Blake wrote:
I did some reading on this. Nagle was originally trying to stem
the flood of 'tinygrams' sent during lightweight network sessions
such as telnet (where single bytes (keystrokes) were being given
their own packet, and generating lots of overhead). It's my
impression that since many of us are using iSCSI to move large
files, we aren't likely to see the algorithm affect us in the way
previously described. Is this perhaps why the algorithm is enabled
by default?
Please take a look at the node where the iSCSI Initiator is running
and determine if the following parameter is being set.
[ /kernel/drv/iscsi.conf ] add the following:
#
# Disable Nagle within the iSCSI Initiator
#
tcp-nodelay=1;
The tcp-nodelay setting, alters the iSCSI Initiators socket options
by setting the well-known option TCP_NODELAY, or more specifically
turning off the "Nagel Algorithm". For performance reasons, this
setting is hard-coded in the iSCSI Target and also Availability
Suite's SNDR, both being TCP/IP based data movers of disk blocks.
For both iSCSI and SNDR, the Nagel algorithm interacts badly when
using TCP delayed acknowledgments, of which both iSCSI and SNDR
relies upon. These performance issues have been noted by many others,
including the author John Nagel. See link at bottom of this page:
http://en.wikipedia.org/wiki/Nagle's_algorithm
Jim
Blake
On 10/24/07, Jonathan Edwards <[EMAIL PROTECTED]> wrote:
On Oct 24, 2007, at 14:07, Patrick Sullivan wrote:
> I recently posted about a serious performance issue when running a
> single thread against my Solaris iSCSI Software Initiator connected
> volume. The cause of this performance issue is due to the Nagle
> algorithm being used for the network traffic sent to and from the
> iSCSI target. Simply disabling the Nagle algorithm solved the
> problem.
>
> Easy enough fix, but it raised a question. If having the Nagle
> algorithm enabled could cause this serious a performance
> degradation (300k/sec I get with it enabled, 40MB/sec with it
> disabled), should it be disabled by default?
>
> I guess I would be curious as to the reason it is enabled by
> default. Does the performance advantage on some targets outweigh
> the performance degradation seen when it is enabled for other
> targets? 300k/sec is pretty dreadful performance, and for most
> applications it could be considered unusable. Granted my testing
> was to the raw device, but any application that accesses the raw
> device directly is also affected by this (think Solaris Volume
> Manager, and I wonder how ZFS would fair, I can test if necessary).
>
> How would I go about filing a request to change the default to not
> use the Nagle algorithm?
File an RFE here:
http://www.opensolaris.org/bug/report.jspa
pick the latest nevada release
---
.je
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss
