From: Julian Tarkhanov <listbox@...>
Date: 2007-09-29T12:40:01+09:00
Subject: Strange difference in performance (30x) between PPC and Intel

Hello Rubytalkers

I have a very odd problem here and thought that I should ask your  
opinion. I got a test suite here

http://julik.textdriven.com/svn/tools/micro-apps/pasaporte/trunk/test/

it's for a little Camping app for OpenID. This is not PR, I mind you.  
However, I have some issues running it on two of our machines. When I  
run it
on my staging server:

processor       : 0
vendor_id       : GenuineIntel
cpu family      : 15
model           : 2
model name      : Intel(R) Xeon(TM) CPU 2.40GHz

The test suite takes 3 seconds to run. Fine is that.

Then I run it on my home G4 (about 700 Mhz).
It takes 19-26 seconds (20ish). Ok.

Then I make a checkout on my dev machine at work (a quad G5, 5G RAM )  
and run the suite here.

It takes a friggin' minute.

The test suite does nothing fancy except some crypto. When I do ruby- 
prof both, I see the following at the top of the graph on the G5:

Thread ID: 3348220
Total: 12.55

  %self     total     self     wait    child    calls  name
  22.07      2.77     2.77     0.00     0.00      155  IO#read  <<<<  
this is suspicious!
   4.30      0.54     0.54     0.00     0.00       45   
OpenSSL::BN#mod_exp
   3.51      0.44     0.44     0.00     0.00     1495  <Class::Dir>#[]

On the Ubuntu box the same shows:

Thread ID: 3084113660
Total: 4.92

  %self     total     self     wait    child    calls  name
   8.54      0.42     0.42     0.00     0.00       45   
OpenSSL::BN#mod_exp
   3.86      0.19     0.19     0.00     0.00      155  IO#read

I immediately suspect that something is blocking somewhere. Almost  
anytime I try to see what's going on by hard-stopping a test with ctrl 
+c I see a reference to
the ruby-yadis library trying to do an HTTP call.

Now comes the question. How should I proceed to eliminate this  
problem? Mock the hell out of the HTTP calling part of ruby-yadis?

I mean, a quad 64bit machine like that cannot be so slow at running a  
test suite like that and the only processor-dependent
thing I am using here is OpenSSL (the G5 and the dual xeon come  
pretty close with the openssl-speed test). Did anybody have a problem  
with Net::HTTP waiting too long
and/or blocking before it fails to make an HTTP connection on OS X?

-- 
Julian 'Julik' Tarkhanov
please send all personal mail to
me at julik.nl