Performance of 4.x vs 5.x (Re: Lifetime of FreeBSD branches)

[Kris Kennaway]

On Wed, May 25, 2005 at 06:44:37PM +0200, Matthias Buelow wrote:
> Kris Kennaway wrote:
> 
> >>interrupt                          total       rate
> >>irq1: atkbd0                         586          0
> >>irq13: npx0                            1          0
> >>irq14: ata0                           94          0
> >>irq17: wi0                            54          0
> >>irq20: fxp0 atapci1                62079         99
> >>irq21: uhci0 ehci0                     1          0
> >>irq22: uhci1                        1102          1
> >>lapic0: timer                    1246549       1994
> >>lapic1: timer                    1246427       1994
> >>Total                            2556893       4091
> >>The only relevant conflict I could see is irc 20; but I had already 
> >>tested that by removing fxp0 from the kernel.
> >
> >I wonder if USB is causing the problem all on its own..since that was
> >the culprit in other situations when it was being triggered by virtue
> >of interrupt sharing.  Any chance you can try a non-USB mouse and
> >remove USB from your kernel?
> 
> Ok, now USB (both uhci and ehci) is gone.  The problem is still the 
> same.  vmstat -i:
> 
> interrupt                          total       rate
> irq1: atkbd0                        1324          3
> irq12: psm0                         8562         21
> irq13: npx0                            1          0
> irq14: ata0                           94          0
> irq17: wi0                           381          0
> irq20: fxp0 atapci1                61956        154
> lapic0: timer                     801433       1993
> lapic1: timer                     801292       1993
> Total                            1675043       4166
> 
> To be frank, I do not believe it's got anything to do with locking or 
> interrupts.  It somehow seems just like the scheduler is doing a bad job 
> of balancing interactive processes vs. disk i/o.  I've seen the same 
> stuff for years on NetBSD (until they changed scheduling around 1.5 or 
> so) and Linux (until 2.4 kernels).  During that time FreeBSD didn't 
> exhibit these symptoms and only in 5.x have I seen that kind of 
> behaviour creep back in.  Has the classic scheduler been changed 
> somehow?  Maybe I should try and see if the problem persists with the 
> ULE scheduler?

Others don't see this though, and in other cases it was *definitively
proven* to be caused by the issue I mentioned.  I'll have to think
more about what to try next..thanks for running the tests.

Kris
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-stable/attachments/20050525/f7f4da00/attachment.bin