[1465] in linux-scsi channel archive
Disk array latency theory (was: EATA performance issues)
daemon@ATHENA.MIT.EDU (Dario_Ballabio@milano.europe.dg.co)
Tue Feb 25 04:31:48 1997
Date: Tue, 25 Feb 1997 09:27:35 +0100
From: Dario_Ballabio@milano.europe.dg.com
To: linux-raid@vger.rutgers.edu, linux-scsi@vger.rutgers.edu,
lma@varesearch.com
%UATTACH
> In-Reply-To: <199702221835.TAA06022@milano.europe.dg.com>
> \011(Dario_Ballabio@milano.europe.dg.com)
> Dario,
>
> Thanks for your response. Here are the Iozone numbers for a single
> (non-RAID) disk on the DPT PM3334UW.
>
> Iozone Size Rec Length Read Rate Write Rate
> Directory Mbytes bytes Kbytes/sec Kbytes/sec
> /u1 244 512 7867.00 8267.90
> /u1 244 512 7795.82 8278.86
> /u1 244 512 7759.50 8298.11
> /u1 244 512 7752.28 8251.52
> Average 244 512 7793.65 8274.10
>
> This numbers look normal to me. They numbers are consistent with
> other Ultra Wide controllers like Buslogic, Adaptec, and Symbios.
> This is with a 2GB extfs filesystem with 1K blocks.
>
> Now if I configure 5 of these drives on this controller with RAID-5,
> 32M cache, and 32K stripes I get:
>
> Iozone Size Rec Length Read Rate Write Rate
> Directory Mbytes bytes Kbytes/sec Kbytes/sec
> /u1 244 512 4935.91 1742.86
> /u1 244 512 5825.51 1950.63
> /u1 244 512 4908.76 2145.05
> /u1 244 512 4945.69 2147.08
> Average 244 512 5153.97 1996.41
>
> As you said, the disks are not spinning in sync so there will be a
> latency waiting for the last disk. Shouldn't this just be 1/2 a
> rotational delay on average?
>
> I haven't tried the rev. 3.00.00 driver yet. I will let you know what
> results I get with it.
>
> Thanks,
>
> Larry
>
>
Just a little bit of theory about disk latency. If all the disks
in an array are spinning in phase (i.e. sector 0 of each track
is under the head at the same time for all the disks) the average
rotational latency is always T/2, where T is the revolution time.
Disk syncing is a feature supported by some high quality disk arrays
like the CLARiiON (see http://www.clariion.com).
When n disks are not in phase, the completion times are random
variables t1, t2, ....tn uniformly distributed in the time interval [0, T].
If Ln is the latency for n disks, the probability of a latency
lower than t is:
n
P(Ln <= t) = (t / T)
The average latency for n disks is, by definition, the average value
of the random variable Ln over a period, so:
T T n-1
/ / nt n
E(Ln) = | tdP = | t ----- dt = ---- T
/ / n n+1
0 0 T
So, for example for RAID 1 (mirroring with 2 disk images), n = 2 and
the average latency is (2/3)T.
When increasing n, E(Ln) -> T, so when using a real RAID configuration
one should always expect a full revolution latency, instead of the
half revolution of a single disk.
+--------------------------------+--------------------------------------+
| Dario Ballabio, | Phone : +39 2 580031 |
| Technical Director, | Fax : +39 2 58003258 |
| Data General Italy | Mob. phone: +39 3356322300 |
| | Addr. : via Mecenate 90 |
| | 20138 Milano Italy |
| | Email: dario@milano.europe.dg.com |
| | |
| | "si vis pacem, para bellum." |
+--------------------------------+--------------------------------------+
