From: Lars Rzymianowicz Newsgroups: comp.parallel.pvm Subject: Re: linux parallel-PVM benchmark program Date: Mon, 16 Nov 1998 16:35:11 +0100 Organization: Dept. of Computer Engineering, University of Mannheim, Germany Message-Id: <3650462F.96BAB718@mufasa.informatik.uni-mannheim.de> References: <72mb8h$29t$1@nnrp1.dejanews.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Clark Dorman wrote: > Well, there isn't a Linux specific one. However, I would recommend > using NETPIPE, which uses either PVM, MPI, or TCP, if you want to see > the bandwidth capability of PVM or want to know information about > optimal packet sizes on your system. > Nice tool. But results are strange. E.g., if you look at the results for ASUS boards (200MHz PPro) with TCP/IP over Gigabit Ethernet or Myrinet, both networks reach about 150 Mbit/s bandwidth. But hardware offers 1Gbit/s! Where's all the bandwidth gone? Are the protocol implementations so bad that you can only get 15% of peak throughput in real apps? A short test of netpipe with TCP/IP here with our local 10Mbit/s Ethernet resulted in 7.5Mbit/s for large blocks. 75% in real apps, that's a number i can accept. But 15%??? Any comments? Lars -- Homepage: http://mufasa.informatik.uni-mannheim.de/lsra/persons/lars/