NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/comp/unix/wizards/3628

Text File | 1992-08-15 | 1.9 KB | 40 lines

Newsgroups: comp.unix.wizards
Path: sparky!uunet!haven.umd.edu!socrates!berry
From: [e-mail] (Joseph Berry)
Subject: Re: MP CPU-time phenomena
Message-ID: <[e-mail]>
Organization: University of Maryland University College
References: <[e-mail]> <[e-mail]> <[e-mail]> <[e-mail]>
Date: Sun, 16 Aug 1992 04:23:52 GMT
Lines: 29

>>>There's no reason that a processor switch could not occur anywhere in
>>>the code... In a 10 second window, all processors could have a chance to
>>>execute the code stream.
>>

Let me try and clarify the original question better (I work with the
person who posted the question).  We have a monitoring tool (that's
what our company writes -- performance tools) that works on a couple of
different MP UNIX machines. We observed the following 'weird' phenomona:
while monitoring the CPU stats of a given process (the very simple CPU
loop example where there is nothing really to parallelize assuming the
compiler (standard UNIX SVR3 cc) can parallelize anything anyway, we
noted that during a 10 second window, this process frequently consumed
far more than 5 seconds worth of CPU time (sometimes as high as 8 sec).
Most of the time, however, it was using just about 5 secs which is what
one would expect it to use (1/2 of 2 cpus).  Is there any way to explain
this phenomona?  This was observed on both SCO MPX and SUN's 600MP
systems.

joe

+------------------------------------------------------------------------+
| Joe Berry                                                              |
| Strategic Software Group, Ltd., 3402 Shelburne Rd, Baltimore, MD 21208 |
| E-mail:  [e-mail]  or  uunet!ssgltd!joe                          |
| Phone:  [phone];  Fax: [phone];  Pager: 410-796-6861         |
+------------------------------------------------------------------------+
t