David Miller wrote:

From: Karl Hiramoto <redacted>
Date: Thu, 10 Sep 2009 23:30:44 +0200

  

quoted

I'm not really sure if or how many packets to upper layers buffer.
    

This is determined by ->tx_queue_len, so whatever value is being
set for ATM network devices is what the core will use for backlog
limiting while the device's TX queue is stopped.

I tried varying tx_queue_len by 10, 100,  and 1000x, but it didn't seem 
to help much.  Whenever the atm dev called netif_wake_queue() it seems 
like the driver still starves for packets  and still takes time to get 
going again.


It seem like when the driver calls netif_wake_queue() it's TX hardware 
queue is nearly full, but it has space to accept new packets.  The TX 
hardware queue has time to empty, devices starves for packets(goes 
idle), then finally a packet comes in from the upper networking 
layers.   I'm not really sure at the moment where the problem lies to my 
maximum throughput dropping.

I did try changing sk_sndbuf to 256K but that didn't seem to help either.

--
Karl