Average IOPS for SATA, SCSI, SAS, FC, FATA Drives

[Patrick Terlisten]

Hello,

I read the quickspecs for the above listed drives and I calculated the average IOPS for differents disks. How realistic are these values?? Could someone check them?

2,5" 10.000 rpm SAS 113 IOPS
2,5" 15.000 rpm SAS 156 IOPS
3,5" 15.000 rpm SAS 146 IOPS

2,5" 5.400 rpm SATA 71 IOPS
3,5" 7.200 rpm SATA 65 IOPS

3,5" 10.000 rpm U320 104 IOPS
3,5" 15.000 rpm U320 141 IOPS

3,5" 10.000 rpm FC 125 IOPS
3,5" 15.000 rpm FC 150 IOPS

3,5" 10.000 rpm FATA 119 IOPS

Thanks for advice.

Kind Regards,
Patrick

[Stan Sieler]

The answer is "it depends"
It depends upon your CPU, I/O backplane, the disk
drives themselves, and the size of each I/O.

On an rx2600 (IPF, 1.3 GHz clock), with
HP 36.4GMAS3367NC drive (UltraWide LVD SCSI)
I get:

| Reads/sec | MBs/sec (#pgs = 1)
Test | Best Average | Best Average
------------| --------- ---------| --------- ---------
Random | 175.4 r 173.1 r| 0.7 M 0.7 M|
Same | 5260.0 r 5234.8 r| 20.5 M 20.4 M|
Serial | 8560.0 r 8240.0 r| 33.4 M 32.2 M|
Steady1MB | 617.5 r 617.5 r| 2.4 M 2.4 M|

Random...reads from random blocks
Same...reading the same block over and over
Serial ... reading sequential blocks
Steady1MB ... reading mostly sequentially,
skipping one MB on disk between each read
(trying to defeat read-ahead caching)

For random reads of a single page (4KB),
I got about 173 reads per second,
faster than all your estimated drives,
and that's on a several year old computer/drive combo).

That said, your numbers look like reasonable
values to use to compare kinds of drive interfaces.

[Patrick Terlisten]

Hello,

thanks for you reply. I calculated the average IOPS just for me, to have a clue, when I need to calculate IOPS for a disksystem or anything else. I read many times that a 15k upm U320 SCSI disk drive got 150 IOPS, or that you should calculate with 150 IOPS. But no one could tell me IOPS for other drives, so I calculated them on myself. No I'm interested for feedback, how realistic these values are.

Kind regards,
Patrick

[Stan Sieler]

My point was, in part, that you can't
calculate meaningful IOs per second without specifying
more information: sequential access?
random access? how big an I/O? read or write?

At least you have one data point: with one
kind of SCSI disk drive (model number posted earlier,
unknown speed), on an rx2600, I was able to do about 173
random reads a second (reading a block of
4096 bytes).

On a slightly slower machine, 650 MHz PA-RISC, I get:

--------------------------------------------
Opened disk /dev/rdsk/c1t15d0 (DISK 1) [HP 18.2GMAN3184MC]
cRandom of DISK 1; 1 ppr; from 1 : 124.2 Reads/sec; 0.5 MBs/sec
Same of DISK 1; 1 ppr; from 1 : 3073.9 Reads/sec; 12.0 MBs/sec
Serial of DISK 1; 1 ppr; from 1 : 2656.7 Reads/sec; 10.4 MBs/sec
Steady1MB of DISK 1; 1 ppr; from 1 : 429.7 Reads/sec; 1.7 MBs/sec
-----------------------------------------
Opened disk /dev/rdsk/c3t15d0 (DISK 3) [SEAGATE ST373207LC]
cRandom of DISK 3; 1 ppr; from 1 : 165.4 Reads/sec; 0.6 MBs/sec
Same of DISK 3; 1 ppr; from 1 : 4555.4 Reads/sec; 17.8 MBs/sec
Serial of DISK 3; 1 ppr; from 1 : 5574.4 Reads/sec; 21.8 MBs/sec
Steady1MB of DISK 3; 1 ppr; from 1 : 590.6 Reads/sec; 2.3 MBs/sec
--------------------------------------

The ST37320LC is a 10,000 RPM drive.

Hope that's of interest.

Stan

[Patrick Terlisten]

Hello Stan,

thanks for you reply. You're right, however there must be a way to get to a theoretical value. For example:

I need at least 3000 IOPS for an application. So I will nee at least 20 Drives, if one drive will give me 150 IOPS.

I read many times that one disk drive will deliver roundabout xxx IOPS.

Take a look at this thread (http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=1115419), especially the posting from John Kufrovich.

Kind regards,
Patrick

[Greg McDonald]

Patrick,

I was given this link for your IOPS #s. Has anyone checked/confirmed these calculations. If not, how would you suggest they be checked/confirmed? Can they be tested on actual (with all the various dependencies)?

Thanks,
Greg

[Stan Sieler]

Greg asks:
> Can they be tested on actual (with all the various dependencies)?

Our DiskPerf tool produces the numbers
I quoted earlier. Email me at sieler@allegro.com if you want info about it.