Re: PAE and balloon benchmarks

[Jean-Yves Migeon <jeanyves.migeon%free.fr@localhost>]

On 13.07.2010 02:23, Thor Lancelot Simon wrote:
> On Mon, Jul 12, 2010 at 06:09:41PM +0300, Antti Kantee wrote:
>>
>> * Is the cost expected?  Naiively thinking, 15-20% seems quite high.
> 
> It sure does.  Isn't the amd64 MMU, for example, always 3-level in
> approximately the same way?  And, architecturally, almost all the
> current-production processors implementing PAE are 64-bit capable,
> which means the 64-bit atomic ops should be no costlier on the underlying
> execution resources than they are in native 64-bit mode -- or so I would
> expect, anyway.

Hold on: the 15-20% overhead is for one specific benchmark, the memory
bandwidth one from sysbench (and I still do not know precisely where it
comes from).

I made a couple changes in my patch to isolate specific parts (I took
Jeremy's patch back, another test was with the PG_G flag set for the L3
kernel entry): the differences remained in less than 2% - basically, noise.

In every other tests (the one I trust the most being the build.sh
release one, which takes 3-4h long even in MP context), the overhead
between PAE and !PAE is lower (no more than 3%). It represents a real
world use, not a 500 lines program that just measures context/thread
switch latency.

I ran this tests 3 times for each kernel, and results are tightly packed
in the same interval:

## CURRENT
11398,02 real     18658,69 user      3303,09 sys
11321,97 real     18668,48 user      3329,55 sys
11317,80 real     18667,28 user      3332,21 sys

## CURRENT-PAE
11577,67 real     18977,94 user      3580,73 sys
11572,59 real     18971,16 user      3565,14 sys
11584,48 real     18970,06 user      3589,45 sys

-- 
Jean-Yves Migeon
jeanyves.migeon%free.fr@localhost