From: hohndel@informatik.uni-wuerzburg.de (Dirk Hohndel) Subject: Re: German Magazine Benchmarks Linux 0.97 Date: Sun, 1 Nov 1992 16:59:29 GMT
: - What about disks for paging/swapping? These are important for
: multiprocessing performance. If the Sun had to go over NFS,
: that would explain a lot of sluggishness.
Both of the SUNs had local SCSI disks in the 1GB range,
a Fujitsu-M2266SA in the ELC, and something similar in the SS2
On otherwise unused SUNs with 32 or 40 MB of RAM, swapping shouldn't
occur :-)
:
: - For the interactive tests, there aren't as many widely used
: multiprocessing test suites around. But I sure hope that
: Linux and the Suns used the same shells -- zsh on the Sun I'm
: sitting at right now (Sparc) is 320K bytes, which is a lot of
: memory to be sloshing back and forth each fork(). (I thought
: ScumOS finally got copy on write in 4.0.0 -- if it still hasn't,
: *shame* *on* *you* *Sun*!)
The ELC and the Linuxbox were using bash, the SS2 had tcsh
:
: - In comparing different operating systems (even after normalizing
: for differences in raw CPU/memory speed), it's important to keep
: in mind that both generality and additional functionality carry
: a cost, *even* *when* *they're* *not* *in* *use*. The fact that
: ScumOS is essentially machine-independent (versions quite close
: to the most recent, eg 4.1.1, run on Sun 3 = 68020 boxes, and
: a version for the [34]86 is just being released) is an example
: of this -- it has various benefits, but it carries a performance
: and size cost, which certainly shows in tests like these.
: Similarly all the native language support, the streams I/O, etc.
: These probably contribute to a good deal of ScumOS's poor
: performance in these tests.
This is one point and another is, that the Byte Benchmarks give big wheight
to taskswitching latency, and this is where Linux wins.
:
: Of course, Sun being in both the software and the hardware business,
: and hardware traditionally being the major $$$ source, tends to
: make their responce to software-bloat-endengered performance
: problems closer to "buy a bigger (Sun) computer" than "form a
: tiger team to tune or even slim down the software". It's notable
: that /vmunix on the Sun (Sparc) I'm sitting at right now is 1.4
: megabytes in size!
:
The one on the SS2 is 1698406 Bytes (this has LOTS of external drives and
devices), the one on the ELC is 1162555 Bytes, so both are _alot_ bigger than
the 0.97pl? kernel used on the 486 (around 250k)
: It will be interesting (in a pornographic sort of way :-) ) to see
: just how big and slow Solaris 2.0 (= ScumOS 5.0) on 386s is. The
: side-by-side comparison with Linux, Interactive 386, BSDI 386,
: and soon Gnu Hurd on the same hardware will be interesting.
:
:-)
Dirk