Re: Terrible interactivity with 2.6.0-t9-mm3

From: Nick Piggin
Date: Mon Nov 17 2003 - 00:21:49 EST




Gene Heskett wrote:

On Sunday 16 November 2003 23:19, Nick Piggin wrote:

I think you might have confused Andrew a bit more ;)

To start with, you are talking about IO schedulers, while the thread
is about CPU interactivity.


I wasn't aware that such performance issues were divorced. In my admitted limited view, a laggy mouse is a laggy mouse, and its due to irq latency in achieving the context switch to service the mouses data. To me, its sorta like 2=2. :)


I guess its mostly due to scheduling latency. I think IRQ latency is
generally very good these days.

Scheduling latency caused not by long critical sections in the kernel,
but having things scheduled in front of you (X) for a long time.


The problem here looks like something that is caused by something in
mm3, not in mm2, not linus.patch, and not
context-switch-accounting-fix.patch.


Off topic: it would be good if you could try the as disk scheduler
in mm3. I recall you had some problems with it earlier, but they
should be fixed in mm3. Thanks.


Ok Nick. I'll reboot tomorrow without the elevator argument. Right now, amanda is fixin to be fired off in about 25 minutes and I want to see how badly its estimate phase hogs the machine using the cfq scheduler. With the -mm2 as, it was almost psychedelic to watch the mouse move.


OK thanks. I think this problem you were seeing _was_ interrupt latency
due to AS doing millions of WARNs. It should be fixed in mm3.


Also off topic re mouse performance, and I expect this is an X issue, but when its been blanked because I'm typing, it takes about a full seconds worth of hand waving before it becomes visible again. This is an X issue and I should go away, right?


Sounds like it.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/