ANLAB-KAIST / Packet-IO-Engine

A high-performance and batching-oriented device driver for Intel 82598/82599-based network interface cards, the work is done in cooperation with ANLAB and NDSL.
http://shader.kaist.edu/packetshader/io_engine/
Other
144 stars 43 forks source link

ps_send_chunk() stops working after sending random sized packets #8

Closed tonimontana534 closed 11 years ago

tonimontana534 commented 11 years ago

Hi,

I tried testing the module more with the supplied packet_generator.c - if you add "packet_size = 60 + rand() % 1000;" after the for loop at line 399 (in send_packets()) - this makes the packet size random between 60-1060 bytes long - and after less than 1 minute of sending packets, suddenly PPS drops to zero.

Investigating it I saw that ps_send_chunk() returns 0 (0 packets sent) - and even if I try sending regular 60 byte packets after this - it won't send anything, the driver is dead.

The only workaround to this is to do "ifconfig down" and then "up" - then everything works, but if random sized packets are sent - again the same issue happens.

Any ideas?

ngilero commented 11 years ago

I tested and stumbled upon this problem too - my goal is a performance testing application with random packets for our network - and I see the same issue, if packet is random size - it doesn't work after a short while.

Any word from the maintainers of the project?

achimnol commented 11 years ago

Ok, I have confirmed the same problem with modified pspgen. We will look into that.

ngilero commented 11 years ago

achimnol - any updates on this?

achimnol commented 11 years ago

Seems to be fixed now. I will run test for the whole night and reopen it if the problem occurs again.

achimnol commented 11 years ago

After all night test, the problem has occurred again. :( This time, it does not stop completely but only a single interface is stopped among 4 interfaces in the system. And a SandyBridge server with 8 interfaces does not show this problem at all while two Westmere servers with 4 interfaces do. We still need to investigate this problem further, but I'm going to focus on other stuffs first. Any help will be appreciated!

achimnol commented 11 years ago

The last patch seems to work well, sustaining full load more than 7 hours without any problem.