Processor idle time with too few disks

[Pritch5000]

Does anyone know info about the optimum ratio for the number of processors vs hard disks to ensure minimum processor idle time waiting for I/O requests.

thanks

 

[xmario2013]

i think its all depend on the spec of the application you are running on the server

 

[Pritch5000]

Hi. Thanks for the reply.

Let me clarify. I have a server with 8x Xeon processors and 8x 36GB 15K and to get optimum performance out of 8x Xeons, I need about 50 disks on a 6 disks per processor calculation.

I have been told that with so few disks the processors will have massive idle times while waiting for responses from the disks.

By adding more disks to house the same data, you are adding extra heads and plates meaning faster access to the same data.

The trouble with that calculation is that it is quite common to have dual Xeons in a server with just 2 disks RAID1, but then my server is for an MS SQL Server backend.

Can anyone add their opinions (or preferably research!) to this issue?

 

[NickFerrar]

Never seen a calculation like that before, personally I'd have thought it has more to do with the I/O controllers you are using (and how many channels they have) rather than trying to match a number of physical disks to CPUs. You wouldn't partition it as a single array anyhow for best performance.

Also what manufacturer is the box? Dell stopped selling 8-way systems as generally it's quicker to cluster 2 4-way system than have 1 8-way as the CPU's starting interfering with each other.

 

[Pritch5000]

Thanks for your comments.

Its actually a HP (Compaq) server - i know i know, this is a Dell page! but it wasnt really a proprietor question, and and I posted it in the Compaq forum too.

what you have said is what I had thought before someone warned me that my processors will be sitting idle waiting for I/O requests.

Currently I have 2x RAID controllers (with 2c and 128MB BBC) running 4 disks each (4 int and 4 ext). Ultimately there will be one of these controllers running the internal 4 disks and redundant fibre raid controllers for the SAN.

I am currently investigating the best way to divide up the functions of the database server (os, data, logs, pf, etc)

thanks again for your comments, any more would be gratefully appreciated.

 

[johnny99]

I have also never seen this kind of calculations.
If you have a high I/O load go for a faster controller with as much cash as you can get, and then go for the number of disks.

With a server of the size you have, go for a fast disksubsystem.
We use StorageTek D280 disksystem and this is fast.
We had a lot of talk with StorageTek of the number of disks and how fast they should be.
We finaly ended up with 146 GB disks at 10K rpm not going for the 73GB disks at 15K rpm.
We got half number of disks at a lower rpm and what StorageTek told us was that in our installation we would never see the difference (except of the lower cost per GB)
But the main difference is in the controller. Had we gone for the D240 model (it uses the same disks) it would have been a different story.
A very very good high end controller can give you a hit rate in the cache of 95% so you end up with less then 5% of the I/O going directly to the disks.

And if you need very fast I/O always go for Raid 10.

To give you a better direct advise I need to know a little more about the problem you and your users see.
A poor application on a bad tuned SQL will always perform poorly even on the best hardware money can buy.

By the way, what disksubsystem do you use on your SAN?

/johnny