RE: [Xen-users] disk backend performance

["James Harper" <james.harper@xxxxxxxxxxxxxxxx>]

> 
> So: over a dedicated cable with jumbo frames it is actually smarter to
use
> ISCSI than AOE? Is that your conclusion?

My conclusion was more along the lines of "don't assume that just
because iSCSI uses TCP/IP that it is inferior performance-wise to AoE".

I don't believe that there currently exist any hardware implemented AoE
adapters, but there are hardware iSCSI adapters (HBA's).

With a software implementation, Ethernet headers have to be constructed
per packet (+IP+TCP for iSCSI), error detection has to be implemented
for AoE, and packetization has to be done in software (less so for iSCSI
as most hardware has TCP Large Send Offload which from a software point
of view allows sending of 64k TCP packets). Any network adapter that
you'd use in a server these days has TX and RX checksum offloading
available, so you get checksumming of your TCP packets for free when
using software iSCSI, while for AoE you need to calculate checksums
manually in software.

With a hardware iSCSI HBA, the O/S just has to say to the card 'read x
sectors starting at y and put them in memory here'. The card does the
rest. Of course, to do this you have to buy an iSCSI HBA, which
increases the cost of the solution somewhat.

If you want to use non-disk devices then iSCSI is probably a better
choice - a robotic tape library is more likely to work over iSCSI
without problems than it will over AoE, although my information on that
sort of thing may be out of date...

A protocol called HyperSCSI exists, which is basically SCSI over
Ethernet, but I don't know how available it is.

James

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