Testing ubiblock

[Ezequiel Garcia]

Hello,

Here are the numbers on ubiblock, using the posted v2 patch.

The test is mainly focused on eraseblock wearing,
but some throughput information is also available.

If you want some background on ubiblock, read here:
http://lkml.indiana.edu/hypermail/linux/kernel/1211.2/02556.html

---

The test procedure is as follows:

First an image is created for each filesystem that will be tested
(currently ext4, vfat, ubifs).

This image has some files created by 'sysbench --prepare'.

Then an emulated environment is started and for each image
the following is done:

1. Insert nandsim driver (wear=0)
2. Format mtd device with ubiformat
3. Mount device (on ubiblock or not depending on fs)
4. Run sysbench test x 3, with one file-io mode. Report throughput.
5. Unmount device and remove ubi,ubifs,ubiblock drivers.
6. Report nand wear from nandsim.
7. Remove nandsim driver

Sysbench produces very artificial workloads, but it's very simple
and useful enough to get a feeling of ubiblock behavior under
specific conditions.

The results are: (complete reports attached)

Wear:		seqwr		rndwr		rndrw
--------------------------------------------------------------
ext4			5586		6932		5337
vfat			5554		8305		6005
ubifs-none	4965		4188		4141
ubifs-lzo		4104		4104		4104

Test time [s]:	seqwr		rndwr		rndrw
--------------------------------------------------------------
ext4			6.5		8.8		5.4
vfat			6.5		15.9		5.4
ubifs-none	5.0		0.5		0.3
ubifs-lzo		2.5		0.3		0.2

Transfer rate:	seqwr		rndwr		rndrw
--------------------------------------------------------------
ext4			7  M/s		450 K/s		700 K/s
vfat			7  M/s		250 K/s		500 K/s
ubifs-none	9  M/s		7   M/s		12  M/s
ubifs-lzo		18 M/s		10  M/s		16  M/s

---

Conclusions:

Despite being results obtained on a simulated nand, some conclusions
are possible.
First of all it's obvious that compression in ubifs greatly reduces
write I/O, thus reducing eraseblock wearing and improving throughput.
It's expected that implementing some kind of compression in ubiblock
would have a similar impact.

Sequential writes and random writes have a similar wearing in ubifs,
being log structured.
On the other side, block oriented filesystem, have a far worse
performance on random tests.
This is somewhat expected, given the design of ubiblock cache is
sequential-access oriented.
Given f2fs is also log-structured, we can hope to obtain some nice
numbers using it.

Interestingly, sequential writes produce a comparable wear
in ext2, vfat and ubifs-none.
Again, implementing some kind of compression in ubiblock might be
reduce eraseblock wearing numbers.

---

Thoughts?

Thanks,

    Ezequiel
-------------- next part --------------
A non-text attachment was scrubbed...
Name: report_ubiblock.tar.gz
Type: application/x-gzip
Size: 3088 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-mtd/attachments/20121212/9bb0894f/attachment.gz>