How to interpret data from "sar -d"

[pilog]

Hi!

I have difficulties in interpreting the data from "sar -d":
r+w/s is called "number of data transfers", blks/s is called "number of bytes transferred in 512-byte units"

Why is the ratio of these two values not constant? By which factors is the ratio influenced?

[Bernhard Mueller]

Hi,

a single read or write may affect different numbers of physical blocks on your disk.

Regards,
Bernhard

[Tom Geudens]

Hi,
HP-UX does not use a "fixed" blocksize (it depends on the application). For example, an Oracle database does 8K IO's, but an FTP may have much larger blocksizes. So it's not a ratio you are computing, but the average blocksize in a given interval.

Regards,
Tom Geudens

[pilog]

Hi!

Nearly conviced...
In my (too simple??) view of the disk layout, data are lying around in e.g. 8k-blocks. How is it possible for ftp to read with a bigger block size? Can't it be only a multiple of the file system block size?

Best regards,
Dieter

[Tom Geudens]

Hi,
We've recently had a lot of discussion (and consultancy) about this issue over here. Our mainframe system-engineers can "fix" the blocksize there OS writes in (32k seems to be their favorite ;-). On HP-UX you cannot. We've got EMC Symmetrix storageboxes and the WLA-tool shows us blocksizes ranging from 4k (which we can link to the Peoplesoft databases) to alsmost 200k (which is probably datatransfer). The point is that it is NOT fixed on HP-UX. It depends on the application (you can define the blocksize in a C-program, you know), it depends on buffers, it depends on ...
I'm very sure for the blocksizes of the Oracle databases (although there are parameters that can even change that ;-). For the other programs ... it depends.

Sorry I can't define it more cleary, but sar -d can give an "indication" of how big they are.

Regards,
Tom Geudens

P.S. Note that this has also impact on "responsetimes". An I/O of 8k and an I/O of 200k WILL have different responstimes, but that doesn't necessarily mean that they are bad ...

[Tim D Fulford]

Hi

I'm an Informix guy (not Oracle), but I'm sure they have similar features in this department. The 8k size mentioned is simply one page, this is the MINIMUM Oracle can shift. However, the database (or filesystem) can request in 1 IO to be far more data. This results in a system with a spectrum of block sizes which depends on how you use your system.

On our systems our AVERAGE block size is 2.5 kB (pritty small), but as the system is OLTP this is good as we need response/throughput [IO/s] more than we need bandwidth [MB/s].

Here in lies a common argument, "2.5kB is way too small"... "you will not get the bandwidth"... "you are performing too many IOs/throughput"... yadda de ya.

Bottom line, are you happy with the way the system functions on a day-to-day basis? If so great....If not...

Regards