Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!wuarchive!usc!apple!agate!darkstar!rex.cs.tulane.edu
From: bdf@rex.cs.tulane.edu (Brett Fleisch)
Newsgroups: comp.os.research
Subject: Re: Toy Experiment with Shared Memory Programming
Message-ID: <8916@darkstar.ucsc.edu>
Date: 13 Nov 90 16:17:51 GMT
Sender: usenet@darkstar.ucsc.edu
Organization: Computer Science Dept., Tulane Univ., New Orleans, LA
Lines: 41
Approved: comp-os-research@jupiter.ucsc.edu


In article <8754@darkstar.ucsc.edu> jms@central.cis.upenn.edu (Jonathan M. Smith) writes:
>Some Notes on Experiences programming shared memory
>

>In some initial trials, I found that the shared memory program ("mem")
>performed far worse than "piper". This surprised me, so I profiled the code
>to find out where the time was going, and it was being eaten up by the
>calls to "semop()", which are used to implement P() and V(). But the cost
>of the calls was not the problem; it was the *number* of calls! For a file
>(my "mbox") with 184 8192-byte blocks, about 30-40 K calls to semop were
>made! This didn't sem possible from examining the code; and I must admit to
>a lot of hacking trying to "fix" the problem, which I was sure was a BUG.
>
>The problem was that pipes are well-integrated with process synchronization
>and scheduling. While executing a "V()" operation freed the semaphore, it
>did not give up the processor, so that the process which had just executed
>the "V" could now execute a "P" again immediately. After I saw this, I quickly
>rewrote the program with two semaphores so that better control over process
>control (such as in "sleep" and "wakeup") and response to events could be
>achieved.
>

I observed a similar problem in Mirage.  Why not read over the article
"Mirage: A Coherent Distributed Shared Memory Design" in SOSP12 noting the 
yield() instruction which was added to better control UNIX synchronization
and scheduling interactions.

If you have questions contact me:

Brett D. Fleisch
Assistant Professor
Computer Science Department
Tulane University
301 Stanley Thomas Hall
New Orleans, LA 70118-5698

(504)865-5840
bdf@cs.tulane.edu
-- 
brett@cs.tulane.edu