With 1 Gig of real memory and 512 MB per guest, you're probably measuring
the VM paging subsystem or some other overhead phenonmena, which is
probably tunable, not the Linux guest. - with 10x512 MB guest to 1 GB real
memory, you may be overcommited by more than 5 to 1 because VM chews up a
good bit of that memory too! The miracle would probaly be to do something
like add more memory.
I'd tried your perl_bench on an LPAR on a 9672 (G6) and a guest on a 2064
(z116)and the results were very consistent (though I did change the 2nd
loop to 10000).
SLES8 with SP2 applied
IFL
2 GB memory
9672 bogomips 634.06 (G6)
time ./perl_bench
real 11.7-11.8s
user 11.7-11.8s
sys 0.010s
EC guest
non-IFL
SLES8 with SP2 applied
512 GB guest memory/ 24 GB VM memory
2064 bogomips 776.6 (z1 - 116)
time ./perl_bench
real 9.6s
user 9.6s
sys 0.0s
It both cases, the results repeated with a deviation around 1%.
Regards, Jim
Linux S/390-zSeries Support, SEEL, IBM Silicon Valley Labs
t/l 543-4021, 408-463-4021, [EMAIL PROTECTED]
*** Grace Happens ***
Matt Lashley/SCO
<[EMAIL PROTECTED] To: [EMAIL PROTECTED]
te.id.us> cc:
Sent by: Linux on Subject: Re: offloading CPU intensive
loads from zLinux to cheaper pastures
390 Port
<[EMAIL PROTECTED]
IST.EDU>
06/19/2003 08:32
AM
Please respond to
Linux on 390 Port
Just in case anyone is interested --
Though this initial focus of this comparison seemed geared toward S390
Linux running in an LPAR and Linux on an x86, since I no longer have an
S390 Linux LPAR I ran the test under a few different VM guest machines.
The code I used is at the bottome of this note. (Copied from John's
example.)
What I found strange was that the fatsest times (posted under Results1)
came off of the initial run. Subsequent runs were always slower. And,
when I ran the code on all three machines at the same time (posted under
Results2) the times increased fairly dramatically.
Dear IBM - please come up with a way to level the trade-off between
processing power of an IFL and its cache. I fear MTBF, throughput and
white space won't be enough in the near future. The gap in TCO won't be as
wide forever. You guys have worked miracles before...
The system:
9672 G6 machine with a single IFL
1 gig of real mem (for the whole VM system)
VM 4.3
Ten concurrently running Linux guest machines
Guest machine one:
OS: SLES8
Mem: 512M
CPU Share: 2000 (relative)
Heaviest app on the server: Domino 6.5 (not heavily used)
Results1:
tuxd1:~ # time ./perl_bench.perl
real 0m14.014s
user 0m14.010s
sys 0m0.010s
Results 2:
tuxd1:~ # time ./perl_bench.perl
real 0m42.427s
user 0m42.350s
sys 0m0.080s
Guest machine two:
OS: SLES8
Mem: 700M
CPU Share: 3500 (relative)
Heaviest app(s) on the server: DB2, Information Builders WebFocus software,
Apache and Tomcat (not heavily used)
Results1:
tuxib:~ # time ./perl_bench.perl
real 0m14.470s
user 0m14.420s
sys 0m0.010s
Results2:
tuxib:~ # time ./perl_bench.perl
real 0m31.623s
user 0m31.320s
sys 0m0.080s
Guest machine three:
OS: SLES7
Mem: 128M
CPU Share: 1100 (relative)
Heaviest app(s) on the server: Apache, Mysql (moderately used)
Results1:
xapdvlp:~ # time ./perl_bench.perl
real 0m26.164s
user 0m26.160s
sys 0m0.000s
Results2:
xapdvlp:~ # time ./perl_bench.perl
real 0m58.682s
user 0m20.310s
sys 0m21.370s
Code:
#!/usr/bin/perl
#use integer;
$i = 0;
while ($i < 1000)
{
$j = 0;
while ($j < 10000)
{
++$j;
}
++$i;
}
Matt Lashley
Idaho State Controller's Office
