slow on HPUX, SunOS, fast on MS Win2K and Linux?

[Ian Parker]

In article <xqxlm45jvpd.fsf at cola2.ca.boeing.com>, Harry George 
<hgg9140 at cola2.ca.boeing.com> writes
>Is there some reason python code which builds objects should be
>notably slow on HPUX or SunOS? Some known code optimization problem?
>
>We are seeing:
>
>> -----------
>> import cPickle
>> f =
>> open("/acct/mad5486/SpecEngineV0/bac5117-compiled.xml-pickled",'r')
>> xml_dom_object = cPickle.load(f)
>> ----------------------------END-------------------------------
>> ----------
>>
>> Times:
>>
>>  5 seconds on Intel / Linux
>>  5 seconds on Intel / Win2k
>> 33 seconds on HP-UX 11.0  (uname -a: HP-UX dblv018 B.11.00 E 9000/800)
>> 22 seconds on Sun OS (uname -a: SunOS wizard 5.8
>> Generic_112953-02 sun4u sparc SUNW,Ultra-Enterpris)
>> 17 seconds on HP-UX 10.2  (uname -a: HP-UX drtn029 B.10.20 E 9000/819)
>>
>
>The Win2K and Linux boxes are (nearly) identical hardware.  The others
>are serious machines with substantial RAM.  The guy running the tests
>tried to isolate for these issues:
>
>1. No one else on the machine; light load of background tasks.
>
>2. Code and datafile cached in RAM (no delays due to LAN-based file
>   servers).
>
>3. Results similar for Py 1.5.2 and 2.2.1 on each machine.
>
>4. According to "top" process was getting 99% of CPU, and it was CPU
>   bound.
>
>5. While this example is from the pickle, the same effects were seen
>   for the original PyXML parse of the DOM.
>
As a first step, it might be worth checking the relative power of the 
systems, for example have a look at http://www.spec.org.  At least that 
will give you some idea of what differences to expect.

In a previous incarnation as a datacentre manager, I was increasingly 
embarrassed that of all the "serious" machines in our datacentre (Sun, 
IBM and HP, Compaq Alpha)  only the some of Compaq Alphaservers were 
faster, CPU for CPU, than the recent Intel boxes.  We bought the 
"serious" machines for more for their multiprocessing, large memory, i/o 
bandwidth or even OS (e.g. for VMS or Tru64 UNIX clustering).

So if you've got relatively new Intel boxes, you may well find that they 
are seriously faster.

But I wouldn't expect that to explain the whole of the difference.

Regards
-- 
Ian Parker