[linux-mm-cc] interbench result

[Nitin Gupta]

On Tue, May 13, 2008 at 9:51 AM, Nai Xia <nai.xia at gmail.com> wrote:
> ---------- Forwarded message ----------
>  From: Nai Xia <nai.xia at gmail.com>
>  Date: Mon, May 12, 2008 at 11:58 PM
>  Subject: Re: [linux-mm-cc] interbench result
>  To: John McCabe-Dansted <gmatht at gmail.com>
>
>
>
>  On Mon, May 12, 2008 at 10:44 PM, John McCabe-Dansted <gmatht at gmail.com> wrote:
>   > I find the following format easier to understand, and have attached Nai's
>
>   Yes, you are right, :)
>
>
>
>   > result in this format. Additionally I have generated results for my P3-500
>   > with 256MB of ram and 512MB of compcache, and included the scripts I used.
>   >


>   > Anyway, Nai do you have an interpretation  of this data? Also, is the memory
>

Readme file included with interbench contains some information on this.


>   Well, actually I am thinking although interbench was a good
>   Interactivity benchmark tool,
>   it may not distinguish compcache much. For most of the numbers are
>   related to CPU timing,
>   but all the simulated programs do not change RAM usage when they are
>   waked up or sleep.
>   I am thinking that the result is of little relevance to the amount of
>   PHY RAM it has, it is designed
>   for scheduler.(I will do additional test to prove this).
>

Yes, now I also think the same. After interbench tests are completed,
very little swap
space is used. So, it is not much afftected by compcache.

>   By now, from my point of view, what we really need is a
>   famous/standard benchmark that makes anonymous pages compete
>   drastically with file cached pages (like kpdf & firefox )---- that is
>   indeed where compcache stands like a real hero.
>

I think ideal scenario for compcache is when all of these conditions
are satisfied:
 - Working set of apps (anon + page cache) > Physical RAM
 - Phy RAM left over from compcache is enough to hold page-cache part
of working set
 - Compressibility is good

These conditions seem to hold true for "general desktop use" (web
browsers, editors etc.)

>
>
>
>   > used in any sense of "normal" compressibility?
>
>   I did not record the compressibility, since I thought intercepting
>   interbench might change the result .....
>
>

You can note this by checking /proc/compcache after interbench tests
are complete.
In my tests, I see that compressibility is great (NumGoodCompress > 95%) since
maybe its filling memory with 0's. One strange thing to notice is that number
of ramzswap reads/writes is huge (in order of 10^6).


Cheers,
Nitin