Our default strategy for migrating NDT from web100 kernels to tcpinfo kernels was to carry forward network settings established on the web100 platform, including ethtool -K eth0 tso off. Historically, TSO caused excessive jitter for TCP based e2e network measurements. Newer kernels (including the CoreOS 4.19 kernel) seem to have a bug in the TSO path. TSO has been on by default for some time, and improved TSO auto sizing has greatly reduced the jitter it causes. We have anecdotal evidence that the tso off configuration may have impacted flows with speeds approaching 2Gbps.
Our default strategy for migrating NDT from web100 kernels to tcpinfo kernels was to carry forward network settings established on the web100 platform, including
ethtool -K eth0 tso off
. Historically, TSO caused excessive jitter for TCP based e2e network measurements. Newer kernels (including the CoreOS 4.19 kernel) seem to have a bug in the TSO path. TSO has been on by default for some time, and improved TSO auto sizing has greatly reduced the jitter it causes. We have anecdotal evidence that thetso off
configuration may have impacted flows with speeds approaching 2Gbps.