Date: Tue, 14 Jul 2009 09:23:06 -0700 (PDT) From: Richard Mahlerwein <mahlerrd@yahoo.com> To: Matthew Seaman <m.seaman@infracaninophile.co.uk> Cc: Free BSD Questions list <freebsd-questions@freebsd.org> Subject: Re: ZFS or UFS for 4TB hardware RAID6? Message-ID: <719914.10546.qm@web51006.mail.re2.yahoo.com>
next in thread | raw e-mail | index | archive | help
--- On Tue, 7/14/09, Matthew Seaman <m.seaman@infracaninophile.co.uk> wrote= :=0A=0A> From: Matthew Seaman <m.seaman@infracaninophile.co.uk>=0A> Subject= : Re: ZFS or UFS for 4TB hardware RAID6?=0A> To: mahlerrd@yahoo.com=0A> Cc:= "Free BSD Questions list" <freebsd-questions@freebsd.org>=0A> Date: Tuesda= y, July 14, 2009, 4:23 AM=0A> Richard Mahlerwein wrote:=0A> =0A> > With 4 d= rives, you could get much, much higher=0A> performance out of=0A> > RAID10 = (which is alternatively called RAID0+1 or=0A> RAID1+0 depending on=0A> > th= e manufacturer=0A> =0A> Uh -- no.=A0 RAID10 and RAID0+1 are superficially= =0A> similar but quite different=0A> things.=A0 The main differentiator is = resilience to disk=0A> failure. RAID10 takes=0A> the raw disks in pairs, cr= eates a mirror across each pair,=0A> and then stripes=0A> across all the se= ts of mirrors.=A0 RAID0+1 divides the=0A> raw disks into two equal=0A> sets= , constructs stripes across each set of disks, and then=0A> mirrors the=0A>= two stripes.=0A> =0A> Read/Write performance is similar in either case: bo= th=0A> perform well for the sort of small randomly distributed IO=0A> opera= tions you'ld get when eg.=0A> running a RDBMS.=A0 However, consider what ha= ppens if=0A> you get a disk failure.=0A> In the RAID10 case *one* of your N= /2 mirrors is degraded=0A> but the other N-1=0A> drives in the array operat= e as normal.=A0 In the RAID0+1=0A> case, one of the=0A> 2 stripes is immedi= ately out of action and the whole IO=0A> load is carried by=0A> the N/2 dri= ves in the other stripe.=0A> =0A> Now consider what happens if a second dri= ve should=0A> fail.=A0 In the RAID10=0A> case, you're still up and running = so long as the failed=0A> drive is one of=0A> the N-2 disks that aren't the= mirror pair of the 1st failed=0A> drive.=0A> In the RAID0+1 case, you're o= ut of action if the 2nd disk=0A> to fail is one=0A> of the N/2 drives from = the working stripe.=A0 Or in=0A> other words, if two=0A> random disks fail = in a RAID10, chances are the RAID will=0A> still work.=A0 If=0A> two arbitr= arily selected disks fail in a RAID0+1 chances=0A> are basically=0A> even t= hat the whole RAID is out of action[*].=0A> =0A> I don't think I've ever se= en a manufacturer say RAID1+0=0A> instead of RAID10,=0A> but I suppose all = things are possible.=A0 My impression=0A> was that the 0+1 terminology was = specifically invented to=0A> make it more visually distinctive=0A> -- ie to= prevent confusion between '01' and '10'.=0A> =0A> =A0=A0=A0 Cheers,=0A> = =0A> =A0=A0=A0 Matthew=0A> =0A> [*] Astute students of probability will poi= nt out that this=0A> really only=0A> makes a difference for N > 4, and for = N=3D4 chances are=0A> evens either way that failure of two drives would tak= e out=0A> the RAID.=0A> =0A> -- Dr Matthew J Seaman MA, D.Phil.=A0 =A0 =A0= =0A> =A0 =A0 =A0 =A0 =A0 =A0=A0=A07=0A> Priory Courtyard=0A> =A0 =A0 =A0 = =A0 =A0 =A0 =A0 =A0=0A> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0=0A> =A0 =A0 =A0 =A0= =A0 =A0 =A0=0A> =A0=A0=A0Flat 3=0A> PGP: http://www.infracaninophile.co.uk= /pgpkey=A0=0A> =A0=A0=A0Ramsgate=0A> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0=0A> = =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0=0A> =A0 =A0 =A0 =A0 =A0 =A0 =A0=0A> =A0=A0= =A0Kent, CT11 9PW=0A> =0A=0A--- On Tue, 7/14/09, Matthew Seaman <m.seaman@i= nfracaninophile.co.uk> wrote:=0A=0A> From: Matthew Seaman <m.seaman@infraca= ninophile.co.uk>=0A> Subject: Re: ZFS or UFS for 4TB hardware RAID6?