Maglev with Intel DPDK

[Kumar Narayanan]

Hi,
Just a broad-stroke question...

I understand that Maglev processes packets without kernel involvement. Does that mean that you're using Raw Sockets? Or, some other technique?

Is it reasonable to consider Intel DPDK where the driver for several Intel MAC controllers (X540, 82599EB etc.) run in user space and allows packet movements between NIC and user-space without Kernel involvement? This can also completely eliminate the need for context switches.

I've run Intel DPDK and do see PPS number to saturate 10GbE interface. I didn't run 64B packets but more in the 400B to 600B range average packet size.

Thanks,
-Kumar

[Richard Sargent]

Answers from one of the VM engineers:

The maglev socket primitives use the calls

recv(), recvfrom(), & send()

which are defined in linux  /usr/include/sys/socket.h
"kernel involvement" would be a function of how linux implements the above 3 calls and how the linux installation is configured.   You would need

to consult linux  kernel docs or kernel sources to get more details.

 

Thanks,
-Kumar