Change Details

The threshold was introduced in 2004 with the limit of 500. It was bumped to 1000 in 2007, but has not been bumped since. The ixl.4 manual page has documented that the threshold falsely detects interrupt storms on 40Gbit NICs as long ago as 2015, and we have seen similar false positives with the ioat(4) DMA device (which can push GB/s). Arbitrarily bump the threshold to 8000 to avoid false positives on high performance server devices. (Loosely following the log-trend of doubling the limit every 3 years.)

The ixl.4 manual page has documented that the threshold falsely detects interrupt storms on 40Gbit NICs as long ago as 2015, and we have seen similar false positives with the ioat(4) DMA device (which can push GB/s). For example, synthetic load can be generated with tools/tools/ioat 'ioatcontrol 0 200 8192 1 1000' (allocate 200x8kB buffers, generate an interrupt for each one, and do this for 1000 milliseconds). With storm-detection disabled, the Broadwell-EP version of this device is capable of generating ~350k real interrupts per second.

The threshold was introduced in 2004 with the limit of 500. It was bumped to 1000 in 2007, but has not been bumped since. The ixl.4 manual page has documented that the threshold falsely detects interrupt storms on 40Gbit NICs as long ago as 2015, and we have seen similar false positives with the ioat(4) DMA device (which can push GB/s). Arbitrarily bump the threshold to 8000 to avoid false positives onFor example, synthetic load can be generated with tools/tools/ioat high performance server devices. (Loosely following the log-trend of'ioatcontrol 0 200 8192 1 1000' (allocate 200x8kB buffers, generate an doubling the limit every 3 years.)interrupt for each one, and do this for 1000 milliseconds). With storm-detection disabled, the Broadwell-EP version of this device is capable of generating ~350k real interrupts per second.