iSCSI stall/pause/disconnect workaround

[jgreco]

So some people might have noticed that I've looked extensively at the iSCSI disconnect issue many people have seen. The basic problem is that ZFS can get too busy and pause for many seconds at a time, causing iSCSI disconnections.

I've chronicled this for some time in Bug #1531 but at this point I think I'm pretty much done. The answer is to significantly reduce the TXG sizes possible, which allows ZFS to manage the size but to do so within a reasonable window, and then to decrease the time intervals involved.

You will lose SOME write performance in doing this, but you will gain responsiveness like you won't believe.

 

[cyberjock]

If we were smart we'd bookmark this, then when someone posts that their iscsi machines keep stalling/pausing/disconnecting point them to this thread.

 

[jgreco]

Just refer them to 1531. That's why I keep posting stuff to it when I find something useful.

By the way, this topic is not JUST for iSCSI. It works for ANY protocol where you'd like to increase responsiveness and are willing to lose some performance.

 

[Stephens]

And/or put it in the FAQ.

 

[jgreco]

Bear in mind that I don't mind keeping one thing up to date with my findings; as of right now, that's #1531. I'm fine with someone moving that to a FAQ and pointing 1531 there, of course, but in the (at this point unlikely?) event I discover more information, the ability to make additions or changes without substantial work would be advisable.

Isn't there a wiki around here somewhere?

 

[bollar]

Thanks for your great work on this!