Re: [gentoo-user] Intel Atom: architecture, distcc, crossdev and compile flags

[Florian Philipp]

Am 11.12.2012 18:36, schrieb Frank Steinmetzger:
> Hello list
>
> Long time no read... :)
>
> It follows a verbose preamble.  For the actual questions see dashed line below.
> TL;DR summary: it’s all about ricer-performance questions on a netbook.
>
>
> I have the luck of having obtained a used netbook for free (Atom N450, single-
> core with HT, 1 GB memory, 5400 rev HDD).  During the last week I’ve been
> experimenting with 32 and 64 bits on it and am still quite undecided which to
> keep. My reasons:
>
> - They are not as far apart in CPU performance as is the Core2.
>   I posted a 32/64 comparison for Core2 a few months ago, which showed that
>   Lilypond speedup on 64 bit was 50%. On the Atom, it actually took 5% longer.
>   (Sadly, Blender doesn’t build on 32 bit right now).
> - Startup times for hogs like Firefox and KDE are quite equal between the two
>   (that could be attributed in parts to the fact that the 64 bit partition
>   sits on the disk’s first sectors, while 32 bit sits at the other end, I
>   don’t know which end is faster).

The first part (which actually is the outer edge of the physical disk).

> - pro 64: it is very easy to use distcc, as opposed to 32 bits (see below).
> - con 64: it uses about 50% more memory, 32 bit builds are a little faster.
>
> The RAM argument is the most convincing one right now, since more free RAM
> means more cache, which means a faster system in the long run. Currently, KDE
> after logon needs 150 MB on 32 bit, and 250 MB on 64 bit (without akonadi for
> now). But awesome WM rocks on a netbook anyway.
>
>
> ----------[ Questions begin ]------------------------------------------------
>
> So I’m interested in you opinion and own experience about the following
> arising questions:
>
> * From my observations, the benefit of 64 bit over 32 is much smaller for an
>   Atom than it is for my Core2.  Am I right to assume thus that the Atom
>   architecture doesn’t have much to offer to 64 bit (such as extra registers)?
>   I’m not talking about memory here, since it’s limited to 2 GB in any case.
>

It has the same set of registers as your Core2. It's just that the Atom
micro-architecture is terrible with regard to 64bit. That's just about
the only reason that x32 was invented (and now that I've said it, I'm
just waiting for the flamewar about it).

> * The problem of distcc between different architectures:
>   The netbook already had an older 32 bit Gentoo installed.  And since I have
>   a multilib host (march=core2), I though I could upgrade with distcc (using
>   march=atom on the netbook).  But at some point more and more stuff stopped
>   working, eventually I got “invalid instruction” errors during emerge, hence
>   I figured that was a dead end.
>
>   So is it possible to mix architectures in this way at all with distcc?
>   I also have crossdev for i686 installed, which even shares files with the
>   system’s normal multilib gcc. I find that odd.

I don't think you can mix x86_32 and x64 easily but I've never tried.
Did you try adding a CFLAGS="-m32"? From comparing `gcc -Q --help=target
-march=xxx` they should be compatible.

>   I sped up the installation process for 32 bit by using a chroot on the big
>   machine, which worked nicely.  But it’s not a long-term solution, b/c it
>   uses up too much disk space on the host.
>

I do the same using NFS, bind mounts and tmpfs. What do you mean by disk
space? If you can use common CFLAGS, you could try installing binary
packages from your build host on your netbook (use quickpkg and friends).

> * I’m interested in the question of -O2 vs. -Os.
>   Some sources say -Os is bad, b/c it breaks debugging and is mainly untested.
>   I won’t do heavy developing on it anyway, and Atoms do have a puny cache.
>   So I wonder whether -Os would improve execution time and RAM usage
>   noticably.  Diskspace itself is not an issue.
>

I use -Os but have no data to back it up (I'm also using an older Atom).
Gentoo wiki suggests using "-O2 -fno-reorder-blocks
-fno-reorder-functions" as a compromise.

> * I’m also interested in comparing bin packages over self-compiled ones.
>   E.g. I did compile icedtea, even if it’s just for TV browser. :)
>   Can you name a Java benchmark to measure CPU performance?
>

How about something from this site:
http://shootout.alioth.debian.org/

> * The last thing I’m going to set up is filesystem encryption, at least for ~.
>   I already know/think that AES would be the best choice due to limited CPU
>   power, but what else is there to heed besides key size?
>

Nothing, you're good. Hash and key chaining method have negligible
impact. If you stick with an x86_32 userspace I suggest at least using
an x64 kernel so you can use of CRYPTO_AES_X86_64.

> * What other small benchmarks for CPU and memory can you recommend?  So far I
>   tested with nbench and sysbench. The results are so-and-so. Some computation
>   stuff is much slower on 64 bit, some a bit faster.  The applicability to
>   every-day use is of course a wibbly-wobbly argument.
>   I also tested the runtime of some application (packing and unpacking of
>   archives, throughput with dd, mencoder).  If there is interest, I can post
>   the result of 21 runs on each platform, measured with GNU time.
>

How about trying some browser benchmarks. Check out the following for
different aspects:
http://ie.microsoft.com/testdrive/performance/mazesolver/default.html
http://www.webkit.org/perf/sunspider/sunspider.html
http://peacekeeper.futuremark.com/run.action

There is also a Qt render benchmark
http://code.google.com/p/qtperf/
Check out app-admin/eselect-qtgraphicssystem and see how they compare in
appearance and numbers.

Regards,
Florian Philipp