The network's nominal bandwith is 1GB (gigabit with a "G"). Even @ 30% util, that's 300mbit/sec. In your calc bandwith required is ~98mbit -or 1/3 of practical available. Not sure why you asserted that the bandwith is 1mbps. btw, windows says "network utilization" is < 2% during said test.
FWIW, we've discovered that replacing NIO with the old blocking connector improved latency by a factor of 4-5 times. Digging deeper... thanks -nikita Christopher Schultz-2 wrote: > > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA1 > > Nikita, > > On 4/16/2009 3:09 PM, dukehoops wrote: >> In brief: >> I am load-testing an app running on Tomcat 6 NIO and am observing a big >> discrepancy between avg request latency as reported by JMeter (660ms) and >> the profiler (19ms). I would like to understand where is the delta being >> spent. > >> JMeter is simulating 40 users, each repeatedly hitting the server with >> the >> same request (against port 8080), with 50ms think time. The response for >> each request is in plain text, about 10-20k worth. The request hits >> app-level cache 100% of the time, so JDBC connection aren't even being >> checked out of the cxn pool and no DB work is being done. > > Assuming this is true, here is your raw data: > > Simultaneous “users” 40 > Request Think Time (ms) 50 > Response size (kB) 15 > Network speed (ideal kbps) 1024 > >> Given the huge discrepancy, I would like to determine where is 642ms >> (=661-19) being spent. Network latency should not be an issue. > > Not latency, but perhaps bandwidth: > > Request Freq (1/sec) 800 > Response size (bits) 122880 > Response bits per second 98304000 > Network speed (bps) 1048576 > > You are trying to shove ~100 times more data through your network per > second than it can handle. Also don't forget that actual ethernet > bandwidth is more like 30% of the the "ideal", so you're talking about a > network that can actually run at something around 315,000bps. > > Did you think that you wouldn't clog Teh Tubes with all that data? The > Internet isn't like a big truck. > >> Is it >> possible that Tomcat is spending time doing something in NIO poller - >> before >> Tomcat's exec threads are engaged? > > Try backing-off your tests to the point where your network can keep up, > and then see what your test report. > > - -chris > -----BEGIN PGP SIGNATURE----- > Version: GnuPG v1.4.9 (MingW32) > Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/ > > iEYEARECAAYFAkns7BQACgkQ9CaO5/Lv0PCVGACgqA/umJr4kgXBSmPlCHl6LChe > 3ksAn004amc0l6oCUi+u6y9dvhgtxhMM > =mTzH > -----END PGP SIGNATURE----- > > --------------------------------------------------------------------- > To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org > For additional commands, e-mail: users-h...@tomcat.apache.org > > > ----- ---------------- Nikita Tovstoles vside.com ---------------- -- View this message in context: http://www.nabble.com/what-can-Tomcat-be-doing-that-a-profiler-can%27t-see--tp23084469p23145723.html Sent from the Tomcat - User mailing list archive at Nabble.com. --------------------------------------------------------------------- To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org For additional commands, e-mail: users-h...@tomcat.apache.org
