disk IO bendwidth measure

[nanan]

Hi
I want to know how to make decision the disk IO reach the limt of IO bendwith.
there are many factors with which we can analyze the IO performance and status.
ex) IO rate, IO utilization on glance, IO read/write size
what i want to say whether i can got a way to anticipate the limit of IO bendwidth like CPU and memory.

If i can measure the amount of IO rate, is it really useful to prospect the limit of IO.
If yes, is there a criterion of IO rate.
how many IO rates can be able to be allowed on the differnt system and storage
please give me an advice with other things i didn't consider

Best regards


Those resources can see the utilization and usage so it is easy to.

[RAC_1]

The io bandwidths would be detailed in device documents. It rarely happens that these are hit.

I monitor with combination of iostat and sar -d. Looking at these two values over a period of time gives us information about which disks are heavily used and queues on it. Over a period of time, it helps to identify high hitting disks and spread IO over multiple disks. (We consider % busy and run queue and wait and serv times)

[nanan]

Thanks

As you comment, it is useful to gather the data with the commands you recommanded.
I am able to know if disk IO biased some disks.
But what i saying
For exmaple, i am operating a DB system using XPstorage or EMC storage
and the system installed 2FC 2GB cards and connected to XP of EMC FC ports.
there is running disk IO and that is in the bandwidth of disk IO.
As you know some LUNs in the storage and the DB system generate disk IO through FC path. the more time the more increase disk IO, i can't wait for the time when disk bottleneck is happened at LUNs.
Before such event happen to the system, i should do some thing to avoid.

For that i need to know a way to anticipate Disk IO with IO rate.
the reson why i am considering IO rate is this factor is not chnaged, while data trasfer size is not constant, isn't it?

If my thought is incorrect, please give me a solution of my concern

Best Regards

[RAC_1]

There is not a tool (that I know or have come accros) that would anticipate such event. I think it is still a combination of your historical data and database stats that would give you information as to how load is inceasring/decreasing and when you might run into a problem.

Now there is somwthing called prospect (i think HP tool) that might prove helpful, but I have not worked on it.

[Bill Hassell]

It's very important to understand the fdifferenvce between bandwidth and throughput or I/O rate. They are NOT the same! Bandwidth usually means the megabytes per second whereas throughput refers to how much data is actually transmitted or received. The difference is the transaction overhead. If your application writes 1 Kbyte per record, the overhead will be ENORMOUS and throughput will be terrible. If you change to write 10 Mbytes per record, the overhead is very small and overall throughput will be close to the maximum link speed.

All the above assumes that the disks are 100% cached in RAM so there is no seek time. Change the data patterns to scatter the data all over the disks and now seek time enters the equation.

So about all you can use is the sar -d command to monitor the disk queue (always desirable to be below 1) and the service time (less than 10 is desirable). When the seek times go above 10 (on a typical array) then the data is not optimally cached and more array cache is needed. When the disk queue goes up, requests cannot be processed as fast as they are received, indicating a possible array controller overload and/or long service times.

[nanan]

I think that FC cards bandwidth is 2GB so IO to disk can be close to the limit
Never can over the bendwidth, alright?

I can estimate the IO rate in certain amount of time. sar -d
That mean I just need to know the size of a IO rate.

I want to analyze disk IO with IO rate
but i don't know how to estimate the size of a IO

please let know the way to estimate the size of a IO

Best Regards

[Yogeeraj_1]

hi nanan,

Glance plus allows you measure IO as follows:
i - IO By File System
u - IO By Disk
v - IO By Logical Volume

hope this helps too!

kind regards
yogeeraj

[Bill Hassell]

You asked:

> I think that FC cards bandwidth is 2GB so IO to disk can be close to the limit
Never can over the bendwidth, alright?

Correct. But the physical disks are extremely slow compared to the bandwidth of the FC channel. A typical disk can supply about 10 to 100 Mbytes/sec (100 to 1000 Mbits/sec). The problem is that the disk does not do any transfers when it is seeking. So disks will never come close to the speed of the link.

But the disk data is buffered in the array controller with lots of RAM which is faster than the link. So transferring data to/from the array's RAM will indeed exceed the bandwidth of the link. And since the RAM is smaller than all the disks, only recently accessed data will be available from array RAM.

> I can estimate the IO rate in certain amount of time. sar -d
> That mean I just need to know the size of a IO rate.

sar -d is anj average for the measurement period but it does NOT measure I/O's. The blks/sec is computed by sar from the transfers that took place during the measurement period. If it says 100 blks/sec, there may have been 2 large I/O's or 250 small I/O's.

> I want to analyze disk IO with IO rate
but i don't know how to estimate the size of a IO
> please let know the way to estimate the size of a IO

You cannot measure the size of each I/O without extensive knowledge of tghe kernel and programming skills similar to those needed to create the sar program. Once you have the list of I/O's (and it can be enormously large in just one second (hundreds to thousands), the average isn't really useful because large I/O's result in lower overhead per I/O, while small I/O's result in much larger overhead for the same total data.

[The Old Man]

A couple of things you might want to look at in your quest to get a handle on IO:

www.iozone.org
Free utility, has a tarball for HP-UX. Good for learning how many variables go into typical IO. May spark some ideas.

www.sarcheck.com
Tuning/performance software (not free). Has output examples.

Good Luck,
Tom