2(log2(n)+1), this should mean no read-before-write is ever required. This is applicable to prefetch for both sequential reads and sequential writes, and for regionally aligned access for random I/O. Since we don't implement this, even though IOZONE does exactly that kind of access, we still take read hits on writes up to the FS block size. Terry Lambert terry@lambert.org --- Any opinions in this posting are my own and not those of my present or previous employers.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199604151950.MAA09397>