Date: Tue, 25 Aug 2009 09:20:03 GMT From: Ivan Voras <ivoras@freebsd.org> To: freebsd-geom@FreeBSD.org Subject: Re: kern/113885: [gmirror] [patch] improved gmirror balance algorithm Message-ID: <200908250920.n7P9K3XA042099@freefall.freebsd.org>
next in thread | raw e-mail | index | archive | help
The following reply was made to PR kern/113885; it has been noted by GNATS. From: Ivan Voras <ivoras@freebsd.org> To: bug-followup@freebsd.org, zuborg@advancedhosters.com Cc: Subject: Re: kern/113885: [gmirror] [patch] improved gmirror balance algorithm Date: Tue, 25 Aug 2009 11:11:12 +0200 The patch will not increase streaming read performance beyond what's possible with a single drive, it will improve random read performance in certain cases where reads are localized in such ways that reading some of them from one drive and others from the other drive helps. The reason why there is no scalability with streaming read performance vs what can be achieved with RAID0/3/5 is that there is no striping here. For example: if you need to read 4 striped blocks from a RAID0 of two drives, blocks 0 and 2 will be sequentially stored on the first drive, blocks 1 and 3 will be sequentially stored on the second drive. Thus reading the 4 blocks will result in two sequential reads per drive. OTOH, with RAID1, blocks 0 and 2 will be stored with a "gap" between them, containing block 1, and cannot be read sequentially, but a seek is needed. This is why e.g. the "split" method (which effectively does striping on the request level) doesn't help much with performance.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200908250920.n7P9K3XA042099>