Date: Tue, 16 Oct 2001 16:01:40 +1000 (EST) From: Bruce Evans <bde@zeta.org.au> To: Matt Dillon <dillon@earth.backplane.com> Cc: Ollivier Robert <roberto@ns2.freenix.org>, Peter Wemm <peter@wemm.org>, "David O'Brien" <obrien@FreeBSD.org>, <cvs-committers@FreeBSD.org>, <freebsd-arch@FreeBSD.org> Subject: Re: cvs commit: src/sbin/newfs newfs.8 newfs.c Message-ID: <20011016154235.N6619-100000@delplex.bde.org> In-Reply-To: <200110151607.f9FG7QL54919@earth.backplane.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 15 Oct 2001, Matt Dillon wrote: > note: we shouldn't default to anything other then 8K/1K or 16K/2K. No > other combinations are optimal at the moment. This depends on the mix of file sizes, and probably on the drive. My benchmarks (made just after the dirpref changes) showed that 4K/2K, 8K/2K and 16K/2K are optimal for holding /usr/src, as measured by tarring up the src tree on a freshly mounted filesystem after just having written the src tree to an empty filesystem (the write speed is relatively sensitive to the layout). This is presumably caused by the average fragment size being related to 2K. ffs-4096-512: tar cf /dev/null src: 51.50 real 1.09 user 12.61 sys ffs-4096-1024: tar cf /dev/null src: 39.77 real 1.17 user 12.16 sys ffs-4096-2048: tar cf /dev/null src: 37.29 real 1.26 user 12.51 sys ffs-4096-4096: tar cf /dev/null src: 38.92 real 1.00 user 12.51 sys ffs-8192-1024: tar cf /dev/null src: 39.39 real 1.07 user 11.27 sys ffs-8192-2048: tar cf /dev/null src: 32.90 real 0.89 user 11.50 sys ffs-8192-4096: tar cf /dev/null src: 33.53 real 1.02 user 11.33 sys ffs-8192-8192: tar cf /dev/null src: 37.85 real 1.10 user 11.54 sys ffs-16384-2048: tar cf /dev/null src: 28.47 real 1.09 user 10.76 sys ffs-16384-4096: tar cf /dev/null src: 29.64 real 1.12 user 10.86 sys ffs-16384-8192: tar cf /dev/null src: 32.67 real 0.98 user 10.83 sys ffs-16384-16384: tar cf /dev/null src: 44.43 real 1.07 user 11.24 sys Note that the extremes of 4K:512 and 16K:16K are the worst, despite 4K:512 having the 8:1 ratio. The filesystem for this test was rather small (about 1GB) IIRC. Larger block sizes work relatively better for larger filesystems, not because they can be accessed more efficiently (clustering makes this advantage small), but because they reduce the number of cylinder groups by a small factor. The dirpref changes reduced this advantage significantly. Bruce To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-arch" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20011016154235.N6619-100000>