On 20/02/2024 13.08, Sebastian Andrzej Siewior wrote:
On 2024-02-20 11:42:57 [+0100], Jesper Dangaard Brouer wrote:
This seems low...
Have you remembered to disable Ethernet flow-control?
No but one side says:
| i40e 0000:3d:00.1 eno2np1: NIC Link is Up, 10 Gbps Full Duplex, Flow Control: None
but I did this
# ethtool -A ixgbe1 rx off tx off
# ethtool -A i40e2 rx off tx off
and it didn't change much.
| Summary 8,436,294 rx/s 0 err/s
You want to see the "extended" info via cmdline (or Ctrl+\)
# xdp-bench drop -e eth1
with "-t 8 -b 64". I started with 2 and then increased until rx/s was
falling again. I have ixgbe on the sending side and i40e on the
With ixgbe on the sending side, my testlab shows I need -t 2.
With -t 2 :
Summary 14,678,170 rx/s 0 err/s
receive total 14,678,170 pkt/s 14,678,170 drop/s 0
error/s
cpu:1 14,678,170 pkt/s 14,678,170 drop/s 0
error/s
xdp_exception 0 hit/s
with -t 4:
Summary 10,255,385 rx/s 0 err/s
receive total 10,255,385 pkt/s 10,255,385 drop/s 0
error/s
cpu:1 10,255,385 pkt/s 10,255,385 drop/s 0
error/s
xdp_exception 0 hit/s
receiving side. I tried to receive on ixgbe but this ended with -ENOMEM
| # xdp-bench drop eth1
| Failed to attach XDP program: Cannot allocate memory
This is v6.8-rc5 on both sides. Let me see where this is coming from…
Another pitfall with ixgbe is that it does a full link reset when
adding/removing XDP prog on device. This can be annoying if connected
back-to-back, because "remote" pktgen will stop on link reset.
so I replaced nr_cpu_ids with 64 and bootet maxcpus=64 so that I can run
xdp-bench on the ixgbe.
Yes, ixgbe HW have limited TX queues, and XDP tries to allocate a
hardware TX queue for every CPU in the system. So, I guess you have too
many CPUs in your system - lol.
Other drivers have a fallback to a locked XDP TX path, so this is also
something to lookout for in the machine with i40e.
so. i40 send, ixgbe receive.
-t 2
| Summary 2,348,800 rx/s 0 err/s
| receive total 2,348,800 pkt/s 2,348,800 drop/s 0 error/s
| cpu:0 2,348,800 pkt/s 2,348,800 drop/s 0 error/s
| xdp_exception 0 hit/s
This is way too low, with i40e sending.
On my system with only -t 1 my i40e driver can send with approx 15Mpps:
Ethtool(i40e2) stat: 15028585 ( 15,028,585) <= tx-0.packets /sec
Ethtool(i40e2) stat: 15028589 ( 15,028,589) <= tx_packets /sec
-t 4
| Summary 4,158,199 rx/s 0 err/s
| receive total 4,158,199 pkt/s 4,158,199 drop/s 0 error/s
| cpu:0 4,158,199 pkt/s 4,158,199 drop/s 0 error/s
| xdp_exception 0 hit/s
Do notice that this is all hitting CPU:0
(this is by design from pktgen_sample03_burst_single_flow.sh)
-t 8
| Summary 5,612,861 rx/s 0 err/s
| receive total 5,612,861 pkt/s 5,612,861 drop/s 0 error/s
| cpu:0 5,612,861 pkt/s 5,612,861 drop/s 0 error/s
| xdp_exception 0 hit/s
going higher makes the rate drop. With 8 it floats between 5,5… 5,7…
At -t 8 we seem to have hit limit on RX side, which also seems too low.
I recommend checking what speeds packet generator is sending with.
I use this tool: ethtool_stats.pl
https://github.com/netoptimizer/network-testing/blob/master/bin/ethtool_stats.pl
What we are basically asking: Make sure packet generator is not the
limiting factor in your tests.
As we need to make sure DUT (Device Under Test) is being overloaded
100%. I often also check (via per record) that DUT don't have idle CPU
cycles (yes, this can easily happen... and happens when we hit a limit
in hardware either NIC or PCIe slot)
Doing "ethtool -G eno2np1 tx 4096 rx 4096" on the i40 makes it worse,
using the default 512/512 gets the numbers from above, going below 256
makes it worse.
receiving on i40, sending on ixgbe:
-t 2
|Summary 3,042,957 rx/s 0 err/s
| receive total 3,042,957 pkt/s 3,042,957 drop/s 0 error/s
| cpu:60 3,042,957 pkt/s 3,042,957 drop/s 0 error/s
| xdp_exception 0 hit/s
-t 4
|Summary 5,442,166 rx/s 0 err/s
| receive total 5,442,166 pkt/s 5,442,166 drop/s 0 error/s
| cpu:60 5,442,166 pkt/s 5,442,166 drop/s 0 error/s
| xdp_exception 0 hit/s
-t 6
| Summary 7,023,406 rx/s 0 err/s
| receive total 7,023,406 pkt/s 7,023,406 drop/s 0 error/s
| cpu:60 7,023,406 pkt/s 7,023,406 drop/s 0 error/s
| xdp_exception 0 hit/s
-t 8
| Summary 7,540,915 rx/s 0 err/s
| receive total 7,540,915 pkt/s 7,540,915 drop/s 0 error/s
| cpu:60 7,540,915 pkt/s 7,540,915 drop/s 0 error/s
| xdp_exception 0 hit/s
-t 10
|Summary 7,699,143 rx/s 0 err/s
| receive total 7,699,143 pkt/s 7,699,143 drop/s 0 error/s
| cpu:60 7,699,143 pkt/s 7,699,143 drop/s 0 error/s
| xdp_exception 0 hit/s
At this level, if you can verify that CPU:60 is 100% loaded, and packet
generator is sending more than rx number, then it could work as a valid
experiment.
-t 18
| Summary 7,784,946 rx/s 0 err/s
| receive total 7,784,946 pkt/s 7,784,946 drop/s 0 error/s
| cpu:60 7,784,946 pkt/s 7,784,946 drop/s 0 error/s
| xdp_exception 0 hit/s
after t18 it drop down to 2,…
Now I got worse than before since -t8 says 7,5… and it did 8,4 in the
morning. Do you have maybe a .config for me in case I did not enable the
performance switch?
I would look for root-cause with perf record +
perf report --sort cpu,comm,dso,symbol --no-children
--Jesper
