<br><font size=2><tt>Roland,</tt></font>
<br>
<br><font size=2><tt>Roland Dreier <rdreier@cisco.com> wrote on 05/24/2006
02:03:34 PM:<br>
<br>
> Shirley> My observation is the atomic operation is
not that<br>
> Shirley> expensive.<br>
> <br>
> It's just about the worst thing to do. For example, on x86/x86-64
an<br>
> instruction with the lock; prefix is quite slow. If you look
at an<br>
> instruction level profile you can see that quite clearly.<br>
</tt></font>
<br><font size=2><tt>Compared to have a single thread handling AH, I don't
think this atomic operation is expensive.</tt></font>
<br><font size=2><tt><br>
> Shirley> I thought the path holding another AH reference
to<br>
> Shirley> prevent it to be freed?<br>
> <br>
> If that were true then why would we want to reference count sends
at<br>
> all? The whole point is that a path might be destroyed before
the<br>
> send is executed.<br>
> <br>
> - R.</tt></font>
<br><font size=2 face="sans-serif"><br>
It is true for unicast, it has a reference count before ipoib_send(). I
need to look at multicast.</font>
<br>
<br><font size=2 face="sans-serif">Thanks<br>
Shirley Ma<br>
IBM Linux Technology Center<br>
15300 SW Koll Parkway<br>
Beaverton, OR 97006-6063<br>
Phone(Fax): (503) 578-7638</font><font size=2><tt><br>
</tt></font>