[Iscsitarget-devel] Effect of read-ahead on the volume exported from ietd / cpu usage and iowait

[Pasi K. <pa...@ik...>]

Hello list!

I'm trying to understand what's the problem with my setup.. 

I have a lvm volume (from a single SATA disk) exported with 
ietd v0.4.16 like this:

Target iqn.2001-04.com.test:vol1
        Lun 0 Path=/dev/vg1/test1,Type=fileio
        InitialR2T              No
        ImmediateData           Yes 

I'm accessing the volume with open-iscsi (CentOS 5.1) initiator. 

I'm seeing weird performance changes when I change the value of read-ahead
for this volume on the initiator.

Please see this thread on open-iscsi list for more information and the
benchmark values:

http://groups.google.com/group/open-iscsi/browse_thread/thread/79d10a3dfa1a453

My latest post today shows the CPU usage information on both the initiator and
on the target.

Basicly I'm seeing that setting the read-ahead value to 4096 (2 MB) on the
initiator causes the CPU on the _target_ being maxed out and the performance 
drops to ~half.. most of the time will be spent on iowait on the target.

Bigger read-ahead on the initiator seems to cause more cpu usage (more iowait) 
on the target..
 
Does someone know why this happens and is there some way to prevent it? Some
settings to change on ietd? 

Target is running CentOS 5.1 x86 (with all the latest updates installed).
Target hardware is Intel P4 3.0 GHz (hyperthreading on) with serverworks
chipset and broadcom tg3 NIC.

Thanks!

-- Pasi