And until these two issues viewtopic.php?f=150&t=30126#p215699
aren't fixed IMO it's not a good idea trying to 'masquerade' low performance by sending eth0 IRQs to the big cores instead of solving the problems. At least on my board RX performance is pretty low and this won't change until correct RX/TX delays are configured (I mean, all the hard work has already been done, it's just pulling in patches into N1 kernel, running the test script and then adjusting the DT)
Very sorry for late response.
I was tackling this topic from the wrong way up. Sorry.
Right, I agree with all of you that it's not a right approach, but just deception.
We're trying to optimizing network performance based on you and other users' knowledge and suggestions.
In the meantime I wonder how the LanTest results shown in viewtopic.php?f=7&t=29932
have been created? Was this a different board revision (trace length differences between SoC and RTL8211)? Or other DT settings? At least I get very low RX values while TX is well above 100 MB/s.
For test, I tried to find better settings with LanTest and the trick I mentioned was also included.
Among items that I tried, the following two settings affect the performance conspicuously.
Just for information, I'm sharing them here.
1. The irq eth0 is assigned to big cores, CPU4-CPU5.
Code: Select all
# echo 4-5 > /proc/irq/24/smp_affinity_list
2. smb.conf is adjusted as following.
Code: Select all
server signing = No
write raw = Yes
read raw = Yes
getwd cache = Yes
socket options = TCP_NODELAY IPTOS_LOWDELAY SO_RCVBUF=131072 SO_SNDBUF=131072
min receivefile size = 2048
comment = share directory
path = /media/disk
browseable = Yes
guest ok = Yes
read only = No
create mask = 0777
write cache size = 1024000
use sendfile = Yes
strict allocate = Yes
allocation roundup size = 4096
oplocks = Yes
aio read size = 16384
aio write size = 16384