*Really* bad I/O latency with md raid5+dm-crypt+lvm

[Christian Pernegger]

[Please keep me CCed as I'm not subscribed to LKML]

Summary: I was hoping to use a layered storage setup, namely lvm on
dm-crypt on md raid5 for a new box I'm setting up, but that isn't
looking so good since a single heavyish writer will monopolise any and
all I/O on the "device". F. ex. while cp'ing a few GB of data from an
external disk to the array it takes ~10sec to run ls and ~2min to
start aptitude. Clueless attempts at a diagnosis below.

Hardware:
AMD Athlon II X2 250
2GB Crucial DDR2-ECC RAM (more after testing)
ASUS M4A785D-M PRO
4x WD1000FYPS
        connected to onboard SATA controller (AMD SB710 / ahci)

Software:
Debian 5.03 (lenny/stable)
Kernel: linux-image-2.6.30-bpo.2-amd64 (based on 2.6.30.5 it seems)

The 4 disks are each partitioned into a 256MB sdX1 and a $REST sdX2.
The sdX1s make up md0, a raid1 w/ 1.0 superblock for /boot.
The sdX2s make up md1, a raid5 w/ 1.1 superblock, 1MiB chunk size and
stripe_cache_size = 8192.
On top of md1 sits md1_crypt, a dm-crypt/luks layer using
aes-cbc-essiv:sha256 and a 256 bit key. It's aligned to 6144 sectors
(=3MiB / 1 stripe)
The whole of md1_crypt is an lvm PV with a metadatasize of 3008KiB.
(That's the poor-man's way of aligning the data to align the data to
3MiB / 1 stripe. The lvm tools in stable are too old for proper
alignment options.)
The VG consisting of md1_crypt has 16GiB root, 4GiB swap, 200GiB home
and $REST data LVs.
All filesystems are ext3 with stride=256 and stripe-width=768. home is
mounted acl,user_xattr, data acl,user_xattr,noatime. Readahed on the
LVs is at 6MiB (2 stripes).

So, first question: should this kind of setup work at all or am I
doing something pathological in the first place?

Anyway, as soon as I copy something to the array or create a larger
(upwards of a few hundred MiB) tar archive the box becomes utterly
unresponsive until that job is finished. Even on the local console the
completion time for a simple ls or cat is of the order of tens of
seconds, just forget about launching emacs.

Now I know that people have been ranting about desktop responsiveness
for a while but that was very much an abstract thing for me until now.
I'd never have thought it would hit me on a personal streaming media /
backups / multi-user general purpose server. Well, at the moment it's
single-user, single-job ... :-(

Here's what I tried:
changing scheduler from cfq to deadline (no effect)
tuning proc/sys/vm/dirty*ratio way down (no effect)
turning off NCQ (some effect, maybe)
raising queue/nr_requests really high, e. g. 1000000 (helps
noticeably, especially when NCQ is off)

Ideas:
According to openssl speed aes-256-cbc the CPUs encryption speed is
~113 MiB/s (single core, est. for 512b blocks). Obviously the array is
much faster than that. I can't find the benchmarks ATM but the numbers
seemed plausible for 70 MiB/s (optimistic est. for sequential access)
disks at the time. So lets say at least 50% faster. Wouldn't this move
the bottleneck for requests away from the scheduler queue thus
rendering it ineffective?

Also, running btrace on the various block device layers I never see
>4k writes, even when using dd with a blocksize of 3 MiB. Is this
normal? btrace on (one of) the component disks shows some merged
requests at least. Am I wrong or would scheduling/merging lots and
lots of 4k blocks effectively, take an *insane* queue length?

All comments and suggestions welcome

Thank you,

Chris
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/