If you have any variable latency issues in your network (more than 20ms RTT) you will see a large difference in your transmittion troughput.
Despite the latest TCP congestion control features (see http://dsd.lbl.gov/TCP-tuning/background.html), we are seeing about 300% on -p 16 for large files (more than 2G) The biggest improvements are seen on large latency links with no TCP tuning. Terry Rankine -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of John Bresnahan Sent: Tuesday, 19 June 2007 9:41 PM To: Seshachalapathi Cc: [email protected] Subject: Re: [gt-user] globus-url-copy , -p option (parallelism) > I have been testing globus-url-copy between two systems in a node. I > tested with no parallel (i.e with out -p option), and, tried with -p 2 > and -p 4. I dont see any difference in transmission times? parallel tcp streams is a network optimization when transmitting data over networks where packets are lost. it doesn't necessarily translate into higher throughput over all networks. > When I use -p option with globus-url-copy, > 1) will there be multiple ftp processes running on receiving or > sending system (checking with ps -ef)? No, i think you are thinking of server striped transfer. no matter what the parallel streams count, they are formed between 2 end points. > 2) Will there be any noticable change in transmission time before and > after using -p option? often times yes, but it depends on your network. > 3) when I ftp a directory with 4 files in it, and use -p 4 option , > then will these 4 files ftp'ed at same time with 4 threads? no. That is not how parallel streams work. Parallel streams just break 1 logical pathway into serveral logical pathways to lessen the penalty of a dropped TCP packet. What you are talking about here is known as 'client side striping'. > > Please help with my doubts? > > Seshachalapathi > > > > ______________________________________________________________________ > _____________ You snooze, you lose. Get messages ASAP with AutoCheck > in the all-new Yahoo! Mail Beta. > http://advision.webevents.yahoo.com/mailbeta/newmail_html.html >
