http://bugzilla.xensource.com/bugzilla/show_bug.cgi?id=691

Summary: iperf concurrently to dom0 & guest1 yields bad net performance. Product: Xen Version: unstable Platform: x86 OS/Version: Linux-2.6 Status: NEW Severity: major Priority: P2 Component: Unspecified AssignedTo: xen-...@lists.xensource.com ReportedBy: stev...@savvis.net

Test Setup: dom0 and xen1 are the root and guest domain on the xen machine and are running iperf TCP servers. iperf clients are running on various machines. A TCP test from any machine to either domains yields similar results. If I connect two iperfs on the same machine to dom0 and xen1, dom0 gets twice the bandwidth as xen1 but both are steady. If I connect machine A to dom0 and B to xen1 and start the tests at nearly the same time, one domain gets all the bandwidth and the other gets nearly nothing (perhaps some spikes to 10% or a trickle). It seems that the first to start the test gets the bandwidth (doesn't matter if it is dom0 or xen1). It doesn't matter which machine calls which domain. Also, if I have A connected to dom0, wait a bit, and connect B to xen1, B gets nothing until A finishes.

I have tried this enough different ways to think it should be trivial to reproduce for anyone (with three computers). I would think this would be a big enough, and obvious enough, problem to be noticed before - but...

Configuration: Clean FC5 install with xen 3.0.2-2 / vanilla kernel 2.6.16 downloaded from xensource about 6/14/2006. Guest OS is one created from the xenguest-install.py in the FC5 distro (xen). All the setups are (pretty much) default.

Tools: get iperf at http://dast.nlanr.net/Projects/Iperf/ and a gui front end jperf at http://dast.nlanr.net/projects/jperf/