Gentoo Archives: gentoo-user

From: Michael Mol <mikemol@×××××.com>
To: gentoo-user@l.g.o
Subject: Re: [gentoo-user] OT: HW-raid1 & SSD?
Date: Mon, 09 Jul 2012 19:30:29
Message-Id: CA+czFiCQK8DjMQ-hjuZpGzYt94YrXtTHiQiptm_ZbNF4h6sxUA@mail.gmail.com
In Reply to: Re: [gentoo-user] OT: HW-raid1 & SSD? by Florian Philipp
1 On Mon, Jul 9, 2012 at 3:12 PM, Florian Philipp <lists@×××××××××××.net> wrote:
2 > Am 09.07.2012 19:26, schrieb Jarry:
3 >> Hi,
4 >>
5 >> I'm building small server for VPS-hosting purpose (not decided
6 >> yet, probably KVM or ESXi). Althought it is "non production",
7 >> reliability is still issue for me. I'm considering using two
8 >> small SSD (ie. Intel 313, 24GB, SLC) in hw-raid1 (Adaptec 3805)
9 >> just for hypervisor, and 2xHDD/raid1 for VPS.
10 >>
11 >> I started looking for info on internet but results are not
12 >> conclusive. Opinions vary from "no problem, use SSD&raid1"
13 >> to "using SSD in raid is not recommended no matter if it is
14 >> SW or HW". Mostly missing trim-command is mentioned as reason
15 >> for not using SSD in any raid.
16 >>
17 >> Does anyone have any experience with SSD in raid (SW or HW)?
18 >> Is it safe to use it on server?
19 >>
20 >> Jarry
21 >
22 > I don't think trim is currently supported in mdaid. AFAIK dmraid
23 > supports it as long as the RAID controller plays along. I'm unsure about
24 > LVM mirrors.
25 >
26 > But trim is not your only issue: mdraid and LVM as well as most (all?)
27 > hardware RAIDs will mirror all blocks from one disk to the other when
28 > you assemble the RAID-1. This means that all blocks on your second SSD
29 > will immediately be marked dirty which incurs a significant performance
30 > penalty.
31
32 Well, yes and no. Yes, there will be a performance penalty from having
33 all the blocks marked dirty. But this would eventually happen anyway!
34
35 On the other hand, once all blocks are marked dirty, you have
36 performance _stability_. If you're in a position to care about
37 predictive performance metrics, stability is quite useful.
38
39 --
40 :wq