Hi! On 14:54 Fri 17 Apr , Devesh Sharma wrote: > thanks for replying michi, BTW how does a layer above network device > driver manages skb cache? how can we tune this skb cache size? any > idea? See http://lxr.linux.no/linux+v2.6.27.7/net/core/skbuff.c#L179 The struct sk_buff and the data area are allocated seperately. The size of the sk_buff struct is alwaly constant and does not grow with the packet size. The data area size is variable. 4096 may be a "magic number", because 4096 byte is the size of a memory page and you need some header+trailer for every packet. If you allocate more than one page at the same time, the allocation may take very long in some cases. The "normal" way to avoid this is to do allocate 2 non continuous pages and do scatter-getter IO. However, this is not done in the networking subsystem. If this is the source of the problem, the benchmark should also slow down at 4096 byte packet size, if you use gigabit ethernet with the some MTU. -Michi -- programing a layer 3+4 network protocol for mesh networks see http://michaelblizek.twilightparadox.com -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
