Hi Peter, Thanks for your comments. I am using 1Gb ethernet.
My loopback testing included the motherboard network circuitry, TCP/IP stack, OS, CPU and client and Tomcat App code. The best I got for a single request round trip that reused the http connection was 600 microseconds. As I added more threads running on the client sending requests it got slower. When not using loopback I got around 1.2msec so the extra components included my 1Gb switch and 1Gb router. When I ran the client from my laptop on the network that had 1Gb ethernet support I got over 2msec. I googled on the internet and I was seeing 2msec as the fastest people were saying for a network card even one with a built-in CPU. I would love to eliminate the network portion from the path and go directly from client to Tomcat to see what Tomcat truely can do. Any recommendations are appreciated. I did move on to other projects since I could not think how more timing info. Regards, -Tony ----- Original Message ---- From: Peter Crowther <peter.crowt...@melandra.com> To: Tomcat Users List <users@tomcat.apache.org> Sent: Mon, March 14, 2011 1:12:12 AM Subject: Re: Performance 5.5 vs 6 vs 7 On 13 March 2011 21:01, Tony Anecito <adanec...@yahoo.com> wrote: > As someone mentioned the network can imit you. If your bandwidth > utilization is > at 60% or over you are in trouble since collisions start to become a > serious > issue. > > Collisions may or may not be an issue, depending on the exact mode of operation of the congested network link. Collisions are a feature of the original Ethernet. They're still an issue on many Ethernets, but there are provisos even here. If your switch does store-and-forward, or you simply have a switched network and all the traffic is between two stations, then collisions are relatively unlikely and the Ethernet usage can go considerably higher than 60% before collisions and retries start to become significant. Sure, if you're on wireless or using an old 10baseT or Cheapernet hub, 60% is about the knee in the curve. But I'd be a little surprised if Dave were using such ancient technology for a critical piece of infrastructure such as this server :-). Point-to-point links, such as most broadband connections, do not suffer from collisions as the routers at each end store and forward traffic across the link. That said, by the time you get to 80% usage then each packet is typically going to have to queue behind several others. Even if this only adds a few milliseconds of latency each time, it builds up when you multiply it by the TCP 3-way handshake, 4-way FIN/ACK sequence, plus any data you're transferring. Dave, could you give us any more information about your network? What is the piece that's at 80% utilisation when you see the trouble? Is it a point-to-point connection, or an Ethernet LAN, or what? If it's Ethernet, what hardware are you using for connection? Tony - I also don't believe 2ms is the best you can get on Ethernet. If it was, most big machine rooms would have silted up with network traffic long ago ;-). It depends on what network technology you're using. For example, gigabit Ethernet latencies are considerably lower than good ol' 10 meg. What link speeds do you have to your router? 100 Mbit/s? - Peter --------------------------------------------------------------------- To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org For additional commands, e-mail: users-h...@tomcat.apache.org