=0A> To= : mahlerrd@yahoo.com=0A> Cc: "Free BSD Questions list" <freebsd-questions@f= reebsd.org>=0A> Date: Tuesday, July 14, 2009, 4:23 AM=0A> Richard Mahlerwei= n wrote:=0A> =0A> > With 4 drives, you could get much, much higher=0A> perf= ormance out of=0A> > RAID10 (which is alternatively called RAID0+1 or=0A> R= AID1+0 depending on=0A> > the manufacturer=0A> =0A> Uh -- no. RAID10 and R= AID0+1 are superficially=0A> similar but quite different=0A> things. The m= ain differentiator is resilience to disk=0A> failure. RAID10 takes=0A> the = raw disks in pairs, creates a mirror across each pair,=0A> and then stripes= =0A> across all the sets of mirrors. RAID0+1 divides the=0A> raw disks int= o two equal=0A> sets, constructs stripes across each set of disks, and then= =0A> mirrors the=0A> two stripes.=0A> =0A> Read/Write performance is simila= r in either case: both=0A> perform well for the sort of small randomly dist= ributed IO=0A> operations you'ld get when eg.=0A> running a RDBMS. However= , consider what happens if=0A> you get a disk failure.=0A> In the RAID10 ca= se *one* of your N/2 mirrors is degraded=0A> but the other N-1=0A> drives i= n the array operate as normal. In the RAID0+1=0A> case, one of the=0A> 2 s= tripes is immediately out of action and the whole IO=0A> load is carried by= =0A> the N/2 drives in the other stripe.=0A> =0A> Now consider what happens= if a second drive should=0A> fail. In the RAID10=0A> case, you're still u= p and running so long as the failed=0A> drive is one of=0A> the N-2 disks t= hat aren't the mirror pair of the 1st failed=0A> drive.=0A> In the RAID0+1 = case, you're out of action if the 2nd disk=0A> to fail is one=0A> of the N/= 2 drives from the working stripe. Or in=0A> other words, if two=0A> random= disks fail in a RAID10, chances are the RAID will=0A> still work. If=0A> = two arbitrarily selected disks fail in a RAID0+1 chances=0A> are basically= =0A> even that the whole RAID is out of action[*].=0A> =0A> I don't think I= 've ever seen a manufacturer say RAID1+0=0A> instead of RAID10,=0A> but I s= uppose all things are possible. My impression=0A> was that the 0+1 termino= logy was specifically invented to=0A> make it more visually distinctive=0A>= -- ie to prevent confusion between '01' and '10'.=0A> =0A> Cheers,=0A>= =0A> Matthew=0A> =0A> [*] Astute students of probability will point ou= t that this=0A> really only=0A> makes a difference for N > 4, and for N=3D4= chances are=0A> evens either way that failure of two drives would take out= =0A> the RAID.=0A=0ASorry, you are correct. Thanks for clearing that up. = =0A=0AI *have,* by the way, stumbled across them a couple of times in the c= onsumer/on-board market, and that's why I tend to remember that and include= it even though it's incorrect now. IIRC (which is NOT certain :), I remem= ber once perhaps back around 2000 that a major mag tested some and found th= at it was only nomenclature differences: all RAID10/1+0/0+1 that were avail= able were all RAID10. =0A=0AAnd, if I recall, that was back in the PATA da= ys.=0A=0AAnyway, NP. I could also be off my rockers.=0A=0A(Oh, and thanks = for the addendum, I actually was following and thinking "...now wait a minu= te..." and then you clarified that last bit. :) )=0A=0A=0A
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?719914.10546.qm>