If anyone is doing some simple I/O tests, like dd, for example with
iSCSI and noticing abysmal performance I just found something.
Internally someone was setting up a machine with iSCSI and noticed
the following:
time dd if=<iscsi_target> of=/dev/null bs=1k count=200
reported time would be 12 seconds.
time dd if=<iscsi_target> of=/dev/null bs=8k count=200
reported time would be less than a second.
For the test the iSCSI target backing store was a file in /tmp, so
it's a ram disk. The link between the two machines is a 10GbE.
Clearly something odd was happening since there's no way 1k reads
should be slower than 8k reads. Well, after poking around a little
bit it turns out this simple test case is showing why the "Nagele
Algorithm" is not a good fit for iSCSI. By disabling this feature the
times for the 1KB case are just a fraction of a second quicker than
the 8KB which is expected.
I just wanted to point this out to folks in case you've noticed the
problem and where wondering why or saw this problem and thought the
Solaris iSCSI Target was a piece of junk. I'll be fixing this with my
next set of changes.
----
Rick McNeal
"If ignorance is bliss, this lesson would appear to be a deliberate
attempt on your part to deprive me of happiness, the pursuit of which
is my unalienable right according to the Declaration of
Independence. I therefore assert my patriotic prerogative not to
know this material. I'll be out on the playground." -- Calvin
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss
