On 22.4.2021. 13:08, Hrvoje Popovski wrote: > On 22.4.2021. 12:38, Alexander Bluhm wrote: >> It is not clear why the lock helps. Is it a bug in routing or ARP? >> Or is it just different timing introduced by the additional kernel >> lock? The parallel network task stress the subsystems of the kernel >> more than before with MP load. Having more results from machines >> with different timing would help. >> >> bluhm >> > > > I'm running this diff on Dell r620 with 6 x Intel(R) Xeon(R) CPU E5-2643 > v2 @ 3.50GHz, 3600.52 MHz, 06-3e-04 and ix and forwarding went up from > 1Mpps to 1.8Mpps
fun fact, with parallel diff on top of fresh source i'm getting 2.2Mpps of plain forwarding and 3.5Mpps over veb(4) with NET_TASKQ=4 on 6 x Intel(R) Xeon(R) CPU E5-2643 v2 @ 3.50GHz, 3600.03 MHz