[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference decwet::ntaxp

Title:Windows NT For Alpha AXP
Moderator:TARKIN::LINEIBER
Created:Mon Sep 27 1993
Last Modified:Thu Jun 05 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:1058
Total number of notes:4567

1043.0. "Named Pipes and SMP" by NNTPD::"[email protected]" (Genne Roger) Tue May 06 1997 11:37

Have a customer with dual processor 4100 with 4GB RAM who says his 
application written to use named pipes is slower than when he ran it on only
one processor and 2GB RAM.  He knows this by measuring what he calls "atomic
units of work", which is the client side sending 8 bytes to the server side
and then the server side sending 8 bytes back.  There is no disk I/O or
network traffic involved.  He gets 2300/sec in the one CPU/2GB systems and
only 1900/sec in the two CPU/4GB system.  PM shows CPUs as evenly utilized at
low %.  
Have checked out all firmware levels, and have applied hotfix to NT 4.0 SP2. 
Same result.  Now we are going to experiment with taking out the memory and
swapping CPUs to see what happens.  
Anyone have any ideas on what could be happening?  Customer is willing to
rewrite application if MS named pipes is the problem.
Thanks for any hints or tips.

[Posted by WWW Notes gateway]
T.RTitleUserPersonal
Name		DateLines
1043.1See also MVBLAB::ALPHASERVER_4100 note 583NPSS::GLASERSteve Glaser DTN 226-7212 LKG1-2/W6 (G17)Tue May 06 1997 17:196
    See also MVBLAB::ALPHASERVER_4100 note 583.
    
    No solution there either, but it's a similar situation (2 processors
    running slower than 1 when using named pipes under NT).
    
    It might even be the same customer.
1043.2Question about appwagtake.zso.dec.com::wagnerWed May 07 1997 14:165
One question: Is there an attempt to pipeline the send/receive?
In other words, does the client attempt to send another 8 bytes
before getting the reply from the server for the previous 8
bytes?
1043.3Named Pipes/SMPNNTPD::"[email protected]"Genne RogerThu May 15 1997 16:2010
No, according to the writter of this Visual C++ app.  We wait for the reply
back from the server.  It is all a moot point now, because the customer has
decided not to try and fix his code, but to run the app on Sun.  The consensus
from a number of folks around Digital was that this was an app problem and
that the customer needed to expend more resource to make it work on NT/SMP. 
Much of the basis for this statement was that the app behaved the same way on
Intel, and that programs like SQL do scale reasonably on Alpha as well as
Intel.

[Posted by WWW Notes gateway]
1043.4Huh?wagtake.zso.dec.com::wagnerMon May 19 1997 20:376
I might be missing something, but how are they expecting 
multiple CPUs to help in what is, essentially, a single
threaded application.

Dave Wagner
1043.5KEIKI::WHITEMIN(2�,FWIW)Mon May 19 1997 22:584
    
    	One wonders if they even tried multiple clients?
    
    						Bill