Re: 2.6.0 performance problems

[William Lee Irwin III]

On Tue, Dec 30, 2003 at 04:14:13PM -0500, Thomas Molina wrote:
> I also get 90+ percent iowait under 2.6 and 0 iowait in 2.4.  I'm not sure 
> how the alleged suckiness of 2.6 paging fits into this.  On this system 
> the execution times are almost the same.  On this machine, in addition to 
> the iowait differences, there are cpu use statistics as reported by top.  
> On 2.4 idle time is 70 percent while on 2.6 the idle time is near zero 
> percent.  I'm not sure what the significance of this is.

2.4 does not report iowait; all iowait is reported as idle time on 2.4.

On Tue, Dec 30, 2003 at 04:14:13PM -0500, Thomas Molina wrote:
> CPU: PIII, speed 648.072 MHz (estimated)
> Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a unit mask of 0x00 (No unit mask) count 324036
> vma      samples  %           symbol name
> c0115e20 22498    22.6776     mark_offset_tsc
> c0110080 12707    12.8084     mask_and_ack_8259A
> c018eec0 7115      7.1718     ext3_find_entry
> c010ff60 4013      4.0450     enable_8259A_irq
> c0168d50 2650      2.6712     __d_lookup
> c015eb10 1727      1.7408     link_path_walk
> c010afd0 1482      1.4938     irq_entries_start

Well, it looks like Linus said various things along these lines in
various ways before I finished writing this, but in case hearing it a
second time is any reassurance:

There's a slight problem here in that you're io-bound, not cpu-bound,
so profiles won't actually tell us much about remaining overheads.

One thing here is that since turning off all the debugging options got
you down to about a 15% degradation, things aren't actually
looking anywhere near as problematic as before when you had a near 90%
degradation. One possible explanation is that the extensive padding
done by CONFIG_DEBUG_PAGEALLOC created significant memory pressure.

If you'd like further speedups, logging the things I suggested earlier
and trying fiddling with swappiness might help.

In fact, you are down to such a small margin of degradation that the
remaining degradation vs. 2.4 may in fact be due to using oprofile,
which has significant, though not overwhelming overhead.


-- wli
-
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/