iostat bug 4962

Hi Jean ,

Thanks for your reply.

-I am not sure it is impossible. I have four 10,000rpm Ultra-320 SCSI hard -drives (and two 7,200 rpm EIDE hard drives on the same channel) on my - -machine. If I had eight, and stuck four on each SCSI -controller, I imagine I might get it up there if the kernel and the -applications could keep up with it. If I spun off 8 processes copying hard -drives todev-null, it might work. I have two hyperthreaded 3.06 Xeon

-processors with 1 MByte L3 caches, so they can probably keep up with the -interrupt load.

I use EMC disk array , and i have 1 HBA with fibre channel. This can provide 100MB. per second io.

-Does vmstat tell you the same thing? It is unlikely that both are in error -unless the kernel is reporting incorrectly.

I dont check the vmstat part but i checked sar -B which shows page in rate. I see the same values. 500MB. data is paged in per second.(Probably this is file io for page cache)

-As a practical matter, I seldom see over 50 Megabytes-second transfers with -the disk drives, but I see that fairly often.

My test sql is simple : select count(*) from tablea

The table size is 1gb.

When i run this by using aio , (which shows reasonable values in iostat , 100MB.read per second) it takes 15 sec. ( This is also expected with this io rate)

On the other hand , for the above case , it takes 30 sec. If 500mb. iorate is correct , this must take 3-4 secs.(including the overhead of file system)

Comparing the r-s values , the above output says :

r-s : 14344.44 rkB-s : 568785.19 %util : 1851.85

Issuing 14344.44 read request per second? Isnt 200-250 read per second io enough to saturate a scsi disk ? (The %util value is also strange)

And , according to the iostat

avg-cpu: %user %nice %sys %idle 20.83 0.00 68.98 10.19

%sys is 68.98.

is generated.

Most importanly , as i sad above , it takes longer .

This is a datawarehouse system . Most of the queries are full table scans and i always use parallel query .

The datafiles are not raw , i use ext3. ( This is development )

Direct-io which bypbuttes the page cache and puts data directly to the user process , is supported but not enabled on my server dues to an oracle bug.

I tuned the oracle in order to increase the perfomance of full table scans.

Oracle on each io operation via pread or kio , requests 1MB. data from the operating system.Oracle parallel query bypbuttes its own cache , so each parallel query requests data from the operating system.( Whether this data is supplied from the page cache or from the disk array depends on the io type.) I dont use directio , so page cache is used for my data files. I think this explain the sar -B page in values.

So , i still think the possibility of iostat problem .

Kind Regards, tolga tolga