What will the network interrupt handler do when NIC recieve data? - sockets

Question

As far as I know, when a packet arrives at the NIC, the DMAC will copy the packet to the kernel space. When the DMAC completes its work, it notifies the CPU, and then the CPU copies the data to the user space. Doing so will cause the memory to be read once and to be written twice. I wrote a simple program to simulate this process. This is the code:
# server.py
import socket
import sys
server = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
host = "70.202.0.116"
port = 12306
server.bind((host, port))
server.listen(5)
while True:
conn,addr = server.accept()
print(conn,addr)
while True:
data = conn.recv(4096)
if not data:
print("client has lost")
conn.close()
break
server.close()
# client.py
import socket
import sys
client = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
host = "70.202.0.116"
port = 12306
client.connect((host, port))
data = ''
for i in range(4096):
data += 'a'
while True:
client.send(data.encode())
client.close()
My machine has two numa nodes. At the first time, I disabled NIC Multi-Queue by ethtool -L eno1 combined 1, thus there is only one network interrupt left, and set the affanity by ehco 22 > /proc/irq/137/smp_affinity_list. Core 22 is on numa 1. Then I ran server.py. I use pcm-memory to moniter system memory bandwidth, and I got the expected output, the read-write ratio is close to 1:2.
But when I changed the affanity to core 0 which is on numa 0, I got totally different result. The read-write ratio is close to 1:1.
I want to know what does the interrput handler do during this process, why did I get different result?

Answer 1

increase read latency could be because device belongs to different numa_node. Check device where server and client is running belongs to which numa node
# cat /sys/bus/pci/devices/<PCI device>/numa_node