Threshold for Disk IO

[Microsoft Technologies]

Do we have standard values for disk IO counters?

[Leythos]

In article <ae32161a-bb55-451f-b268-6309923a39b5
@x21g2000yqa.googlegroups.com>, ***@gmail.com says...

Post by Microsoft Technologies
Do we have standard values for disk IO counters?

If you're talking about average disk R/W queue, it's normally the number
of platters (some use the number of drives).

So, if you have a 5 drive RAID-5 array, if you see an average READ or
WRITE queue for the array of 5+ for any real length of time, it normally
means you're system is DISK bound.

[Dave Warren]

Post by Leythos
In article <ae32161a-bb55-451f-b268-6309923a39b5
@x21g2000yqa.googlegroups.com>, ***@gmail.com says...

Post by Microsoft Technologies
Do we have standard values for disk IO counters?

If you're talking about average disk R/W queue, it's normally the number
of platters (some use the number of drives).

Platters or spindles? Many drives have multiple platters... Spindles
and drives are (I believe) effectively the same.

Post by Leythos
So, if you have a 5 drive RAID-5 array, if you see an average READ or
WRITE queue for the array of 5+ for any real length of time, it normally
means you're system is DISK bound.

To me, it's a bit more complicated than that. For reads, in a RAID-0 or
RAID-5 array, I'd agree with that but for RAID-1 (and 10/01) it depends
whether the controller uses the redundant copy for reading (in which
case count all the spindles) or uses the redundant copy to detect bitrot
(in which case divide by 2)

For writes, however, RAID-5 is a lot more complicated since there are
multiple reads and writes involved and it really depends on how smart
your controller is as to how long a write actually takes, but at a
minimum every RAID-5 write is one read and two writes across at least
two (and possibly three) spindles. So for RAID-5 I'd probably divide by
3, and for RAID-1/10/01 divide by 2 (since again, two spindles are
writing)

However, it's also possible that the disk IO counters take some or all
of this into consideration so I may be way off base.

[Leythos]

Post by Dave Warren

Post by Leythos
In article <ae32161a-bb55-451f-b268-6309923a39b5
@x21g2000yqa.googlegroups.com>, ***@gmail.com says...

Post by Microsoft Technologies
Do we have standard values for disk IO counters?

If you're talking about average disk R/W queue, it's normally the number
of platters (some use the number of drives).

Platters or spindles? Many drives have multiple platters... Spindles
and drives are (I believe) effectively the same.

Post by Leythos
So, if you have a 5 drive RAID-5 array, if you see an average READ or
WRITE queue for the array of 5+ for any real length of time, it normally
means you're system is DISK bound.

To me, it's a bit more complicated than that. For reads, in a RAID-0 or
RAID-5 array, I'd agree with that but for RAID-1 (and 10/01) it depends
whether the controller uses the redundant copy for reading (in which
case count all the spindles) or uses the redundant copy to detect bitrot
(in which case divide by 2)
For writes, however, RAID-5 is a lot more complicated since there are
multiple reads and writes involved and it really depends on how smart
your controller is as to how long a write actually takes, but at a
minimum every RAID-5 write is one read and two writes across at least
two (and possibly three) spindles. So for RAID-5 I'd probably divide by
3, and for RAID-1/10/01 divide by 2 (since again, two spindles are
writing)
However, it's also possible that the disk IO counters take some or all
of this into consideration so I may be way off base.

If you notice, I didn't mention RAID-0, 1, 1+0 or 0+1.

The same would hold true, if you look at how reads/writes are processed.
While some controllers have very nice settings to prioritize
reads/writes, caching, etc... If you're numbers for Avg Disk Write Queue
are hovering around 10 and you have 3-5 disks in your array (I picked 10
out of the air for discussion) then you've got a disk bottleneck.

[Dave Warren]

Post by Leythos

Post by Dave Warren

Post by Leythos
In article <ae32161a-bb55-451f-b268-6309923a39b5
@x21g2000yqa.googlegroups.com>, ***@gmail.com says...

Post by Microsoft Technologies
Do we have standard values for disk IO counters?

If you're talking about average disk R/W queue, it's normally the number
of platters (some use the number of drives).

Platters or spindles? Many drives have multiple platters... Spindles
and drives are (I believe) effectively the same.

Post by Leythos
So, if you have a 5 drive RAID-5 array, if you see an average READ or
WRITE queue for the array of 5+ for any real length of time, it normally
means you're system is DISK bound.

To me, it's a bit more complicated than that. For reads, in a RAID-0 or
RAID-5 array, I'd agree with that but for RAID-1 (and 10/01) it depends
whether the controller uses the redundant copy for reading (in which
case count all the spindles) or uses the redundant copy to detect bitrot
(in which case divide by 2)
For writes, however, RAID-5 is a lot more complicated since there are
multiple reads and writes involved and it really depends on how smart
your controller is as to how long a write actually takes, but at a
minimum every RAID-5 write is one read and two writes across at least
two (and possibly three) spindles. So for RAID-5 I'd probably divide by
3, and for RAID-1/10/01 divide by 2 (since again, two spindles are
writing)
However, it's also possible that the disk IO counters take some or all
of this into consideration so I may be way off base.

If you notice, I didn't mention RAID-0, 1, 1+0 or 0+1.

No, you didn't not. Nor did the original poster mention RAID-5 at all.
What's your point?

Your numbers make sense for RAID-0 reads and writes, and possibly for
RAID-5 reads, whereas RAID-5 writes are far slower and so the IO
counters might need to be interpreted differently in this situation.