Expected performance of RAID5
Mikhail Teterin
My measurements of the performance of a RAID5 array of 6 disks with iozone
(http://www.iozone.org/) show, that it is about the same as that of a
single disk.
Is that to be expected? I know, I'm gaining the redundancy of the RAID5, but
I expected the IO to be faster too -- if not 5 (6-1) times faster...
The disks are all Serial ATA 200Gb <WDC WD2000JD-00FYB0>. The six of them,
that constitute the RAID5 are attached to <LSILogic MegaRAID SATA 150-6D>.
The single one is attached directly to the SATA controller on the mainboard
-- <SiI 3114 SATA150 controller>.
If it matters, the system is FreeBSD/amd64. The motherboard supports two
Opterons, but only has one processor right now. The RAID card is of the
``PCI-X'' kind.
Thanks!
-mi
Pete French
> Is that to be expected? I know, I'm gaining the redundancy of the RAID5, but
> I expected the IO to be faster too -- if not 5 (6-1) times faster...
On writing you should expect RAID-5 to be a lot slower. Reading rather
dependson the controller and the stripe size. When this comes up on
other mailing lists the general conclusion is that FreeBSD doesnt
make large enough read requests to the disc to actually parallelise the
reading ofmultiple stripes(i.e. most requests fall in a single stripe)
so I would expect the maximum performance to be the same as a single drive.
Impressive if you get that performance when writing though...
-bat.
Mikhail Teterin
> Mikhail Teterin <use...@aldan.algebra.com> writes:
>> Is that to be expected? I know, I'm gaining the redundancy of the RAID5,
>> but I expected the IO to be faster too -- if not 5 (6-1) times faster...
>
> On writing you should expect RAID-5 to be a lot slower. Reading rather
> dependson the controller and the stripe size. When this comes up on
> other mailing lists the general conclusion is that FreeBSD doesnt
> make large enough read requests to the disc to actually parallelise the
> reading of multiple stripes(i.e. most requests fall in a single stripe)
> so I would expect the maximum performance to be the same as a single
> drive.
>
> Impressive if you get that performance when writing though...
Ok, it seems, iozone does not use a big enough file by default. On this
machine -- with 2Gb of RAM -- reading and writing speeds were largely
affected by the caching.
Once I switched to 2gb file size, the story changed. With short record
length (up to 8 bytes) the RAID is faster and with larger record length --
much slower in writing, than the single disk. Disappointing.
> When this comes up on other mailing lists the general conclusion is that
> FreeBSD doesnt make large enough read requests to the disc to actually
> parallelise the reading of multiple stripes(i.e. most requests fall in a
> single stripe) so I would expect the maximum performance to be the same as
> a single drive.
When reading, the RAID now hovers around 100Mb/second (declining somewhat
with higher record length), while the single drive sustains about
44Mb/second. Encouraging, but am I not supposed to see 4-6 times the read
speed of a single drive?
Interestingly, increasing the stripe size used by the RAID controller from
the default 64Kb to the maximum of 129Kb virtually killed the performance
of both reading and writing. I'll try the 32Kb now :-)
-mi
-mi
Pete French
> Once I switched to 2gb file size, the story changed. With short record
> length (up to 8 bytes) the RAID is faster and with larger record length --
> much slower in writing, than the single disk. Disappointing.
Thats what I would have expected. RAID-5 almost always slows things down.
> When reading, the RAID now hovers around 100Mb/second (declining somewhat
> with higher record length), while the single drive sustains about
> 44Mb/second. Encouraging, but am I not supposed to see 4-6 times the read
> speed of a single drive?
Not unless you can make a single request to the controller which is four
times larger than the stripe size. So if you have a stripe size of 1k then
making a 4k read will need to access 4 stripes and can thus run the four
discs in parallel (assuming the controller is smart enough to do that).
Anything below that wiill only need to read from one stripe and hence from
one disc so there is no speed up.
So whats the smallest sizeof stripe you can have oon the controller, and whats
the largest size of block that the reeBSD kernel can request from a device
driver at once ? I have a feeling that the latter is smething like 64K
so you would need a really small stripe size to get an increase.
-bat.