RE: [Xen-users] AoE vs iSCSI

[Jeff Sturm <jeff.sturm@xxxxxxxxxx>]

> -----Original Message-----
> From: Markus Hochholdinger [mailto:Markus@xxxxxxxxxxxxxxxxx]
> Sent: Thursday, March 18, 2010 10:17 AM
> To: Jeff Sturm
> Cc: xen-users@xxxxxxxxxxxxxxxxxxx
> Subject: Re: [Xen-users] AoE vs iSCSI
> 
> Example: One server with vblade exported one block device over 1GBit/s
NIC. On the
> other server, the client, i got ~100MByte/s as expected. If i
configured 10 vblades on
> the server, connected all 10 to the client and then testet one ether
device, i got only
> ~20MByte/s throughput. Then i configured 100 and got only ~1MByte/s
throughput!

Unfortunately vblade is little more than a toy program.  Its beauty lies
in its simplicity.  Its drawbacks lie also in its simplicity.  It's nice
to have as a reference program for those who want to tinker with the
protocol or understand how AoE works, but you can't really draw any
conclusions about performance of the AoE protocol from using vblade in
general.

Vblade is single-threaded and can only issue one outstanding I/O request
at a time per device.  Multiple vblade processes can run on the same
adapter to export multiple disks, and cooperate via packet filtering.  I
don't know if the packet filtering was working optimally in the version
you tested... based on your results I could guess it was not.

You should have better overall results testing with another open source
implementation like qaoed, or using Coraid's commercial product.  I have
routinely demonstrated 200MB/s throughput performing sequential
transfers on an AoE target, multipathing over 2 GigE adapters.

There's nothing wrong with iSCSI either, and many users have perfectly
valid reasons to require iSCSI.  But you can get comparable performance
with AoE for Linux hosts, often spending far less.

This is getting a bit off-topic for a Xen list, I'm afraid.

-Jeff



_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users